Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Rocketsonde buoy system : observing system simulation experiments Spagnol, Giancarlo (John) 2005

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
831-ubc_2005-105704.pdf [ 21.14MB ]
Metadata
JSON: 831-1.0052550.json
JSON-LD: 831-1.0052550-ld.json
RDF/XML (Pretty): 831-1.0052550-rdf.xml
RDF/JSON: 831-1.0052550-rdf.json
Turtle: 831-1.0052550-turtle.txt
N-Triples: 831-1.0052550-rdf-ntriples.txt
Original Record: 831-1.0052550-source.json
Full Text
831-1.0052550-fulltext.txt
Citation
831-1.0052550.ris

Full Text

Rocketsonde Buoy System: Observing System Simulation Experiments by GrANCARLO (JOHN) SPAGNOL B.Sc. (Physics) Simon Fraser University, 1971 M.Sc. (Atmospheric Sciences) M c G i l l University, 1975 A THESIS S U B M I T T E D F N P A R T I A L F U L F I L L M E N T O F T H E R E Q U I R E M E N T S F O R T H E D E G R E E O F D O C T O R O F P H I L O S O P H Y in T H E F A C U L T Y O F G R A D U A T E STUDIES A T M O S P H E R I C S C I E N C E T H E U N I V E R S I T Y O F BRITISH C O L U M B I A July, 2005 © Giancarlo Spagnol, 2005 Abstract The research presented here provides measurements of the negative impact of the Pacific and Arctic data voids on numerical weather predictability downstream over North America, and the extent to which the detrimental effects can be mitigated with rocket soundings from anchored buoys within the Rocketsonde Buoy System (RBS). The goals include the optimum deployment locations in the Pacific data void, rocketsonde sensor requirements, optimum height of the sounding profiles, number and configuration of the RBS array, real-time management requirements, daily launch times, launch periods during the year, number of rockets per buoy, and resulting buoy size. Observing System Simulation Experiments (OSSEs) are performed that assimilate virtual rocket soundings into an initial weather analysis and then measure the resulting change in forecast skill relative to reference forecasts. OSSEs are performed for 1 day in 2001, 21 days in 2002, 19 days in 2003 and 6 days in 2004. Reference forecast results are studied for 382 days. It is found that the optimum profile height is 6 km. Wintertime-storm forecast-error over western North America averages 18-25% more than summertime, rising to double for high-impact events. The forecasts from 00Z analyses score 6-12% better than the forecasts from 12Z analyses. The RBS engineering constraint is currently 200 of the 6 km altitude rocketsondes allowing a seven-month RBS operational period centered on the winter months delivering a daily 12Z profile. RBS mass and wind observations contribute fairly equally, sometimes cumulatively, to the forecast. A 5% launch tilt is acceptable. ii Compared to eastern North America, western North America has an averaged penalty of 20%, increasing to 35% during storms, caused by the Pacific data void. A 20% averaged improvement from better Pacific initialization is realistic. During high-impact situations, a three to six RBS strategic deployment can achieve 0.70 or more of this goal. On average, a six-RBS deployment delivers a 0.30 improvement; a 12-15 RBS deployment provides a 0.60 improvement. Real-time management and targeted-RBS operations enhance the results. In summary, deployment of an RBS is recommended because it would likely improve forecast skill over much of North America. i n Table of Contents Abstract ii Table of Contents iv List of Tables viii List of Figures xiii List of Acronyms and Definitions xxiii List of Abbreviations xxviii Acknowledgements xxix C H A P T E R 1 Introduction 1 C H A P T E R 2 A Systems Overview 5 2.1 Global 5 2.2 Observations 6 2.3 Four Dimensional Data Assimilation 7 2.4 Forecasts 8 2.5 Technology 9 2.6 Mesoscale Modeling 9 2.7 Summary 10 C H A P T E R 3 Data Coverage 11 3.1 The Current Observational Network 11 3.2 The Unbalanced Observational Network 13 3.3 Observations Within the Domain of Interest 14 3.4 Impact of Data-Void Areas 16 3.5 Costs to Society 19 3.6 Data-void Mitigation Efforts 21 3.7 New In-situ Observing Systems 22 3.8 Satellite Observing Systems and GPS Meteorology 25 3.9 Real Time Management 26 3.10 Summary 27 iv CHAPTER 4 Rocketsondebuoy System 29 4.1 Rocketsonde Usage 29 4.2 The Concept 30 4.3 Design Objectives 33 4.4 Progress to Date 33 4.5 Real Time Management 34 4.6 Other RBS Advantages 35 4.7 Multi-Rocketsonde Buoy Payload 36 4.8 Summary 36 CHAPTER 5 Data Assimilation, Numerical Weather Prediction 37 and Experimental Issues 5.1 Foreword 37 5.2 Initial Conditions 37 5.3 Physical Law Constraints 39 5.4 Data Assimilation 39 5.5 Comparison of the Assimilation Schemes 42 5.6 Quality Control 43 5.7 Model Error 44 5.8 Regional Models and Boundary Conditions 44 5.9 Data Impact Experiments 45 5.10 " Linear vs Non-Linear Error Growth 46 5.11 Verification 47 5.12 Summary 47 CHAPTER 6 The MM5 Modeling System 48 6.1 Foreword 48 6.2 Domain 48 6.3 Preprocessing 49 6.4 Model Numerics 49 6.5 Model Physics 50 6.6 MM5-SCM Implementation 50 6.7 MM5-3DVAR Implementation 51 6.8 Initial and Lateral Boundary Conditions 53 6.9 Visualization 55 6.10 Summary 55 CHAPTER 7 Experimental Methodology 57 7.1 Foreword 57 7.2 Basic Experiments 57 7.3 OSSE Methodology 58 v 7.4 Reference and Benchmark Atmospheres 59 7.5 OSSE Atmospheres 61 7.5.1 OSSE-PAC Atmospheres 62 7.5.2 OSSE-WNA Atmospheres 64 7.5.3 OSSE-ARC Atmospheres 66 7.6 Verification Atmospheres 66 7.7 Groups 67 7.8 Impact Metrics 68 7.9 Summary 70 CHAPTER 8 OSSE Results 71 8.1 Research Objectives 71 8.2 The Study Approach 72 8.3 Cases Overview 73 8.4 OSSE Case Examples 74 8.4.1 Test Period WNA Case: From 12 Dec(S )12Ztol5 75 Dec@00Z (2001) 8.4.2 Test Period ENA Case: From 20 Feb @ 12Zto23 Feb 84 @00Z(2002) 8.4.3 Study Period WNA Case: From 15 Oct @ 12Z to 19 86 Oct @ 00Z (2003) CHAPTER 9 General Findings 95 9.1 Foreword 95 9.2 EDAS 97 9.2.1 EDAS Failures 97 9.2.2 EDAS Value 98 9.3 RBS NWP Support 100 9.3.1 Synoptic Hour Support 100 9.3.2 Summer Period Denial 102 9.3.3 WNA Quiet Periods 104 9.4 Regional NWP Forecast Errors 105 9.4.1 Sources of Forecast Errors 105 9.4.2 Initialization & Model Attributes 107 9.4.3 Effect of Boundary Conditions 109 9.5 Data Assimilation 111 9.5.1 SCM Sensitivity Studies 111 9.5.2 3DVAR Sensitivity Studies 113 9.5.3 SCM/3DVAR Comparison 118 9.5.4 Quality Control 120 9.6 Validity of the OSSE Procedure 122 9.6.1 The Study Period 122 9.6.2 Basic Screening Test 122 VI 9.6.3 NWP Error Linear Growth 123 9.6.4 The Uneven Pacific Analysis 125 9.7 Impact of the Data-voids 126 9.7.1 Pacific Data-void 126 9.7.2 Polar Data-void 130 9.8 Pacific RBS Operations 131 9.8.1 Location 131 9.8.2 Buoy Array Size 135 9.9 RBS Data Issues 141 9.9.1 Profile Height 141 9.9.2 Meteorological Variables 147 9.9.3 Observation-error Sensitivity 149 9.10 RBS Targeting 149 9.11 RBS Deployment 154 9.12 Arctic Operations 156 9.13 Observation System Issues 158 9.13.1 Extending the Network 158 9.13.2 Balanced Operational Network 159 9.14 The Composite Observing System 164 9.14.1 Possible Improvements 164 9.14.2 A Composite Mix Simulation 164 9.15 Summary 169 CHAPTER 10 Conclusions 176 10.1 Summary of Findings 176 10.2 Critique oftheOSSE Method 179 10.3 Tips for Future OSSE Work 181 10.4 Weather Observations in an Ideal World 181 BIBLIOGRAPHY 184 Appendix A Dropsonde Sensors 189 Appendix B U B C Computer Systems 196 Appendix C Case Dates and Assorted Information 197 vu List of Tables 3.1 For 4 February 2004 at 00Z, the number of worldwide observations 9 available at E C M W F . The first column is the code name. The second column categorizes the data as in-situ or remotely sensed. The third column states whether the data is surface (sfc) or upper air (ua). The fourth column indicates i f the data is valid at a single level (hor) or at multiple levels (ver). The last column has the amount of data available at the 00Z synoptic hour. This information is taken from http://www.ecmwf.int/products/forecasts/dVcharts/monitoring/coverage/dc over/ 6.1 The errors are based on those used in the F N M O C N O G A P S model and 49 the E C M W F model. 6.2 The attributes of the E T A model, the ETA104 data product and the M M 5 54 model implementation. 8.1 Quality of the Pacific analyses for the 13Dec01 case, using root mean 73 square difference (RMSD) and correlation coefficient (COR) as the metrics, for forecasts at various hours after data addition ( A D A ) . Verification is for V E R W E S T in the VERwest region. Rows are for atmospheres with different amounts of soundings available (see text). The REFOO forecast starting at OOZ verifies better than the REF12 forecast starting at 12Z (12 hrs before). The B E N O S S E atmosphere, with the virtual radiosonde soundings added at OOZ also verifies better. From these statistics, it is concluded that the OOZ Pacific analysis was updated in a proper manner. 8.2 The 13Dec01case. Forecast error for the various P A C Group 78 configurations. R M S D is shown as forecast error. 8.3 Rainfall data, 15-21 October 2003. The data was provided by . 87 Environment Canada. 8.4 The REF12 forecast was started 15 October 12Z. The P A C Group data 92 was added on 16 October OOZ, 12 hrs later. The rainstorm moved onshore between 16 October 12Z and 17 October OOZ. The P A C Group data was assimilated with M M 5 - 3 D V A R . viii 9.1 REF12 and REFOO HT70 R M S D results every 12 hrs of the forecast 96 duration. Lower R M S D is better. The REF12 results start with the 12 In-forecast. The same valid time REFOO results start with the initial time. The 07-11 February 2004 forecast failures showing a "series of busts". The heavier solid arrows follow the same time group. REFOO is worse than the REF12 even though REF12 is 12 hrs longer. Heavy downward sloping lines show that successively, shorter-length forecasts do not get better. Lower R M S D is better. The shorter downward sloping lines compares REF12 and REFOO for the same forecast length. REF12 is often markedly worse than REFOO. A n interesting observation when viewing the synoptic charts: the W N A atmospheric flow was shifting from a zonal flow on 07 February 2004 to a meridional flow with a major change occurring on 10 February 2004 9.2 E T A First Guess (EFG) is the M M 5 forecast starting with the E T A 12 hr 97 forecast. REFOO is the forecast from the subsequent E T A OOZ E D A S update. A comparison estimates the average value that E D A S has provided. The E T A 12 forecast also carries on from its own 12 hr forecast. The average improvement for the same 143 cases is shown. The E F G VERwest improvement is comparable to the E F G VEReast improvement. The E T A 12 VERwest improvement is higher than the E T A 1 2 VEReast improvement. Lower R M S D represents a better forecast. 9.3 VERwest and VEReast R M S D averages of all 5 winter 01/02 cases. 99 REFOO verifies better than R E F 12 suggesting a diurnal effect. The VEReast forecasts verify better than the VERwest forecasts, indicating the effects of the Pacific data-void. 9.4 The Study Period VERwest HT70 R M S D results show that the forecasts 100 started at OOZ are better than the ones started at 12Z by 6 to 12%, which is almost half of the forecast improvement possible. Over VEReast, the effect is less pronounced. VERwest starts with a nearly 20% penalty over VEReast due to the Pacific data-void. The penalty reverses itself after the 48 hr forecast period due to other influences (weather regime?). 9.5 For the Test Period VERwest winter cases, M S L P R M S D of about 0.2 101 kPa for 24 hr forecasts increases to about 0.4 kPa for 60 hr forecasts. The 60 hr R M S D summer values are less than half of the winter values, showing that weaker synoptic systems in summer are easier to forecast, even with the Pacific data-void lack of data. The Study Period R M S D values for the summer cases are 18-25% lower than winter cases. 9.6 VERwest HT70 R M S D results during this quiet time were much better 103 than the winter average (including the quiet periods), and tending towards the summer values. i x 9.7 A comparison of the forecast skill of E T A and MM5 shows that E T A 106 forecast skill is much higher than the M M 5 . An estimate of the error growth rates shows that the VEReast error growth is greater than the VERwest error growth. Error growth rates for the averaged VEReast synoptic situation and one period (6-22 January 2004) where there was small cross-mountain flow. There was little difference between the R M S D results. 9.8 The lateral boundary tendency updates are varied from every 3 hrs to 107 every 12 hrs. VEReast forecasts are affected more than VERwest forecasts. 9.9 Analysis of S C M sensitivity to the 5, 10, 15 ROI. The average is over 7 109 cases. The RMSD does not increase until the ROI is more than 10. The 23Jan03 case shows that when bad Pacific vROB data is used, the R M S D can increase rapidly with increasing ROI. 9.10 3D V A R sensitivity to the observation scale length for the NineBuoy 115 15Oct03 OSSEs The optimum result seems to be a scale length between .18 and .25. It was set to .20. 9.11 M M 5 - 3 D V A R WS70 RMSD results for the 16Oct03 case. Temperature 115 and humidity (TTD) are taken together as the mass variables. Wind is indicated by WND. The three vROB modes used are the regular observations including both WND and TTD and then WND and TTD alternately denied. Scale lengths are alternately set at .20 and .25. The B E N vRAOBS continue to include both WND and TTD at the same scale length setting as the vROBs. The results showed that the WS70 deteriorated with increasing WND scale length. 9.12 The HT70 RMSD forecast comparison between the results of S C M - D A 116 and 3DVAR-DA data assimilation for 3 cases. The S C M - D A scored better. 9.13 The effect of S C M quality control on HT70 RMSD for the 7Nov02 case. 118 For most periods, the quality control has lowered the RMSD. 9.14 The B E N SCM-DA (42 cases) and 3DVAR-DA (11 cases) are averaged 122 and compared to the same case REF12 and REF00. For M M 5 - 3 D V A R linear growth cannot be assumed after 60 hrs A D A . x 9.15 VERwest results for the 20Feb02 case, showing RMSD values after the 123 addition of vROBs for various atmospheres (rows) at various forecast times (columns). Adding "irrelevant data" from the ThreeBuoy and SixBuoy arrays deteriorates the forecast. Adding the NineBuoy vROBs mitigates the effects of the bad data and improves the forecast for some time periods. The best result is by not adding data from the ErnPAC_RSB array, and adding data from only the C e n P A C R S B array. 9.16 VEReast R M S D values for the 13Dec01 case for different atmospheres 125 (rows) and different forecast durations (columns). A comparison of W N A void with PACvoid results show a deterioration of about 35% in G A I N during the 24 and 36 hr A D A forecast period. Including the WNAall vRAOBs to fill PACvoid results in a 20% improvement in the forecast. Including the B E N area results in a loss. This result suggests that a data void can cause a 20% loss in forecast accuracy and as much as a .35 loss in GAIN. 9.17 The winter 02/03 M M 5 - S C M and the winter 03/04 M M 5 - 3 D V A R results 127 for the data-void area experiments. Relative to the W N A void atmosphere results, the PACvoid atmosphere makes most of the G A I N and % improvement. The B E N atmosphere does not improve the forecasts as much. 9.18 The Arctic data-denial eliminates the vRAOBS from the top 10° of B E N 128 to become BENs. The results are an average over 15 cases from 06-20 January 2004. Except for the longest forecast period, the results are indeterminate 9.19a Winter 02/03 results PAC Group OSSE results with M M 5 SCM-DA. 134 Most of the G A I N is made by the smaller arrays. Larger arrays deliver more G A I N but with diminishing returns. The forecast improvement for the 6-12 RSBs are about half of what is possible. 9.19b Winter 03/04 results PAC Group OSSE results with M M 5 - 3 D V A R D A . 137 Study Period PAC Group OSSE results. Most of the G A I N is made by the smaller arrays. Larger arrays deliver more GAIN but with diminishing returns. The TwelveBuoy array was available to improve the forecast results by half of what is possible. 9.20a Winter 02/03 M M 5 - S C M ThreeBuoy OSSE results for 19 cases. 141 Compared to the 13Dec01 case (Figures 9.13), there is much smaller GAIN. 9.20b Winter 03/04 MM5-3DVAR ThreeBuoy OSSE results. There was little 142 G A I N for the 8 cases analyzed. x i 9.20c Table 9.20c. Variable height TwelveBuoy v R O B OSSEs assimilated with 143 M M 5 - 3 D V A R . Most of the benefit of the 8 cases was delivered with 6 km profiles. 9.21a The impact of the wind (WND) and mass ( T T D for temperature and 145 humidity) observations on the O S S E results. The observations for the 20Nov02 and 24Mar03 cases were assimilated with M M 5 - S C M . 9.21b The impact of the wind (WND) and mass ( T T D for temperature and 146 humidity) observations on the O S S E results. The observations for the six cases were assimilated with M M 5 - 3 D V A R . 9.22 The 15Oct03 VERwest HT70 R M S D results for various observation-error 147 magnitudes. For this case the results do not seem sensitive to the observation errors. 9.23 The R M S D improvements that a R B S targeting strategy can provide are 151 less but comparable to the E r n P A C _ R S B TwelveBuoy benefits. 9.24 The results of the phased deployment arrays. The winter 02/03 S C M - D A 154 (3 cases) showed GAINs of near 0.25 for the 6Cb6k buoy array. The winter 03/04 3 D V A R - D A (5 cases) showed GAINs of near 0.35 for the 36 hr period. 9.25 The arctic deployed R B S units provided an averaged (7 cases) G A I N of 155 about .15 up to and including 48 hr A D A . The data became erratic after 48 hrs A D A . 9.26 For the Study Period, 8 cases of S C M - D A (top) and 3 cases of 3 D V A R - 160 D A v R A O B s are presented at 5,6,7,8,9,10 grid-point increments. The averaged R M S D results are a slowly varying function of grid-point spacing. 9.27 Relative impacts of simulated aerosonde, R B S , aircraft, surface and 166 dropsonde data. xii List of Figures 3.1 Data sets available from the existing observing systems, ordered 13 approximately by their valid level within a data assimilation 3.2 (a) Radiosonde sounding locations (dots) over North America. The region 14 of mid-tropospheric in-situ sounding-data paucity, identified as the "Pacific data-void", is (b) sandwiched between data-rich layers of surface observations and near-tropopause observations. Typical numbers of in-situ observations per analysis period is given for each region. A I R E P = Aircraft Report (manual); A M D A R = Aircraft Meteorological Data Reporting Relay; and A C A R S = Aircraft Communications Addressing and Reporting System. Satellite-derived data ( S C A T , S A T O B , A T O V S ) is of variable coverage, density and frequency. The states/provinces marking the eastern edge of the data void are: A K = Alaska, B C = British Columbia, C A = California, O R = Oregon, and W A = Washington. 3.3 (a). A normal zonal flow at 12Z, 9 November 2002. Weather 16 disturbances over the Pacific data-void affects both W N A first and E N A later, (b) A meridional flow at 12Z January 2003. Weather disturbances over the eastern and central Pacific move to the Alaska areas not affecting W N A and E N A . Weather disturbances originating in the arctic data-void affect E N A later. 3.4 (left) Data rejections (circled stations) during the windstorm of 14 17 December 2001, and (right) during the extreme rainfall event of 16 October 2004 are shown. Both cases are discussed in Section 8.4. 4.1 A rocketsonde launch. The rocket will carry a sounding package (sonde) 30 to a maximum of 8 km altitude before the sonde, with parachute, separates from the rocket and falls back to the ocean surface. While falling, the sonde will sense the atmosphere for its ambient temperature, humidity and pressure. It could also receive GPS signals to determine winds and altitudes. The sonde will transmit its data to the buoy. The data stream signal will be amplified and retransmitted to a satellite that will relay the data to shore where it is added to the G T S and carried to the data users. 6.1 The E T A 22 km domain is the larger area bordered by the curved solid 52 line. The M M 5 45 km domain is bordered by the inside rectangular area. The E T A 104 product domain is bordered by the inside dashed line. xiii 7.1 The OSSE Timeline. The control or "reference-atmosphere" is the REF12 56 forecast that excludes any new data. The OSSE forecast has new data assimilated at 00Z. After data addition, data impact is measured as the difference in two-to-three-day evolution of the REFXX and OSSE atmospheres from the verification atmosphere. 7.2 The heavy rectangle outlines the BEN region, which for this OSSE is 59 filled with a uniformly distributed, high-density array of vRAOBs to 16 km altitude (over both continent and ocean). For the PACvoid experiments, virtual sounding data are excluded from the mostly-oceanic region west of the idealized west coast of N. America, and south of the idealized US-Mexican border. For WNA void experiments, the simulated west coast of N. America is shifted further east, with sounding data excluded everywhere west and south of that shifted coastline. 7.3 The PACRBS atmosphere consists of only low-altitude vROBs at the 61 locations indicated, which are at the intersection of every 5° parallel and meridian. Subsets include the eastern Pacific (ErnPAC_RBS) and central Pacific (CenPAC_RBS). 7.4 The P A C T A R atmosphere consists of only low-altitude vROBs at the 62 locations indicated, which are at the intersection of every 5° parallel and meridian. When referenced the ErnPAC_RBSs will be preceded by an E. The CenPAC_RBSs will be preceded with a C. 7.5 The WNARBS atmosphere consists of only low-altitude virtual 63 rocketsonde soundings at the locations indicated, which are at the intersection of every 5° parallel and meridian. There are no soundings elsewhere in the domain. 7.6 The ARCRBS atmosphere consists of 6 km apogee vROBs assimilated at 64 the locations indicated. 7.7 OSSE verification domains over the data-rich continent are divided into 65 VERwest and VEReast, which are placed just east of the PACvoid shoreline and the WNAvoid-shifted shoreline, respectively. The verifications over VERwest and VEReast will represent the NWP results over WNA and ENA respectively. 8.1a Mean sea-level pressure (MSLP dark isobars every 0.4 kPa) from the 74 REF12 atmosphere for the intense mid-latitude cyclone of 14 December 2001. Overlaid is the difference (DIF = VER - REF12) between the verification atmosphere and the reference atmosphere. DIF contours every 0.1 kPa are thin dashed for positive and dotted for negative differences, (a) Valid 00Z 13 Dec showing the 12 hr forecast. A 99.8 kPa storm near 50N 140 W is moving into a long wave trough positioned off the BC coast, with MSLP values that are already 0.5 kPa too high. xiv 8.1b Val id OOZ 14 Dec showing the 36 h forecast. Forecast cyclone movement 74 is too slow. DIF values (-0.9 kPa over B C , and +0.9 kPa west of Oregon) indicate that the actual low center has moved onshore. 8.1c Val id OOZ 15 Dec showing the 60 hr forecast. This forecast is quite poor, 75 with M S L P ridges and troughs out of phase as indicated by the large positive and negative DIF values over BC/Washington and the N E Pacific, respectively. 8.2a Effect of adding virtual rocketsonde soundings (ErnPAC_RBS 75 ThreeBuoy) to the otherwise Pacific data-void forecast for the 12Dec01 cyclone. Plotted are M S L P (dark isobars every 0.4 kPa) and the M S L P difference (DIF, thin lines every 0.1 kPa dashed for positive and dotted for negative) between the P A C _ R B S atmosphere and the R E F 12 atmosphere. Val id OOZ 13 Dec at 12 hr into the forecast, at. the time of R B S data addition. DIF fields show that the cyclone center is better captured, with more accurate (0.7 kPa lower) central pressure and better location (further west). 8.2b Val id OOZ 15 Dec at 60 hr into the forecast (48 hr after v R O B data 76 addition), showing only the M S L P DIF field for clarity. The improvements due to v R O B D A over the eastern Pacific have moved over the continent, and indicate that the forecast better captures a deeper low that moves east faster. Meanwhile, near zero DBF values over the ocean now indicate that data-void effects from further upwind have moved.over the N E Pacific. 8.2c Close-up of the western N . America verification area valid OOZ 15 Dec 76 (60 hr forecast, 48 hr A D A ) , but showing the improvement in R M S D associated with the addition of the ThreeBuoy v R O B data. Plotted is the difference between (VER-REF12) R M S D minus ( V E R - P A C _ R B S ) R M S D , where positive values (dashed contours) indicate R M S D improvement in the O S S E forecast compared to the R E F 12, and negative (dotted contours) indicate degradation. Evident is a region of improvement over B C , W A and ID, and a dipole region of improvement and degradation across the Rocky Mountain crest in the northeastern portion of the figure. 8.3 Growth of forecast errors with time after data addition ( A D A ) of various 77 numbers of R B S sounding locations, for the 12 Dec storm. Forecast error metric is root mean square difference (RMSD) between the O S S E forecasts and corresponding V E R analyses, valid at the same time, and verified in the VERwest region. Initialization is from 12Z on 12 Dec, with data addition at 00Z on 13 Dec. Smaller R M S D is better. xv 8.4a Changes in 60 hr forecast quality of M S L P measured over the continental 78 VERwest region, for different numbers and arrangements of sounding buoys upstream over the Pacific, for the Decl2 storm. Various metrics of forecast quality valid at 00Z on 15 Dec (48 h A D A ) are used: (a) R M S D of the O S S E forecasts vs. the REF12 forecasts valid at the same time, where smaller R M S D is better. 8.4b Same as (a), but normalized to give the relative gain in accuracy, where 79 G A I N = 0.0 for forecasts started from the REF12 background-state atmosphere of no soundings added, and G A I N =1 .0 for the idealized scenario of a dense array of added soundings covering all of the Pacific (BEN). Larger G A I N is better. 8.4c Relative G A I N per sounding buoy, where the data from (b) is divided by 79 the number of buoys. Larger G A I N per buoy is better. 8.5 The effect of increasing numbers of sounding buoys on M S L P forecast 81 quality, for different forecast durations for the 13Dec01 case. C O R is the spatial correlation of the O S S E forecast with the corresponding verification analysis (VER) valid at the same time. The forecast initialization time was 12 Dec at 12Z, and data addition occurred on 13 Dec at 00Z. A l l values for 36 hr after data addition ( A D A ) are interpolated, because the windstorm caused a power outage and a computer failure resulting in the missing verification data. Larger C O R is better, with C O R = 1 representing a perfect forecast. 8.6 Forecasts of M S L P (solid lines every 0.4 kPa) for the P A C v o i d _ R S B 80 atmosphere, where vROBs up to the mid troposphere were assimilated at the ThreeBuoy E r n P A C R B S sites, and v R A O B s at every fifth grid point (i.e., 225 km horizontal spacing) up to 16 km altitude were retained over land. Dashed and dotted contours represent the positive and negative difference (DIF) between the REF12 and the ThreeBuoy PACvoid_RBS atmosphere in 0.1 kPa units, (a) Valid at 00Z on 13 Dec 2001, (12 hr forecast, 0 hr A D A ) . Comparison with the REF12 atmosphere of Figure 8.1a shows improvements, with a weaker ridge over B C and a sharper trough along the coast, (b) Valid at 00Z on 15 Dec 2001 (60 h forecast, 48 hr A D A ) . Comparison with the 60 hr REF12 forecast (Figure 8.1c) shows improvements due to R B S data addition; namely, the low over Alberta is located further east, and the ridge over eastern B C is stronger. xvi 8.7 Impact of adding/denying v R A O B s over land to a scenario where vROBs 82 would already exist over the Pacific. Forecast errors ( R M S D , ordinate) of M S L P over VERwest are plotted for various numbers of sounding buoys (abscissa), with all results for 48 hr A D A in the 13Dec01 storm. The P A C _ R B S curve has only sparse vROBs over the Pacific (with no soundings over land), while the PACvoid_RBS curve is for the sparse R B S soundings over the Pacific plus v R A O B s spaced 225 km apart over land. For the VERwest region, R B S soundings were more important than continental soundings for this storm. Smaller R M S D is better. 8.8 Increase in forecast errors with time similar to Figure 8.4, but for the 83 20Feb02 case of Figure 8.9, with a virtual coastline shifted east (Figure 7.2), with vROBs added at the W N A R B S locations of Figure 7.5, and with results verified in the VEReast domain of Figure 7.7. Smaller R M S D is better. Results show that the effects of a poor Pacific analysis update have moved over the VEReast area. 8.9 O S S E result for the 20Feb02 storm,.but where the virtual west coast of 84 North America is shifted 2 0 + ° eastward. ThreeBuoy vROBs are added at the W N A _ R B S locations of Figure 7.5, with v R A O B s retained covering the remainder of the continent east of the shifted coastline (Figure 7.2). Shown are M S L P (solid contours) for a 12 hr forecast valid at 00Z on 21 Feb 2002. Superimposed are difference DIF contours ( W N A _ R B S -REF12) of the O S S E forecast from the verifying analysis. Although positive DIF values (dashed) appear where the R B S soundings were added, negative values (dotted) west of the B C coast corresponds to deepening of the approaching low, which is a change in the wrong sense. 8.10 G O E S water-vapour image composite for the Pacific. White regions 86 indicate large moisture content in the mid to upper troposphere. The images shows the connection between the intertropical convergence zone at about 30N 150E and the mid-latitude westerly flow at about 45N. This image was produced by N O A A and posted on their web site. 8.11 Rutherford Creek between the villages of Pemberton and Whistler, B . C . 88 Floodwaters washed out the bridge on October 18 t h near 4 a.m. Three cars drove off the bridge into the river before the washout was noticed. Lives were lost. Photo courtesy of McCollor (2004). xvn 8.12 (a) 36 hr rainfall accumulations valid 16 Oct OOZ. (b) 48 hr rainfall 89-accumulations valid 16 Oct 12Z. (c) 60 hr rainfall accumulations valid 17 90 Oct OOZ. (d) 72 hr rainfall accumulations valid 17 Oct 12Z. The solid lines are the rainfall accumulations for the R E F 12 forecast starting at 15 Oct OOZ. The dashed lines are the added rainfall accumulations predicted by the TwelveBuoy P A C _ R S B forecast starting at 15 Oct OOZ. The TwelveBuoy R B S forecast advanced the rain faster and resulted in higher early accumulations closer to the actual event. 8.13 The dark lines are the 70 kPa 12 hr forecast height contours valid 16 91 October OOZ by the R E F 12 forecast started at 15 October 12Z. The dashed lines are the DIF between V E R and the R E F 12 forecast. The negative values show that the forecast was 20 to 50 m too high in the vicinity of the low near 50N 145W and over W N A . V E R was the E T A OOZ analysis that was probably analyzing heights too high in the eastern Pacific since E D A S was slow to react to actual events. 9.1 HT70 valid OOZ, 11 February 2003, the middle of the quiet period 2-16 102 February 2003. The locations of the R B S arrays are shown. The arrow indicates that the HT70 flow over E N A is primarily from the arctic. 9.2 The Study Period averaged M M 5 error growth (after 12 hrs) 104 corresponding to results shown in Table 9.4. After the first 12 hrs, the error growth rates are approximately linear. The VEReast error growth has a larger magnitude than the VERwest error growth. The average is over 375 cases. Smaller R M S D is better. 9.3 A comparison of the E T A vs VERwest forecasts. This sample was taken 108 during the winter of 02/03. Except for the initial periods, the VEReast forecasts were worse than the VERwest forecasts. The VEReast error growth rates were higher. This is a reflection of the respective weather regimes (El Nino winters over VERwest). The averages are over 143 cases. Smaller R M S D is better. 9.4 (a) For the 23Jan03 case, the 72 hr ThreeBuoy HT70 R M S D DIF with a 111 S C M - D A ROI of 10. There was hardly any impact left over VERwest. Most of the positive benefit went to Alaska, the negative effect went to the Hudson Bay area. The arrow shows the approaching direction of the cyclone with the head indicating the position at OOhr A D A . (b) Varying the ROI from 5 to 20 grid points resulted in appreciable W N A HT70 R M S D scatter. xviii 9.5 3 D V A R HT70 innovations for the 15Oct03 case from a single R B S 114 sounding located at SON 150W. The units are in m. The background error scaling parameter is set at 2.0. The observation scaling parameter from frames a to f is increased. The influence of the single observation increases dramatically as the scale length increases. 9.6 Variation of forecast errors (ordinate) with increasing area coverage 126 toward the west of R B S buoys (abscissa), for different forecast durations (plotted curves). Error metric is M S L P R M S D over VERwest averaged over all winter cases. Results show that VEReast M S L P forecast errors are reduced as more virtual sounding data are added to the OOZ forecast, but only through PACvoid along the abscissa. Further extension of R B S coverage westward does not improve forecast quality, and even reduces it for the 60 hr forecast. Smaller R M S D is better. 9.7 M S L P forecast-error magnitudes (kPa) averaged over the four high- 130 impact winter storms (12Dec01, 17Feb02, 15Mar02, and HApr02) . (a) M S L P R M S D of 12 hr forecasts valid at OOZ. The subjectively-drawn heavy lines highlight maximum errors, and indicate that R B S data could have been added there to maximum advantage in forecasting these winter 2001-2002 storms, (b) R M S D of 24 hr forecasts valid at 12Z. Xs are added at the error-maxima locations from the previous 12 h, with arrows toward the new maxima showing how the uncorrected errors propagated for these cases. 9.8 Potential R B S sites (Xs) optimized to reduce the maximum forecast errors 131 shown in Figures 9.7. Compare these potential R B S networks with the existing, denser radiosonde sounding network over land (Figure 3.2a.). 9.9 HT70 12 hr forecast error magnitudes (m) averaged over 220 Study Period 132 winter cases. No virtual soundings have been incorporated into these runs. The maximum near 50N 145W is the outstanding Pacific feature. Another maximum occurs near 5 ON 160W, both agreeing with the high-impact analysis of Figure 9.7. The near coast maximum is missing suggesting that may only appear with high impact cases. The analysis shows the high and variable forecast error caused by the mountains. 9.10 Reduction of M S L P R M S D forecast errors (ordinate) over VERwest with 133 increasing number of sounding buoys over the Pacific (abscisa), averaged over the two high-impact storms (12Dec01 and HApr02) known to have good analyses. For both the 36 and 48 A D A forecasts, i f only a reasonably small number of R B S sites is economically feasible, then most of the benefit was achieved with the SixBuoy array. Smaller R M S D is better. xix 9.1 l a The S C M - D A O S S E HT70 R M S D results for winter 02/03 are averaged 135 over 24 cases. This graph corresponds to the results shown on Table 9.19a. 9.11b The S C M - D A O S S E G A I N per buoy after 48 hrs A D A . Most of the 136 G A I N is made with the SixBuoy Array. 9.12a The 3 D V A R - D A O S S E HT70 R M S D results for winter 03/04 are 138 averaged over 8 cases. This graph corresponds to the results shown in Table 9.19b. 9.12b Figure 9.12b. The 3 D V A R - D A O S S E G A I N per buoy after 48 hrs A D A . 138 Most of the G A I N is made with the Three and SixBuoy Array. 9.13a Forecast error growth (MSLP R M S D along the ordinate) in the VERwest 140 area as a function of R B S sounding altitude (multiple curves), for various forecast durations (abscissa). These OSSEs are for the 13Dec01 storm, and they all use the E r n P A C R S B ThreeBuoy atmosphere of Figure 7.3. Smaller R M S D is better. The B E N curve is shown for comparison, which would fill the whole Pacific region of Figure 7.2 with about 100 virtual radiosonde sites with soundings up to 16 km 9.13b For the high-impact 13Dec01 storm, results indicate that three buoys 140 launching rocketsondes to only 4 or 6 km would have eliminated near 20% of the forecast error and would have achieved a G A I N near 0.50, compared to the much more extensive and expensive soundings of B E N . 9.13c G A I N from increasing the R B S sounding altitude for a ThreeBuoy 141 configuration for the 12Dec01 storm. The 48 hr A D A results of Figure 9.13b, normalized to show the gain ( R M S D error reduction) normalized between 0.0 gain (highest error, associated with zero v R A O B S of REF12) and maximum gain of 1.0 (least error, B E N scenario of hundreds of vRAOBs) . The most G A I N for the least altitude was achieved with soundings up to 4 to 6 km. 9.14a The ThreeBuoy results of Table 9.20a. For the 19 cases analyzed, there is 142 little G A I N . 9.14b The ThreeBuoy results of Table 9.20b are shown. Three was little G A I N 143 for the 8 cases analyzed 9.14c The TwelveBuoy results of Table 9.20c are shown. The results became 144 erratic at the end of the forecast. 9.14d The TwelveBuoy 3 D V A R - D A G A I N 48 hrs A D A . A G A I N of near 0.60 144 was achieved by 6 km profiles (8 cases). A G A I N of 0.70 was made by 8 km profiles. xx 9.15a D D H H is the day and hour group in October 2003. The low-pressure 148 centre positions are indicated by the ellipses. The low pressure trough positions are indicated by the dashed lines. The 1500 features are the REF12 M M 5 12hr forecast positions. The grey line indicates the system cloud taken from a N O A A G O E S image. The R S B E2, E7, E l l , E14 positions are indicated. Profiles are taken at 1500 to 6 km. 9.15b The impact of the R S B targeted profiles is shown. The heavy solid lines 149 are the Targeted R B S reanalysis valid at OOZ on 15 October 2003. The thinner dashed lines are the HT70 (m) difference between the Targeted O S S E analysis and the REF12 12 hr forecast. R B S targeting lowered the heights ahead of the trough which resulted in a more intense system and a better forecast. 9.15c The 1600 features are the REF12 M M 5 12hr forecast positions of the 150 surface low and low pressure trough valid 16 Oct at OOZ. The grey line indicates the system cloud taken from a N O A A G O E S image valid half hr before. The R S B E5, E6, E7, E10, E13 positions are indicated. Profiles to 6 km are provided at OOZ 16 October 2003. 9.16a Comparison of R M S D error growth for the two R B S configurations of 152 Figure 9.8 using maximum sounding altitude of 6 km for both. Results are averaged over the storms of 13Dec01, 17Feb02, 15Mar02, and HApr02 . Smaller R M S D is better. Both R B S arrays have reduced the errors close to the B E N values for the 36 and 48 hr A D A forecast periods. 9.16b The average M S L P R M S D 48 hrs A D A for the storms of 13Dec01, 153 17Feb02, 15Mar02, and HApr02. The profiles are to 6 km. Smaller R M S D is better. The absolute improvement is about 15%. Both R B S arrays have reduced the errors close to the B E N values. The G A I N is close to 0.70. 9.17 The Study Period 12hr HT70 R M S D errors (m) over the arctic areas from 155 an average of 220 cases. Over Alaska, Yukon and the Northwest Territories, the 12hr forecast errors are greater than the Pacific data-void forecast errors (Figure 9.9). The requirement for more observations is obvious. x x i 9.18 Impact of adding/denying virtual radiosonde soundings over land to a 156 scenario where R B S soundings would already exist over the Pacific. Forecast errors (RMSD, ordinate) of M S L P (kPa) over VERwest are plotted for various numbers of sounding buoys (abscissa), with all results for 48 hr A D A for the 13Dec01 storm. The P A C _ R B S curve has only sparse R B S soundings over the Pacific (with no soundings over land), while the PACvoid_RBS curve is for the sparse R B S soundings over the Pacific plus v R A O B S spaced 225 km apart over land. For the VERwest region, v R O B S were more important than continental v R A O B s for this storm. 9.19a This figure shows the 16Mar02 VEReast M S L P R M S D (kPa) associated 158 with reduced amounts of v R A O B thinning. BEN10 is the coarsest spacing, and BEN05 is the finest spacing, corresponding to the actual average radiosonde spacing over North America. Smaller R M S D is better. 9.19b The 16Mar02 G A I N after 48 hrs A D A for the OSSEs that are shown in 159 Figure 9.19a. BEN10 is the coarsest spacing, and BEN05 is the finest spacing, corresponding to the actual average radiosonde spacing over North America. Larger G A I N is better. 9.20 The results of the Network Thinning OSSEs shown on Table 9.26. The 161 winter 02/03 (8 cases) S C M - D A results and the winter 03/04 (7 cases) results show a slowly varying function of v R A O B spacing. 9.21 The actual W S R dropsonde locations (1999-2004) are shown. This 163 graphic is courtesy of James Charbonneau and Jenn Mundy of the U B C Prediction Research Team. 9.22 Aerosonde flight paths and commercial aircraft flight paths used for the 164 composite mix simulation. These are non-dropsonde flights, so horizontal flight-path data is gathered only at the altitude of the aircraft. xxn List of Acronyms and Definitions Page % Difference % Difference between REF 12 and OSSE RMSD values 68 relative to a perfect score 3DVAR 3 Dimensional Variational 41 4DVAR 4 Dimensional Variational 42 ACARS Aircraft Communication Addressing and Reporting 12 System ADA After Data Addition 58 ADD After Data Deletion 58 AIREP Manual Aircraft Reports 12 AllBuoy ErnPACRBS and CenPAC_RBS combined 62 AMDAR Aircraft Meteorological Data Relay 12 ARC Arctic RBS atmosphere 61 ASAP Automated Aerological Shipboard Program 22 ATOVS Advanced TIROS Operational Vertical Sounder 12 BC British Columbia 19 BEN Benchmark Atmosphere 60 BUOY Buoy data 12 CenPAC_RBS Central Pacific RBS array 60 CFCAS Canadian Foundation for Climate and Atmospheric 29 Sciences COR Spatial Correlation 67 DA Data Assimilation 2 DEW Defense Early Warning 30 x x n i DIF Arithmetic difference between the verification and the 67 forecast fields ECMWF European Centre for Medium-range Weather Forecasts 11 EDAS Eta Data Analysis System 53 EFG Eta First Guess 60 ENA Eastern North America 2 ErnPACRBS Eastern Pacific RBS array 60 ETA12 ETA Reference Forecast starting at 12Z 55 ETA00 ETA Reference Forecast starting at 00Z 57 ETAXX ETA REF12 and REF00 Forecasts 57 ETA104 Eta 104 Data Product 44 ETKF Ensemble Transform Kalman Filter 26 FDD A Four Dimensional Data Assimilation 7 GAIN Fractional Difference between REF12 and OSSE 68 RMSD values relative to BEN GARP Global Atmospheric Research Program 2 GDPFS Global Data Processing and Forecasting Systems 3 GFS Global Forecast System 53 GOES Geostationary Operational Environmental Satellite 86 GMS Geostationary Meteorological Satellite 86 GOS Global Observing System 2 GPS Global Positioning System 25 GRIB GRIdded Binary 54 GrADS Gridded Analysis and Display System 55 GTS Global Telecommunication System 2 HT70 Height of the 70 kPa surface 3 xxiv ISO International Standards Organization 33 M E T A R Meteorological Aviation Reports 12 M S L P Mean Sea Level Pressure 3 M M 5 Mesoscale Model (Generation) 5 3 M S C Meteorological Service of Canada 12 N C A R National Centers for Atmospheric Research 3 N C E P National Center for Environmental Protection 8 N G M Nested Grid Model 54 NMC(s) National Meteorological Centre(s) 3 N O A A National Oceanic and Atmospheric Administration 86 N R L US Naval Research Laboratory 26 N W P Numerical Weather Prediction 1 O l Optimal Interpretation 40 OSE(s) Observing System Experiment(s) 45 OSSE(s) Observing System Simulation Experiment(s) 2 P A C _ R B S Pacific R B S atmosphere 61 P A C T A R Targeted R B S atmosphere 63 P A C v o i d Pacific data-void atmosphere with N A radiosondes 61 P A C v o i d _ R B S A merging of the above two (PAC R B S and P A C v o i d R B S 61 P I L O T Pilot balloon observation 12 P B L Planetary Boundary Level 49 P T H Pressure, Temperature and Humidity 146 P R O F I L E R Radar wind observation 12 R A O B Radiosonde Observations 12 XXV REF12 MM5 Reference Forecast starting at 12Z 57 REFOO MM5 Reference Forecast starting at OOZ 57 REFXX MM5 REF 12 and REFOO Forecasts 57 RBS RocketsondeBuoy System 2 RMSD Root Mean Square Difference 67 RSMC Regional Specialized Meteorological Centres 3 ROI Radius of Influence 40 SATOB Satellite Observations 12 SCAT Scatterometer winds 12 SCM Successive Correction Method 40 SHIP Ship Reports 12 SSMI Special Sensor Microwave/Imager 12 SYNOP Synoptic Reports 11 THORPEX THe Observing System Research and Predictability 2 Experiment TOST(s) THORpex Observing System Test(s) 24 TReC(s) THORpex Regional field Campaigns 24 TTD Virtual temperature and humidity observations 115 UAV(s) Unmanned Aerial Vehicle(s) 22 UBC University of British Columbia 4 US United States 11 VER Verification atmosphere 58 VEReast Eastern North America verification area 66 VERwest Western North America verification area 66 vRAOB(s) Virtual Radiosonde Observation(s) 58 xxvi vROB(s) Virtual RBS Observation(s) 58 WMC(s) World Meteorological Centres 5 WMO World Meteorological Organization 5 WMOSA WMO space activities 5 WNA Western North America 1 WNAall Atmosphere where all western North America 65 radiosondes are available WNARBS Western North America RBS atmosphere 61 WNA_void Western North America data-void atmosphere with 61 radiosondes over eastern North America WNAvoid_RBS A merging of the above two (WNA JRBS and 61 WNAvoid_RBS) WND Vertual wind observations 115 WS70 Wind speeds on the 70 kPa surface 3 WSR Winter Storm Reconnaissance 23 WWW World Weather Watch 5 xxvii List of Abbreviations % deg hr(s) hor K k m m mm N obs Pa sec sfc ua ver vs, vs. degree percent degree hour(s) horizontal Kelvin kilometer metre millimeter north observations Pascal second surface upper-air vertical versus Acknowledgements I am indebted to Professor Roland B. Stull, my Ph.D. supervisor for providing me with the opportunity to enroll in the Ph.D. program. I also am grateful for his support, advice and help with the preparation of this manuscript. I thank my Ph.D. committee members; Professor William Hsieh, Earth and Ocean Sciences, Professor Rich Pawlowicz, Earth and Ocean Sciences and Professor Sheldon Green, Mechanical Engineering for their review and support of my work. I am appreciative to my fellow students in the Atmospheric Science Weather Prediction Research Team, and all the faculty and staff in the Department of Earth and Ocean Sciences for their help, encouragement and friendship during my study period here. I gratefully acknowledge funding support from the Canadian Foundation for Climate and Atmospheric Science, the Canadian Natural Science and Engineering Research Council, Environment Canada, the Cooperative Institute for Coastal and Mountain Meteorology and Hydrology, and the B C Ministry of Water Land and Air Protection. I thank Rolf Langland, Me l Shapiro, and A . J . Thorpe for their endorsement of the Rocketsonde project within THORpex. The OSSEs were performed on the computers of the Geophysical Disaster Computational Fluid Dynamics Centre, created with funds from the Canadian Foundation for Innovation, the B C Knowledge Development Fund, and the University of British Columbia. The C3 Canadian Computer Consortium partially supported analysts Henryk Modzelewski and George Hicks II. Yongmei Zhou ported the M M 5 model to the computers. Rockets were designed and built by Mark Stull, Catherine Readyhough, Sheldon Green, Gary Schajer, and Jenn Mundy. Ronald Shellhorn of the Vaisala Corp contributed sondes and loaned receivers. Mike Dennett of Cesaroni xxix Technologies Inc. provided rocket-motor recommendations. xxx Chapter 1 Introduction For readers not familiar with operational forecasting, Chapters 1 to 6 contain an overview of the weather observation requirements for numerical weather prediction. Original research results start in Chapter 7. During the 1960s subjective weather-forecast methods gradually gave way to numerical weather prediction (NWP) methods, where finite-difference approximations to the equations of atmospheric fluid dynamics are solved on computers. The observed . weather, as analyzed onto a regular array of grid points, is a required initial condition allowing solution of these equations. However, the atmospheric equations are highly nonlinear, which means that forecast skill is sensitively dependent on initial conditions (Lorenz 1975, 1993). Forecast skill is best for short-term forecasts and falls off rapidly after that. To further the accuracy of N W P within the near-term, more real-time data are required to provide better initialization. Chapter 2 presents an overview of the current meteorological observation and forecasting system. Chapter 3 details the current meteorological data coverage, problem areas and mitigation efforts. Providing initial upper-air weather-observation values (i.e., soundings) within the large "data-void" areas lacking the density and type of data required by N W P is one of the most challenging problems confronting N W P progress. The Pacific Ocean 1 is one such "data-void" area. Day-1 and day-2 forecasts over western North America (WNA) and day-2 to day-5 forecasts over eastern North America (ENA) routinely depend on 1 References to the Pacific Ocean should be interpreted as the Northern Hemisphere Pacific Ocean. 1 observations over the eastern and central Pacific Ocean. The lack of upper-air, in-situ data over this area can drastically reduce the practical limit of N W P predictability to even less than day-1, rendering the huge operational N W P effort economically questionable. The northern-hemisphere polar area is another data-void area. Winter snowstorms with origins in the northern North America can sometimes paralyze populated areas of E N A within the day-2 forecast period. Miscasting can have large societal impacts, even if the storms do not materialize. A new Global Atmospheric Research Program (GARP) called T H O R P E X (The Observing-system Research and Predictability Experiment) is underway and designed to improve observing systems, data-assimilation (DA) and operational N W P (Shapiro and Thorpe 2002, 2004). A Rocketsonde Buoy System (RBS) is proposed under the T H O R P E X umbrella to acquire atmospheric sounding data in the oceanic data-void areas. The R B S design and the development progress is contained in Chapter 4. The concepts of real-time management and flexible rocketsonde payload are advanced. A n RBS-equivalent may also evolve for the polar data-void, where manned radiosonde site costs are extremely high. The goal of this research is to determine the likely impact of such an R B S and to aid in its design. Observing System Simulation Experiments (OSSEs) are conducted to test the impacts of virtual R B S data on D A and N W P . The experimental environment, issues and concerns are introduced in Chapter 5. This chapter provides general background information on N W P , D A and regional modeling. More specific modeling information is found in Chapter 6. The National Center for Atmospheric Research (NCAR) M M 5 2 2 Mesoscale Model Generation 5. 2 . mesoscale modeling system used here is described. The description includes M M 5 numerics, physics, and data assimilation methods. O S S E methodology is detailed in Chapter 7. The reference, benchmark and O S S E atmosphere constructs are described along with the reasoning behind the experiments. Verification areas representing the forecast results over W N A and E N A are chosen. Verification parameters used include mean-se-level-pressure (MSLP), height of the 70 kPa surface (HT70) and wind speeds on the 70 kPa surface (WS70). Improvement indices include a measure relative to a perfect forecast and another relative to a benchmark forecast. The primary objective of this research is to support the feasibility, design, development, deployment, operation and cost-effectiveness of the RJ3S system. Specific objectives are detailed in Chapter 8. Reference cases were run every day when resources permitted. Cases with some meteorological significance were chosen for the OSSEs. Three cases are described in detail. Case 12-December-2001 describes an intense cyclone moving over the British Columbia southwest coast. Three R B S 6 km profiles located in the eastern Pacific would have resulted in a large improvement in the N W P result. Case 20-February-2002 simulates a data void over western North America and shows the negative effect over eastern North America. Case 15-October-2003 describes a heavy rain episode and shows the improvement that R B S profiles can make to the timing and intensity of the rainfall forecasts. This case is later used for studying the benefits of RBS real-time management (Chapter 9). General findings are reported in Chapter 9. A study of the reference forecasts (without R B S data addition) shows that the W N A N W P penalty paid by the Pacific data void is 20%. N W P forecasts starting at OOZ deliver better results than the forecasts starting 3 at 12Z. N W P forecast errors for winter cases are 18-25% greater than summer cases. OSSEs show that optimum R B S profile heights are 6 km. At times a three R B S array can have a substantial beneficial impact on N W P results. Routinely, a six-to-twelve R B S array is needed for consistent improvement. The favored position for a six R B S array is centered on 50N 145W with the units about 300 km apart. A detailed summary of the findings in found in Chapter 10. The summary is followed by a critique of the O S S E method. The major weakness is that the RBS profile data is taken from a model that may not resemble real world data. A guideline for any successor research follows. The chapter ends with a glimpse into the possible future and what may be weather observations in the real world. The Appendices provide supporting information. Appendix A includes a summary of the sonde sensors currently available. Appendix B describes the University of British Columbia (UBC) computer resources that were available. Appendix C lists the cases used for various experiments. 4 Chapter 2 A Systems Overview The RBS is designed to fill a gap in worldwide meteorological upper-air in-situ data coverage. To understand its significance, one must first understand how international data gathering works. 2.1 Global The World Meteorological Organization (WMO), through its World Weather Watch (WWW), collects, analyses and distributes meteorological, oceanographic and other environmental information throughout the world. It is comprised of three basic components; the global observing, telecommunication and data processing systems (Vargas 2004). Under international agreements, all countries are responsible for meteorological observations within their borders, and contribute these to the Global Observing System (GOS). Each WMO country is committed to have (or is supposed to have) a national meteorological service. The observations are carried out regularly, simultaneously, and transmitted immediately to the Global Telecommunication System (GTS). These national meteorological services are obligated to provide the real-time data for international usage and are entitled to use the information produced by other countries. Oceanographic information is increasingly made available in a similar manner and through marine reporting systems. International outer-space treaties guarantee accessibility over any part of the Earth's surface for any countries' satellite program. A small subset of WMO members has satellite 5 operations that provide raw and processed satellite data of greatly varying type, spatial and temporal resolution. The W M O space activities group ( W M O S A ) coordinates the space programs and data usage of member countries. The W M O expects these satellite data to be made available to all and promotes programs to encourage maximum usage. The G T S is the backbone carrier that interconnects meteorological telecommunication centers by a combination of terrestrial and satellite telecommunication links. Finally, the Global Data Processing and Forecasting Systems (GDPFS) prepares meteorological analyses and forecast products and makes the products available to W M O members. G D P F S is organized as a three-level system of: World Meterological Centres (WMCs), Regional Specialized Meteorological Centres (RSMCs) and National Meteorological Centres (NMCs). These centers carry out GDPFS functions at the global, regional and national levels, respectively. 2.2 Observations The G O S is the provider of in-situ atmospheric observations such as wind direction and speed, visibility, temperature, humidity, solar radiation, cloudiness, cloud base, pressure, precipitation and oceanographic observations such as sea-state and sea-surface-temperature. The W M O S A provides radiance data measured by satellite-deployed instruments. Many countries have their own radar networks that produce reflectivity data related to precipitation intensity and Doppler observations of related wind components. Details are found in Chapter 3. 6 2.3 Four Dimensional Data Assimilation Receipt of the observations is only the first step in the overall process. After the data arrives at one of the operational centers, the data has to be assimilated into a weather analysis. There are varying degrees of assimilation and various procedures to do so. D A modeling is the process of finding the model state that is most consistent with all available information including the observations. Traditionally, data were assimilated subjectively by simply studying reports or constructing hand-analyses. Today, most D A methods are based on computer models providing a three dimensional representation of the earth-atmosphere system. The results are commonly referred to as an objective analysis1. A n N W P model provides an estimate of the earth-atmosphere's evolution with time. This enables the use of observations distributed in time. The analysis cycle implemented in global operational centers is a real-time driven intermittent data assimilation system termed four-dimensional D A (FDDA). A forecast model makes a prediction for a very short duration, the forecast is stopped, the new observations are assimilated, the forecast continues and the cycle repeats. The N W P model organizes and propagates forward in time the information from previous observations. New observations are used to modify the model state while being as consistent as possible with the previous information and the model physics. Since all information is being represented within the model, it is important that the model should be of sufficiently high resolution and able to embody the information contained in the observations. 1 Hereafter, a reference to an analysis will mean an objective analysis. 7 2.4 Forecasts After the data are assimilated, the resulting analysis is used to provide a starting point for a projection into the future. N W P is now the most common and powerful method to provide this projection. N W P is the process of integrating in time the atmospheric dynamic and thermodynamic governing equations. The resulting forecast products are provided as graphics (weather maps and meteograms), tables of data, and digital electronic files (such as the ETA104 data product provided by the National Centers for Environmental Prediction: NCEP) . One of the major disadvantages of N W P is the large and exacting data requirements needed for initialization. Under many conditions N W P is not accurate, is beset with various deficiencies and often needs to be augmented or supplemented by other techniques. Those other techniques include subjective forecasts, short-term nowcasts using radar and satellite data, climatological and analog (map-typing) methods, and statistical methods including linear (Kalman-filter) and nonlinear (neural-net) processing. Another method to improve N W P is ensemble forecasting, which produces several N W P forecasts from multiple models or multiple initial conditions. The initial conditions can be perturbed for each ensemble member forecast run to represent the "uncertainties" and "errors" inherent in the observations. If all model runs produce a similar forecast, then a forecaster can have greater confidence in the model prediction. If the forecasts vary considerably, then the confidence will be less. Unfortunately, if the initial-condition perturbations are made in a data-void area, the error in the initial field can be greater than the reasonable perturbations. A l l the ensemble-member solutions may not provide results close to that later evolved by the atmosphere. A combination of ensemble and statistical 8 forecasting are used in many numerical operational centers to predetermine sensitive areas in the data-voids. At the local forecast offices, most forecast techniques are "hybrid". Objective aids combine elements of N W P from the national centers with one or more of the non-NWP methods. The elements are blended using weights based on their past performance, and manual refinements can be made by the experienced forecasters. This is especially true where the N W P output can be least trusted, as in and downstream of the data-void areas. 2.5 Technology Great advances in technology have driven breakthroughs in many fields of science, including weather instruments and N W P . These advances include the miniaturization of electronic sensing systems (Appendix A) used for in-situ weather observations, greatly facilitating their deployment and increasing their cost-effectiveness. The rapidly expanding number of observations from remotely sensed land-based and satellite systems are adding greatly to the observational coverage of the earth. The exponential increase in the speed and memory of the computer systems have resulted in large advances in the communications that support all aspects of the forecast system. With no limit in sight, accelerating technology provides the opportunity for further advances and the possibility of unforeseen breakthroughs. The R B S (Chapter 4) takes advantage of some of these advances. 2.6 Mesoscale Modeling One such N W P breakthrough is the advent of mesoscale modeling. Because of the tremendous amount of resources required to develop and operate a N W P model, real-time 9 N W P used to be the sole privilege of a few major operational centers (like the NMCs) . In the most recent decade, mesoscale modeling has experienced phenomenal growth. This transition has occurred because of the recent availability of affordable high-performance computers, community models (such as the M M 5 ) and the real-time accessibility of data products from the N M C s . Such mesoscale models (Chapter 6) are an important tool for the research presented here. 2.7 Summary Where there is sufficient data, computer based D A modeling and N W P provide the best forecast results, and is the hope for continued improvements in forecast skill. However, in and downstream of data-void areas N W P may not provide the best short-term results. The obvious solution is to fill the data-voids. Unfortunately, this obvious choice comes at increased cost. Advancing technology offers cost-effective solutions such as the RBS. 10 Chapter 3 Data Coverage 3.1 The Current Observational Network Data are carried by the G T S to the N M C s , including the European Centre for Medium-range Weather Forecasts ( E C M W F ) in Reading, U . K . , and the N C E P sites headquartered at Camp Springs, United States (US). Large amounts of data reach E C M W F and are compiled, mapped and made available at their Web site ( E C M W F 2005). Similar amounts reach N C E P . Observation Code Type Level Profile Number SYNOP in-situ sfc hor 14303 SHIP in-situ sfc hor 1583 M E T A R in-situ sfc hor 9640 drifter B U O Y in-situ sfc hor 4818 moored B U O Y in-situ sfc hor 151 aircraft-AIREP in-situ ua hor/ver 7037 aircraft-AMD A R in-situ ua hor/ver 14457 aircraft-ACARS in-situ ua hor/ver 28945 RAOB-land in-situ ua ver 600 RAOB-ship in-situ ua ver 7 PILOT in-situ ua ver 273 PROFILER remote ua ver 197 SSMI remote sfc hor 16579 A T O V S remote u a Ver 243747 SATOB remote ua hor 143025 S C A T remote sfc hor 48795 Table 3.1. For 4 February 2005 at 00Z, the number of worldwide observations available at ECMWF. The first column is the code name. The second column categorizes the data as in-situ or remotely sensed. The third column states whether the data is surface (sfc) or upper air (ua). The fourth column indicates if the data is valid at a single level (hor) or at multiple levels (ver). The last column has the volume of data available at the 00Z synoptic hour. This information is taken from http://www.ecmwf.int/products/forecasts/d/charts/monitoring/coverage/dcover/. For example, data available for assimilation for 4 February 2004 at 00Z are shown in Table 3.1. Synoptic reports (SYNOP) are surface observations over land. Ship reports 11 (SHIP) are volunteer surface observations from ships of opportunity. Scheduled airport observations ( M E T A R ) are taken at the end of each hour. Buoy data ( B U O Y ) are automatic surface reports from buoys. Aircraft reports are upper-air observations from commercial aircraft mostly at cruising altitudes, but with soundings near airports. ATREP are manual aircraft reports; A C A R S and A M D A R reports are automatically generated. Radiosonde (RAOB) data are upper-air soundings made with balloon-borne instruments. P I L O T / P R O F I L E R data are locally generated upper-air winds. A T O V S data are satellite-derived temperatures. SSMI data are ocean-surface wind speeds from satellite deployed passive microwave radiometers. S C A T data are ocean surface winds from satellite-deployed scatterometers. S A T O B data are upper-air winds obtained from tracking cloud drifts obtained from satellite-deployed radiometer radiance measurements. Examination of the data amounts shown on Table 3.1 shows that there are a small amount of in-situ observations relative to the large amount of data derived from satellite/radar radiances. The number of in-situ observations reaching the N M C s is about 104 for each D A , totaling to 105 daily, of which the most critical are the vertical soundings. Unfortunately, the numbers of some of these soundings have decreased. In 1988 the number of R A O B s and PILOTs daily was 1660 and 964, respectively. B y 1999 the daily numbers were down to 1209 R A O B s and 576 PILOTs (Dabberdt and Shellhorn 2003). This trend seems to be continuing. The Meteorological Service of Canada (MSC) is considering closure of the Whitehorse, Yukon radiosonde station as a cost cutting move 1. The N M C s have to rely more on satellite-derived information and satellite radiances (discussed below). 1 This information was obtained from a private communication with a MSC manager. 12 3.2 The Unbalanced Observational Network Examination of the data-coverage graphics posted at the E C M W F Web site reveals other undesirable features of the observing system. Intuition would suggest that in three-dimensional space, the distribution of new data and observations should be as uniform as possible. Unfortunately, this desirable feature is not true. Most of the in-situ data are over the land areas and most are clustered in certain geographical areas, like the US , western Europe and eastern Asia. There are many large areas where there is a paucity of observations and others where there are none. These are the in-situ data-void areas3. The E C M W F data coverage charts show that the data-void areas near North America and Europe include the Pacific Ocean, polar areas, Atlantic Ocean and the Gulf of Mexico. Over the Pacific Ocean, the range of coverage starts with no R A O B s to irregular amounts of SHIP, B U O Y , S C A T , SSMI, A T O V S , S A T O B , and AIREP. Over the polar areas, the range of coverage starts with no R A O B s , a few SYNOPs, and a few AIREPs. Unfortunately, satellite-derived data ( A T O V S , S A T O B , S C A T , and SSMI) have high errors delivering low impact. A T O V S retrieval algorithms require complimentary in-situ measurements that currently do not exist. It is difficult to determine the valid level for cloud-drift winds (SATOB). S C A T and SSMI algorithms are based on assumptions that are only approximately valid and do not work within precipitation areas. At the same time there is a large degree of redundancy in the in-situ observing systems over the US, western Europe and eastern Asia. These data-rich areas have a regular network of R A O B s and somewhat redundant number of PILOTs and AIREPs. 2 A reference to the United States means the contiguous 48 states of the U S A . Alaska and Hawaii wi l l be referred to separately. 3 A reference to a data-void area will mean an in-situ data-void area. It is recognized that satellite radiance data covers the entire globe. 13 Moreover, the AIREPs are of vertical profiles rather than the much more limited oceanic and polar flight-level readings. It is possible that much of this extra data results in "data saturation" and adds little value to an objective analysis. Ideally, if concentrations of in-situ observations exist, they should be within active areas such as storm tracks of mid-latitude and tropical cyclones, preferably in the areas where rapid changes are occurring. The current observational network does not satisfy this desirable attribute. Dense coverage of satellite radiance data over data-void areas does not help much. Satellite radiance data does not deliver the type of information most critically needed for N W P . At present, satellite-borne instruments have not provided adequate data for optimum initialization of N W P models. B y themselves, satellite-borne instruments are not likely candidates to fill the data void. Radar and wind-profiler data, extending over the immediate offshore areas, are of local benefit, but of limited range. These systems are not likely candidates to fill the data void. 3.3 Observations Within The Domain of Interest The domain that will be used for this study is shown in Figure 3.1. It can be divided into roughly two areas, the data-rich US and southern Canada areas, and the data-void Pacific and Arctic areas. Over the Pacific region, the real data void is at the mid-troposphere due to the lack of R A O B s (Figure 3.2a). It is the weakest part of the observational network. Figure 3.2b shows the Pacific data void as a mid-tropospheric void of in-situ data sandwiched between relatively data-rich layers near the surface and near the tropopause. Mid-tropospheric A T O V S and S A T O B data is of variable coverage and accuracy. No real-time ocean data is 14 available below the ocean surface, handicapping the coupled ocean-atmospheric models. The coverage over the polar area is somewhat analogous with a few S Y N O P replacing the S H I P / B U O Y reports. N W P within this domain needs in-situ profile information similar to R A O B s . Coupled ocean-atmospheric models need real-time information below the ocean surface. Observing systems providing this profile information are critically needed. 180w 170w 160w 120w BOw 70w 60w MOw 130w 120w llow 100w Figure 3.1. Data sets available from the existing observing systems, ordered approximately by their valid level within a data assimilation. Over the US, the existing radiosonde network averages about 315 km spacing between sites (Dabberdt and Shellhorn 2003). AIREP and PILOT data (Table 3.1) supplement the R A O B s . The same radiosonde coverage is required over the Pacific Ocean and polar areas. Demonstrating this requirement is one of the motivations of this study. 15 180"W ISO W 120"W 9Q*W 12 8 z {km) 4 0 180"W 150"W 120"W Figure 3.2. (a) Radiosonde sounding locations (dots) over North America. The region of mid-tropospheric in-situ sounding-data paucity, identified as the "Pacific data void", is (b) sandwiched between data-rich layers of surface observations and near-tropopause observations. Typical numbers of in-situ observations per analysis period is given for each region. AIREP = Aircraft Report (manual); AMDAR = Aircraft Meteorological Data Reporting Relay; and ACARS = Aircraft Communications Addressing and Reporting System. Satellite-derived data (SCAT, SATOB, ATOVS) is of variable coverage, density and frequency. The states/provinces marking the eastern edge of the data void are: A K = Alaska, BC = British Columbia, CA = California, OR = Oregon, and WA = Washington. 3.4 Impact of Data-void Areas Under normal near-zonal flow conditions, the most important data-void area for North America is the Pacific Ocean, because the storm-tracks are oriented from west-to-east (Figure 3.3a). The negative impacts of the data void travel to the east along with the storms, reaching W N A by day-1 and E N A on day-2. 16 J b ) Observation "Sandwich" , 1 Aircraft ob s . (AIREP, A C A R S , AMDAR) (order of 250 / anal.) P - J 30 ... Pacific Data Void - A few SATOB A few A T O V S • SCAT | v — - y V y — y V — — i ~ ^a^^afi^ H t f i 1 1 1 1 ! ,1 t 1 1 i msa Sometimes the normal zonal circulation gives way to a more meridional circulation (Figure 3.3b). A wintertime ridge sets up over W N A and a cold flow from arctic North America moves over E N A , resulting in a severe snowstorm regime. One example season is the winter of 2002/03. At least 14 eastern North American snowstorms had their origin in the arctic North America. Under these conditions, the polar data void may be the most sensitive upstream area for E N A , where the negative effects of the data void are felt on day-2. The winters of 2003/04 and 2004/05 have experienced similar occurrences. Over data-rich regions such as US, western Europe and eastern Asia, the analysis is dominated by the information contained in the in-situ observations. In some in-situ data-poor regions such as the Atlantic Ocean, the forecast benefits from the information upstream; the model forecast is able to transport information from in-situ data-rich U S to much of Atlantic data void before the value is dispersed. Unfortunately because of the vast expanse, the central and eastern Pacific does not have the benefit of either. The benefit of the data from eastern Asia is dispersed before the flow reaches the central and eastern Pacific. A n F D D A system can then act in a detrimental way. Pacific weather disturbances often deepen rapidly just upstream from W N A . D A quality control algorithms have acceptance criteria that call for data to be within a certain range of the previous forecast providing the first-guess field. A previous poor forecast can trigger a rejection of good data over W N A (Figure 3.4). This rejection renders the output from very expensive operations virtually useless, and moves the effective data void eastward. This loss could be prevented i f some reliable atmospheric soundings were available upstream from the coast. Even day-17 1 to day-2 forecasts can possess major errors ("busts") due to the maintenance and growth of substantial initialization error (Doyle 2001, Hacker et al. 2003). 140W 130W 120W HOW 100W Figure 3.3. (a) A normal zonal flow at 12Z, 9 November 2002. Weather disturbances over the Pacific data void affects the WNA first and ENA later, (b) A meridional flow at 12Z January 2003. Weather disturbances over the eastern and central Pacific move to the Alaska areas not affecting WNA and ENA. Weather disturbances originating in the arctic data void affect ENA later. 18 14 December 2001, OOZ 16 October 2004, OOZ Figure 3.4. (left) Data rejections (circled stations) during the windstorm of 14 December 2001, and (right) during the extreme rainfall event of 16 October 2004 are shown. Both cases are discussed in Section 8.4. 3.5 Costs to Society Extreme weather events are virtually a daily occurrence somewhere in the world, and can result in huge costs to society (Pielke and Kimpel 1997). Wong (2004) documented 2003/04 "wild weather" events affecting British Columbia (BC). A n extreme rainfall event caused floods during the week of 16-23 October 2003 (More details are available in Section 8.4). Roads and railway bridges were washed out; hydro-dam operators had to spill furiously, discarding valuable hydro resources. On 28 October 2003, a windstorm disrupted power on Vancouver Island for 110,000 residents. Heavy snow falling on 23 November 2003 over Vancouver Island disrupted transportation and caused serious car and truck accidents. On the 28 t h of the same month, there was more flooding over southern B C . Windstorms during the first two weeks of December 2003 resulted in deaths. During 28-30 January 2004, snow and wind caused massive power outages and 19 resulted in abandoned vehicles over southern B C . Winter still raged on during 18-19 March 2004 with more wind and snow resulting in power outages, drydock damage, overturned boats and highway closures. On 27 April 2004 B C experienced a severe windstorm; 50,000 B C residents experienced power outages as the 85-95 krn/hr winds shredded trees and downed power lines. Nothing can be done to prevent these extreme events, but adequate warnings can mitigate much of the associated loss. Unfortunately, the opposite seems to be true i f the weather does not materialize in accordance with the warning. Extreme weather events are often preceded by major forecast failures. The M S C heavy snow forecast of 6 January 2004 was good with its timing but poor with the predicted amount; the "heavy snow" portion did not materialize. High-profile busts perpetuate damage far beyond the immediate event; these busts can severely reduce public confidence in all weather products and services. Severe weather events and forecast busts capture the media and public attention, but it is the provision of routine weather information that makes a large contribution to economic activity. In this era of environmental management, the emphasis is on the more efficient use of available natural resources, rather than their expanded exploitation. Reliable weather information can support efficient environmental management. For example, fire-weather models are used to support groups controlling forest-fire damage. These high-resolution mesoscale models need larger-domain forecast data for their boundary conditions. But these larger domain models are the same ones experiencing the data-void initialization errors. 20 As our society becomes increasingly sensitive to its environment, fewer new dams for the production of hydroelectric energy are being constructed. Hydro companies are now emphasizing reservoir planning and consumer electrical efficiency. High-resolution hydro-meteorological models suffer from the same uncertain boundary conditions. Electricity and gas providers must forecast their requirements in order to provide efficient delivery. These estimates are heavily dependent on temperature and sunshine-hour forecasts. Commercial groups involved in the development of Pacific-region offshore resources need accurate site-specific wind-wave and visibility forecasts. Emergency-response groups need detailed site-specific data over their area of interest. The cumulative economic cost resulting from this poor support can be enormous, and must be mitigated. 3.6 Previous Data-void Mitigation Efforts The Pacific data-void problem has seen the rise and demise of profile-type observing systems. Weatherships were an outgrowth of WWII support programs. In 1966 and 1967, two very advanced Canadian weather ships were put into service in the Pacific Ocean at station P (145W, 50N). During these early years of N W P , Station P radiosonde data became a critical "anchor" for the Pacific analyses. Even though the Station P data requirement had not changed, the ships were decommissioned in 1981. The Canadian government became convinced that satellite sensing and buoy observing would be a superior replacement technology. Unfortunately, to this day neither has reached their promised potential. Forecast skill decreased when station P observations disappeared (Spagnol etal. 1980). 21 In the 1980s, a trial Pacific Automated Aerological Shipboard Program (ASAP), delivering radiosonde profile-type observations from volunteer merchant ships was operating in the Pacific Ocean. Unfortunately, A S A P was deemed non cost-effective in the Pacific Ocean due to sympathetic data-denial (i.e., ships avoiding regions of forecast storms). A n analogous A S A P program still operates in the Atlantic Ocean. Because of the shorter distances, Atlantic Ocean merchant ships generally do not try to avoid most mid-latitude storms, resulting in better coverage within the storm tracks. The (few) Atlantic A S A P radiosonde sounding sites can be seen on the posted E C M W F Web graphics. Pailleux et al. (1998) studied the impact of the Atlantic A S A P soundings. They reported that a small number of North Atlantic radiosondes could have a large impact on N W P results. The Atlantic A S A P program will likely be expanded (Shapiro & Thorpe, 2004). Unfortunately, there does not seem to be any interest in reviving the Pacific A S A P program, even though it is needed more in the Pacific than the Atlantic. 3.7 New In-situ Observing Systems To improve numerical-forecast skill of high-impact weather events, T H O R P E X (Langland et al. 2002) is encouraging development of several candidate in-situ systems: driftsonde, aerosonde, smart balloons, dropsondes, commercial aircraft and rocketsondes. The driftsonde would be launched upstream from the data void (e.g., Japan), travel over the data void for its approximate operational period, and deploy the dropsondes at scheduled times and locations or on-demand. Unfortunately, driftsonde trajectories are somewhat unpredictable, uncontrollable and will flow with the prevailing winds. The vertical profiles may take place where they are least needed. There is also the question of 22 the possible unwanted intrusion into a sovereign airspace. Finally, dropsondes from driftsondes are larger than standard radiosondes and have been perceived to be a hazard to commercial aviation. Consequently, an operational driftsonde program may never happen due to safety reasons, not technology limitations. Aerosondes are unmanned airborne vechicles (UAVs) that will be launched from BC, Hawaii or Alaska and gather flight level data within the 4-6 km altitude range. A design innovation may soon make it possible to miniaturize and deploy dropsondes on aerosondes. In theory aerosondes are controllable, but they are prone to icing problems and may become uncontrollable in high wind, turbulent atmospheric conditions. Aerosondes are best deployed in the periphery of high-impact areas. An operational aerosonde program is feasible and may happen. Smart balloons could provide high temporal observations of temperature, pressure, and humidity in the lower troposphere. The problem with the balloon system is the lack of external control. The individual balloons are advected by the winds. In some areas this could lead to a considerable amount of grouping, data duplication, data-sparse regions and ultimately an ineffective observing system. An operational smart-balloon program is feasible and may happen. Dropsonde programs use manned aircraft to deploy dropsondes in sensitive regions. These programs are very expensive. If the targeted areas are not well chosen, there can be a large expense with little benefit. The other problems are the long lead-time (up to two days) and the requirement to use a predetermined flight route. More discussion is found in Section 3.10. The NCEP Winter Storm Reconnaissance (WSR) dropsonde program has been operational since 1999 (Szunyogh et al. 2000, Szunyogh et al. 2002). 23 Commercial aircraft will not deploy dropsondes because the vertical trajectory will occur through lower-level commercial flight lanes and potentially produce a hazard to in-flight aircraft.. Rocketsondes are another potential dropsonde delivery system. A rocketsonde can be launched from a surface-based platform and deploy a dropsonde at its apogee (Chapter 4). The potential of rocketsonde technology has been recognized by T H O R P E X . The potential issues with rocksonde operations could be the requirement for a suitable launching platform and perceived rocket-related hazards. The launching platform could be ships, land, sea-ice (Shapiro and Thorpe 2004) or buoys (Spagnol and Stull 2003). The optimum mix of observing systems will probably consist of the above noted technologies working together with the remote sensing systems discussed below. T H O R P E X will conduct TOSTs ( T H O R P E X Observing System Tests) to evaluate the operational performance and forecast impacts of these observing systems. T H O R P E X will also carry out T H O R P E X Regional field Campaigns (TReCs), which are evaluations of the interactive component of these systems. The final acceptance of the R B S system could be based on its performance and cost relative to these other candidate systems. In the aftermath of the Pacific tsunami event in December 20044, additional tsunami-warning buoys are planned for Pacific deployment, which could also serve as cost-saving platforms for rocketsondes. Research and experimentation with real and idealized observations must precede development and deployment. The research documented by this dissertation constitutes an initial portion of the R B S effort. 4 In the following months, the tsunami was followed for two earthquakes affecting the same region. 24 3.8 Satellite Observing Systems and GPS Meteorology For observing-system development, significant resources continue to be spent on satellite-deployed instruments. In spite of much evidence to contrary, the perception remains among politicians and resource managers that satellite technology will provide the final solution to all data-void problems; Among the meteorological research community, it is recognized that satellites by themselves likely will not provide the total answer. Reasons include (1) radiance observations admit an infinite set of possible temperature soundings, a failing that can be resolved only if supplementary data (first-guess soundings or cost functions using N W P forecasts) are included: and (2) upper-air winds can be found only at a limited number of poorly-estimated altitudes. Pailleux et al. (1998) reported that the impact of satellite observations is much smaller than that of radiosonde observations. Graham et al. (2001) reached the same conclusion. Since then, anticipated new satellite sensors have renewed the hope of greater satellite data impact. This new generation of satellite deployed infrared sounders will increase the spectral resolution by several orders of magnitude and the number of satellite observations available to operational N W P will increase greatly. Unfortunately, the problems with the applicability to N W P will remain, as do problems with data inversion and non-independence of neighboring soundings. Satellite instruments have difficulty measuring the atmosphere below clouds (McNally 2000), so current forecast systems utilize only cloud-free clear-column radiances, and cloud-top winds. Satellite sensors cannot deliver much information below the ocean surface. Because of the variable emissivity over ground, water, ice and snow, current forecast systems do not use radiances over the polar data void areas. New variational D A 25 techniques (Section 5.4) for satellite radiances are in the early stages of development. Early positive results have been recorded, but further future improvements remain uncertain. Shapiro and Thorpe (2004) proposed an innovative procedure to obtain thermodynamic observations of the atmosphere by measuring the atmospheric delay of the radio-frequency signals from Global Positioning Satellites (GPS). The GPS system does provide large coverage and the procedure does show promise. However, these observing techniques are relatively new and the results are uncertain. 3.9 Real-Time Management In an attempt to maximize the impact of available resources, there is currently an emphasis on adaptive observing strategies, commonly termed targeting. Targeting implies the use of real-time management of observing systems. Adaptive observation systems are directed to, or managed within, the probable sensitive areas where new observations will have the greatest impact for the downstream forecast area of interest. The first step in this process is to determine the initial-time sensitive area that will impact the area of interest. One method favored by the US Naval Research Laboratory (NRL) is the adjoint method (NRL 2002). This targeting strategy relies on first making an N W P forecast for the area of interest, and then integrating an adjoint model backwards (in time) to determine where the model results are most sensitive to initial conditions. Another method favored by N C E P is the Ensemble Transform Kalman Filter technique (ETKF) , an improvement over basic ensemble forecasting (Section 2.4). The results of ensemble perturbations are weighted by statistical methods to produce the most likely outcome. N C E P uses this procedure during their operational W S R program. The 26 weather events with a large societal impact are identified. Interpretations of the E T K F results constrain the choices for a predetermined flight route for dropsonde aircraft. Aircraft are tasked out of Alaska and Hawaii to make the drops. To produce useful results, the sensitive-area prediction procedure needs a good initial field. O f course, this condition is usually not present over the data voids. Without a good first-guess, the sensitivity prediction procedure may make a wrong prediction, sending the targeted operations to the wrong area. There would be a great expense for little benefit. Ensemble methods can produce better first-guess fields, but they too can have significant errors in data-void regions. Under these conditions, locations of sensitive areas could be better found through hybrid procedures similar to the ones described in Section 2.4. What is needed is a fixed, coarse spacing of in-situ soundings, which allows a better resolution of the larger synoptic regime. This will provide a greater probability that the N W P forecast would be more representative and a better chance that the targeting procedures will be able to locate the sensitive area. Real-time management of many components of the observing system could be used to stop costly operations where the observations will have little impact. Figure 3.3b shows a large amplitude high-pressure area over the W N A coastal areas. R A O B data taken within the core of the high-pressure area probably had negligible impact on the N W P results anywhere. If those preplanned radiosonde launches could be cancelled, resources would be saved with little impact on the final results. RBS real-time management will be discussed in Chapter 4. 27 3.10 Summary The current observation network suffers from the lack of direct in-situ observations in large data-void areas, and possible redundancy in a few data-rich areas. Fixed sets of radiosonde-like profiles are needed in the data-void areas. Development of new in-situ observing systems is endorsed by T H O R P E X . The driftsonde system includes a stratospheric balloon with gondola that deploys a payload of dropsondes. Aerosondes (miniature-robotic aircraft) provide flight-level measurements of meteorological parameters. More buoys provide more surface observations. Smart balloons provide high temporal resolution observations in the lower troposphere. Dropsonde campaigns rely on manned aircraft to deploy the sounding packages. Rocketsondes measure soundings over fixed buoys. Each proposed system has its strengths and deficiencies. It is likely that a mix of in-situ observing and remote sensing systems will eventually provide the optimum solution. Real-time management of many components of the observing system should improve its cost-effectiveness. 28 Chapter 4 Rocketsonde Buoy System 4.1 Rocketsonde Usage Rocketsonde technology may be one of the innovative technologies leading to future in-situ soundings in data-void areas. Rockets with sensors have been used for meteorological research since the early 1960's (Clark 1965), particularly for stratospheric soundings. Currently, the United States Navy routinely uses rocketsonde low-altitude (boundary layer) soundings to calibrate radars and other shipboard instruments. Recently, Shapiro and Thorpe (2004) and Spagnol and Stull (2004) have suggested that medium-altitude rocketsondes could support the earth-observation network. A major issue seems to be the proposed launch platform. Shapiro and Thorpe have suggested ships and sea-ice. Unfortunately, a mobile rocketsonde launch platform would not work in the north Pacific. A ship host program would produce the same sympathetic data-denial effect seen with the Pacific A S A P trial (Section 3.6). A sea-ice platform is not realistic in the north Pacific; there are no large icebergs available. Even i f there were large enough icebergs, the results would have the same deficiency as the driftsonde program. The rocketsonde profiles may not occur where the data are most urgently needed. Spagnol and Stull have suggested that rocketsonde launches could be from fixed-location anchored buoys, providing many advantages. The locations could be chosen to intersect likely storm trajectories, and piggyback on the soon-to-be-deployed tsunami buoys to save money. Fixed locations would allow authorities to design restricted airspace around the locations from the surface up to the rocketsonde apogee. The locations could be 29 added to marine charts, eliminating accidental contact by marine craft. This approach would address the perceived hazards to aviation and marine traffic from rocketsonde launches. 4.2 The Concept The University of British Columbia (UBC) has been promoting the concept of a Rocketsonde Buoy System since McTaggart-Cowan (1998) provided an initial investigation. In 2001, funding for the feasibility study and preliminary development of the R B S was obtained from the Canadian Foundation for Climate and Atmospheric Science (CFCAS) . Detailed information on the project is available at the U B C Rocketsonde Buoy Web site (Spagnol 2005). A n overview will be provided below. The ocean-deployed RBS is designed to support its primary meteorological mission and a secondary oceanographic mission. A n RBS tertiary mission could be as additional sensing stations in a seismic-tsunami warning system. The R B S technology could be adapted for arctic deployment, and may be the only option available to replace the closed radiosonde stations. T H O R P E X has endorsed the concept of land-based rocketsonde stations. For its Pacific meteorological mission, the proposed R B S would launch small, unguided rocketsondes daily to about 6 km apogee from a buoy platform (Figure 4.1). At apogee, the sonde will separate from the rocket, and while descending with parachute will measure and transmit atmospheric temperature, humidity, pressure, and GPS position to get winds and altitude. The sonde sensors are described in Appendix A . Each deep-ocean moored R B S will carry at least 200 rocketsondes, each in their own sealed launch tube. The R B S launch-control system will initiate the launch when the rocking buoy passes close 30 to vertical during the synoptic observation window near 12Z or OOZ. The R B S is being specifically designed to operate in the severe winter-storm conditions of the north Pacific Ocean, including high seas and icing. Deployment will likely be phased as resources become available. A tender ship will replace the launch-tube canister and recondition the buoys during yearly maintenance cruises. For its polar meteorological mission, an array of land based RBS-equivalents1 could ring the northern limit of North America, similar to the decommissioned Distant Early Warning (DEW) line located on the North American Arctic Ocean coast. The R B S launch system will be designed to withstand cold temperatures, snow and icing. For its Pacific oceanographic mission, oceanographic equipment such as sea-surface-temperature and wave-height sensors will be deployed on the buoy structure (Stull et al. 2003). The mooring line will be equipped with devices for measuring temperature and current speeds providing profiles. For the seismic-tsunami mission, ocean-floor pressure sensors could be mounted on or near the anchor. The hundreds of thousands of deaths resulting from the 26 December 2004 Indian Ocean tsunami support the desirability of doing so. The 8 January 2005 edition of the Vancouver Sun featured a front page entitled "what will happen i f a tsunami hits our coast". The expert (Dr. John Clague from Simon Fraser University) predicted waves as high as 15 meters hitting the Vancouver Island coasts by a tsunami triggered by the sudden repositioning of the locked Juan de Fuca and North American tectonic plates. He speculated that in the near future, these plates would suddenly shift, generating a huge offshore earthquake. 1 The arctic RBS-equivalents will be referred to as RBS as well. 31 . To complete the R B S mission, an on-board processor would compile and compress all data and retransmit it to a communications satellite, which will relay the data to the G T S . rocket sealed launch tubes Figure 4.1. A rocketsonde launch. The rocket will carry a sounding package (sonde) to a maximum of 6 to 8 km altitude before the sonde, with parachute, separates from the rocket and falls back to the ocean surface. While falling, the sonde will sense the atmosphere for its ambient temperature, humidity and pressure. The sonde can also receive GPS signals to determine winds and altitudes. The sonde will transmit its data to the buoy. The data stream signal will be amplified and retransmitted to a satellite that will relay the data to shore, where it is added to the GTS and carried to the data users. 32 4.3 Design Objectives The R B S mission statement is: the RBS system will provide operational meteorological, oceanographic and seismic observations reliably, economically and safely. The R B S design objectives can be placed into two categories. They are the design objectives intended to meet the data requirements and the design objectives intended to meet the system requirements. Much of the information in the following synopsis is taken from Readyhough (2003). The R B S system units must provide the following data: upper-atmospheric soundings at least once a day coincident with one of the main synoptic hours (00Z and/or 12Z) when requested; surface meteorological data on the hour, 24 hours per day, surface and subsurface oceanographic data on the hour, 24 hours per day, and seismic data both in a standard and event mode. The RBS system must be able to function with 99% reliability even in harsh conditions; provide data within standard error limits and be deployed at predetermined fixed strategic positions. The R B S system operation cannot add significantly to residual deleterious substances in the oceans. A l l RBS components and system procedures must comply with International Standards Organization (ISO) safety standards. Finally, the R B S data must be cost effective when compared to equivalent data from other in-situ sounding systems. 4.4 Progress To Date The year-2001 feasibility study concluded that the R B S units could be built. The year-2002 proof-of-concept phase successfully tested the components (Stull et al 2003). Over 130 test launches were made of various rocketsonde designs during 2002/03 by the 33 U B C rocket-design group. Readyhough designed a launch control system in 2003. Due to a lack of funding the scheduled prototype component integration was not initiated and the engineering development was (temporarily) ceased. Only the data-impact component (this research) of the project continued after 2003. The following engineering design constraints relevant to the data-impact portion of the study are taken from Readyhough. Tradeoffs must be made between the desired altitude and the number of rockets on the buoy, rocket mass and rocket diameter. Rocketsondes can be designed to reach 6 km altitude, but only 200 can be stored on a large commercial buoy. The rocketsondes will be able to reach only 3-4 km if 400-rocketsonde capacity is desired. The launch control system will provide near-vertical autonomous rocket launches given a buoy motion corresponding to the worst-case winter storm weather in the North Pacific. The maximum launch tilt design error is 5 degrees. 4.5 Real-Time Management For the dropsonde programs, real-time management through targeting was introduced in Section 3.9. With interactive two-way communication capabilities, it is possible to provide similar real-time management for the RBS array. The reasons for this design feature are straightforward. During active periods, targeting procedures through short-range forecasts and sensitivity analyses (Section2.4) may suggest that extra launches from certain R B S units will provide significant analysis improvements. Finally, there are the maintenance periods and emergency scenarios when the RBS must be shut down. Real-time management would likely increase the RBS cost-effectiveness. 34 4.6. Other RBS Advantages A l l the T H O R P E X observing systems can make contributions to the composite data mix. A l l have their strengths and weaknesses. The main RBS weakness is sociological; it may be a perceived hazard to air and marine craft. Mitigation strategies include keeping the maximum apogee below the commercial flight levels and posting the R B S positions on aviation and marine charts. If necessary, air and marine craft can avoid the locations. Here are the R B S strengths; the RBS: • will collect in-situ soundings of pressure, temperature, humidity and wind; these observations can be directly assimilated into the models and are the ( R A O B equivalent) data delivering large forecast benefit. • will provide vertical sounding profiles; aerosondes will collect single-level data along their flight routes. • will provide subsurface oceanographic and seismic data; only conventional buoys can do so now. • will provide synoptic meteorological data; the dropsonde and aerosonde systems will collect asynoptic data valid over a period of many hours. • sounding locations will be fixed; the aerosonde and dropsonde coverage is variable. • targeting will need minimal lead time; aerosondes and dropsondes may need 24 hours or more, increasing the risk of misdirection. • will work in all weather conditions, including storms with high seas; the aerosonde and dropsonde programs are sensitive to the weather conditions. 35 • equivalent will work in arctic conditions, lessening the human and economic difficulty of operating remote radiosonde stations there. 4.7 Multi-Rocketsonde Buoy Payload Attached to the buoy could be a rocketsonde cannister with rockets of various capabilities. For example, 8 km rocketsondes grouped in the center could be ringed with 6 km rocketsondes that are in turned ringed by 4 km rocketsondes on the outside. A pre-launch management analysis would determine which rocketsonde profile option will deliver the optimum impact. A 4 and 6 km multi-rocketsonde R B S would increase the R B S capacity to more than 200 launches. 4.8 Summary A Pacific RBS array could support meteorological, oceanographic and seismic missions. The RBS has advantages over other observing systems. Arctic-deployed R B S -equivalents are possible serving to fill the polar data void. A real-time management capability could increase the RBS cost-effectiveness. 36 Chapter 5 Data Assimilation, Numerical Weather Prediction and Experimental Issues 5.1 Foreword Like any other experimental scientist, the D A and N W P model user must be aware of the strengths and weaknesses of the tools employed and the procedures used. For example, the largest contribution to short-range N W P error is uncertainty in the initial conditions. The more accurate the estimate of the initial conditions, the better the quality of the forecasts. Paradoxically, the better forecast sometimes results from initial conditions that are the more compatible with the model, not from initial conditions more representative of the atmosphere. A n N W P modeler has to be aware of this modeling peculiarity, when it applies and how it influences the results. This chapter examines the general characteristics of D A and N W P . Specific M M 5 discussions are left to the following chapter. 5.2 Initial Conditions N W P is an initial-value problem. Most grid-point models use wind, temperature, moisture, and pressure as the basic dependent variables. The values of these parameters must be known at the integration start time. This is the purpose of the D A model (Section 2.3). The data required is the number of grid points times the number of dependent variables. For a typical D A - N W P system, this number could be of the order of 107. The total number of actual observations (Table 3.1) is of the order of 105 of which 104 are high-quality in-situ observations. The observation coverage is much smaller than the 107 model 37 requirement. As well, the observations are usually concentrated in certain areas and not spread evenly over most sectors (Section 3.2). For areas like the Pacific data void (Figure 3.2), the D A model uses vertical structure relationships to fill in the levels representing the mid-troposphere. These extrapolation procedures do not reproduce realistic atmospheric structures during some synoptic situations. Observations are often incomplete and imperfect. Errors can occur from having a less than adequate number of observations, and from not being able to resolve meteorological features of interest. Errors can occur as a result of not receiving observations frequently enough to monitor phenomena evolution. Observational errors result from inaccurate instrument readings or data-transmission problems. Some observations may be correct, but represent local phenomena too small to be resolved by the analysis; the resulting representation for a larger area is biased by extremely localized phenomena. A n estimate of these errors must be provided to the D A model. The first-guess field is a mandatory requirement for almost all D A methods (Section 2.3). Errors can be propagated by the N W P model and be present within the subsequent forecast used as a first-guess (Section 3.4). But, the best estimate of model state at every one of the 107 values must be provided, imperfect as it may be. Similarly, D A of indirect observations (satellite radiance, etc.) needs a model first-guess in order to determine how closely the model state can reconstruct the observations. This results in a "cyclic dependency"; the D A depends on a reasonable first-guess field; the first-guess field depends on the results of the D A or the N W P projection or both. Over data-void areas with the uncertain first-guess fields this dependency is often a fatal 38 weakness. During highly dynamic situations, weaknesses in the D A can lead to a series of forecast failures (Section 3.4 & 3.5). 5.3 Physical-law Constraints In order to mitigate some of the errors resulting from imperfect observations, data initialization and cyclic dependency , D A schemes try to constrain the data by use of physical laws. Examples are the use of the continuity equation that constrains the velocity field; the hydrostatic equation that constrains the pressure and height fields, and the balance equation that constrains the wind-mass fields. Unfortunately the implementation of physical constraints is often approximate. For example, the wind-mass balance constraint used is typically a linear approximation to the more representative non-linear constraint. Furthermore, a good first-guess forecast can be altered by non-appropriate physical constraints. The use of a linearized balance equation after D A of a few observations during a forecast may compromise the non-linear accuracy of the rest of the model state within the forecast domain. 5.4 Data Assimilation Each individual observation has several dimensions. A temperature observation will have a magnitude and an error estimate. A wind observation will have a magnitude, a direction and error estimates for both; The errors can be random or systematic. Systematic errors can sometimes be determined and eliminated. Random errors are usually assumed to have a normal distribution. The goal of objective analysis through D A is to improve the first-guess at the model grid points. Innovations are calculated by subtracting the first-guess state from the 39 observation state. Then a weighted amount is added to the first-guess. In mathematical terms, this can be stated as (Kalnay 2003) x a = x b + W(y°-H(x b)) (5-1) where x a = analysis state vector at the model grid points x b = background state vector at the model grid points (i.e., first-guess from previous forecast) y° = observation vector at the station locations H(x b) = forward operator acting on the model state y° - H(x b ) = the innovations (difference between actual observations and the model values interpolated to the observation locations) W = weight matrix that will determine the impact of the innovations at each model grid points The data assimilation schemes differ on how to compute and apply W, the weight matrix. In the Cressman scheme, observations outside a radius of influence (ROI) around any grid point are not used to update the field variable at that grid point. Those observations within the ROI are weighted according to how close they are to the grid point, with the closer observations receiving a greater weight. The Cressman scheme assumes that the observations are perfect. The observations are given full weight after being interpolated to the grid points. A n observation that happens to coincide with a grid point will determine the field variable at that grid point. The usual implementation is via a Successive Correction Method (SCM). The ROI is set to decrease for each pass through the updated first-guess field, so that the field is corrected to larger scale features during the first iterations, and conforms to smaller scale features during latter iterations. Unfortunately the Cressman scheme can directly assimilate only observations that correspond to the model variables. Physical models (required to 40 transform indirect observations to the model variables) must be applied independent of the S C M - D A , and are often inaccurate. Application of physical constraints such as a hydrostatic approximation and wind-mass balancing is carried out separately from the data assimilation and subsequent to it. Optimal interpolation (01) takes into account model and observation errors. Using the method of least-squares, background and observational data are combined in a statistical sense, resulting in a 'best-fit' analysis. 01 allows higher quality data to receive more weight in the analysis. Observations containing random errors receive more weight than those containing systematic errors. In addition, 01 takes into account data distribution by assigning less weight to those data lying within a cluster of observations. Unfortunately, large matrix inversions are required to perforin OI-DAj but modern numerical techniques have diminished this deficiency. The set of approximations needed to make the method practical means that the results are sometimes less than optimal (Kalnay 2003). The Bratseth technique is a successive correction scheme that converges to 01; it differs from S C M in that it allows for the specification of multivariate error covariance of the background and observations. Given sufficient iterations, the analysis converges to an optimal analysis based on some user-specified error variance. The Bratseth method delivers the benefits of 01 without the large matrix-inversion requirements (Bratseth 1985). Methods, commonly employed for both direct and indirect observations such as satellite radiances are termed variational, the most common is the three-dimensional variational (3DVAR) procedure. The model state is converted to the same space as the indirect observation by the use of a representation of the physics H(x b). For example, the vertical temperature and humidity profiles from the model at grid points are used to 41 calculate emitted and transmitted radiation leaving the top of the atmosphere at observation locations. Then 3 D V A R achieves data assimilation through the iterative minimization of a prescribed cost function. Differences between the analysis vs. observations and analysis vs. first-guess are penalized according to their perceived error. The benefits of 3 D V A R are that physical constraints such as geostrophy and hydrostatic balance can be included when minimizing the cost function. A n additional initialization step is not necessary. The quality of the observations is considered through the evaluation of errors in the first-guess field and the observations. Unfortunately, errors are introduced during the 3 D V A R - D A . The observations may have bias errors. The physical model required to transform the model state to the observational state adds complexity and may not be accurate. Application of the physical constraints and the minimization algorithm will introduce more approximations sometimes moving the updated analysis away from physical reality. Some of these problems are addressed with a four-dimensional variational procedure (4DVAR) that has a time component built into the procedure. A full non-linear model is employed as a dynamic constraint in the analysis of the observations. The downsides to 4 D V A R include the extremely high computational costs and associated time delay. 5.5 Comparison of the Assimilation Schemes S C M is easy to implement, flexible to experiment with and takes little computer time. The influence and range of the observations is easy to visualize. The influence of the observations can be made to be flow-dependent. Each observation is treated individually. Unfortunately, S C M does not have a theoretical basis. The results can be erratic due to the 42 possible introduction of unfavorable initial modes, such as gravity waves. Further processing has to be carried out in order to suppress these unfavorable modes. 3 D V A R does have a theoretical basis, and should deliver a better result than an empirical scheme like S C M . That has been the conclusion of the various research and operational groups that have implemented 3 D V A R . Major advantages of the 3 D V A R is the ability to assimilate observations related in a complex manner to the standard model variables, and the imposition of dynamic balance through the use of physical laws. With 3 D V A R all available data are used simultaneously. This avoids "bulls eyes" and other erratic results. However, the range of influence of any one observation is difficult to determine. 3 D V A R still requires many simplifying assumptions. One assumption is that the error covariances are homogenous and isotropic and not flow dependent, which is not very realistic in areas of strong gradients or steep mountains. The physical models providing the physical constraints are linearized and possibly quite inaccurate in critical areas. 5.6 Quality Control It is essential that quality control be performed to avoid the assimilation of erroneous observations. Quality-control procedures are based on direct comparisons between observations and the first-guess field, the difference being referred to as observational increments. Observational increments beyond a preset limit cause rejection of the observation. Most procedures assume that the first-guess field is reasonably representative of the true atmosphere at that time. However, a bad first-guess can cause good observations to be rejected. Typically, the bad first-guess propagates the error and 43 good-observation rejection continues through the next cycle (Section 3.4). This happens most often in data-void areas. 5.7 Model Error Both initial-condition and model errors contribute to total N W P error at extended time periods. Model errors result from the incomplete formulation of the equations of motion, the improper formulation of the physics, the intrinsic error that grows with increasing forecast length, and the error introduced from poor parameterization of sub-grid scale processes that can influence the larger scale. 5.8 Regional Models and Boundary Conditions As well as being an initial-value problem, N W P is also a boundary-value problem. The values at the boundary have to be known or assumed for the entire integration period. When using a global N W P model, the boundaries are typically the condition at the earth's surface as the lower boundary, and the top of the atmosphere as the upper boundary. Surface values of temperature, albedo, sea-state roughness, heat and moisture fluxes, soil moisture, etc. are needed. In addition, subsurface conditions of soil temperature, soil moisture, ocean temperatures and current profiles are increasingly important for coupled models. Over the Pacific Ocean, the correct parameterization of surface fluxes is a major N W P issue. This is one of the problems that the R B S is hoping to mitigate. The use of regional mesoscale models (Section 2.7) for weather prediction has arisen from the desire to reduce the model errors through an increase in horizontal resolution that cannot be provided in a global model. Unfortunately, there is a downside to the use of regional models. They require the use of predetermined lateral boundary 44 conditions obtained from a larger regional model or a global model. A global model or larger domain regional model must be available to generate the boundary conditions for the smaller domain regional model. The ETA104 data product, used in this research to provide time-varying lateral boundary conditions, will be described in Section 6.8. Over the inflow areas (See Figures 3.3), the boundary condition from the host model penetrates into the regional model, while at the outflow areas the regional model solution leaves the domain. After a case-dependent integration period, most initial regional information is swept out of the domain and the information from the larger model dominates the NWP solution. Errors arise because these boundary values come from another model which itself has errors. These values may be incompatible with what the regional model is forecasting near the boundaries. The effect can be spurious phenomena affecting the results. 5.9 Data Impact Experiments Observing system sensitivity studies (Arnold et al. 1986) fall within two groups: observing system experiments (OSEs) and OSSEs. OSSEs guide the design of future observing systems like the RBS, and provide an assessment of their likely potential. OSSEs are complicated to perform and have a number of limitations that will be pointed out later in this document. OSEs are real-data experiments; they assess instruments or observing systems that already exist. An OSE is essentially a data-denial exercise, with the observations either added or removed. Data-addition OSSEs determine the impact of adding virtual data to the forecast runs while data-denial OSSEs determine the impact of withholding data. When the NWP model produces both the first-guess field and the forecast, then the OSSEs are called 45 identical-twin experiments (Arnold et al. 1986). If the first-guess field is produced by another model, then the OSSEs are referred to as fraternal-twin experiments. This study used identical-twin OSSEs. 5.10 Linear vs. Non-linear Growth A small variation in initial conditions is often treated as a perturbation on a basic state trajectory (Kalnay 2003). The assumption that "small" N W P perturbations start their evolution linearly is commonplace in N W P studies. The magnitude of operational perturbations are commonly held to continue evolving approximately linearly for at least two days; afterwards the integration will progress into the non-linear regime and the forecast will eventually descend into chaos. The directional change in the trajectory is usually ignored. These assumptions are common in studies concerning ensemble forecasting, data assimilation, model sensitivity and observation targeting. Other studies have suggested that the linear regime is much shorter than two days and that a perturbation may evolve in different directions as well as with different magnitudes (Gilmour et al. 2001). When interpreting the O S S E results, it is important to know where the linear/non-linear threshold is and where a large change in direction may have occurred. In practice, both are difficult to determine. The OSSEs in this study assume that model error growth is the same for both the identical twins, and that differences between the O S S E results are due to initial conditions only. 46 5.11 Verification Verification measures must be relevant in order to make a reasonable data impact assessment. Proper verification should be multi-faceted, including a provision for subjective assessments. For this study, subjective assessments were not attempted except for example cases (Section 8.4). Objective verification techniques are needed to quantify the results but objective verifications alone do not fully reveal the quality of the forecasts. Therefore they cannot support a full interpretation of the O S S E results. However, resource managers need quantification for their decision-making process, making objective verification the only practical option. 5.12 Summary The above discussions have outlined various issues that must be considered while performing the OSSEs and interpreting the results. Data used for NWP initialization should have appropriate density. D A must produce a reasonable update and one that is compatible with the forecast model. F D D A will allow errors to propagate in data-void areas. Errors will always be present in the observations. The unfortunate rejection of good data will diminish the impact of the observations and may lead to the wrong conclusions. Boundary conditions obtained from a larger domain model must remain compatible with the NWP regional model solution throughout the integration to prevent development of spurious phenomena. The O S S E forecast must remain reasonably close to the results of its identical twin reference run. Lastly, the verification metrics may not provide reliable results transferable into user-relevant information. 47 Chapter 6 The MM5 Modeling System 6.1 Foreword The Perm State/NCAR mesoscale numerical model MM5-V3 .6 is used for the numerical integrations. Much of the following descriptive material is taken from Dudhia et al. (2003). The model is supported by several pre- and post-processing routines comprising the entire M M 5 modeling system that will be referred to as the M M 5 . N C A R provides user support for the M M 5 . Unfortunately, the M M 5 allows experimentation with only the surface and upper-air levels. A description of the project implementation follows. 6.2 Domain The M M 5 domain for this study is based on a polar-stereographic projection true at 60 degrees north (Figure 3.1). The centre of the domain is at 50N 120W. The horizontal grid has 160 grid points from bottom-to-top and 200 grid points from left-to-right. The grid spacing represents a nominal 45 km on the earth's surface valid at the center of the domain. A n Arakawa-B staggered grid is used. The initial preprocessing routines use pressure as the vertical coordinate with 86 levels. The N W P portion of the modeling system uses the terrain-following vertical coordinate sigma (a), defined as: er = PZP^E ( 6 . 1 } p _ reference — p _ top where cr = sigma, a dimensionless quantity 48 p = pressure p_top - a specified constant top-of-the atmosphere pressure P - reference .= a reference-state pressure, constant in time, but varying with terrain height. Conversion to sigma space results in 42 terrain following sigma levels with a concentration below 40 kPa. The numerical integration occurs in sigma space. The sigma vertical spacing is approximately 0.25 km at the lower levels and 2 km at the upper levels. 6.3 Preprocessing The M M 5 begins by horizontally interpolating regular terrain elevation and land-sea surface information onto the domain grid. Initial and boundary condition data are taken from a previous M M 5 forecast and/or the ETA104 data product. The data on the ETA104 pressure levels are interpolated from the native grid and map projection to the M M 5 vertical pressure grid and map projection. Data from observations can be assimilated at any stage of the forecast run. For M M 5 - S C M , the data are assimilated on the pressure levels before conversion to sigma space. For M M 5 - 3 D V A R , data assimilation occurs in sigma space. Conversion to sigma space requires vertical interpolation, diagnostic computation, and data reformatting. 6.4 Model Numerics Second-order centered finite differences represent the gradients except for the precipitation-fall term that uses a first-order upstream scheme. A second-order leapfrog time-step scheme is used for the advection, Coriolis and buoyancy terms. The nominal time step is 30 seconds. The fast terms responsible for sound waves are handled with a time-splitting scheme. The finer time interval is XA of the nominal time step (7.5 seconds). 49 A forward step is used for diffusion and microphysics. Some radiation and cumulus options use a constant tendency over periods as long as every 30 minutes or so. Implicit schemes are used for vertical sound waves and vertical diffusion. 6.5 Model Physics The Grell cumulus parameterizations scheme is based on a simple single-cloud scheme with updraft and downdraft fluxes and compensating motion determining the heating and moistening profile. Shear effects on precipitation efficiency are considered. The Planetary Boundary Layer (PBL) scheme implementation is similar to that in the N C E P Medium Range Forecast model. The Dudhia ice scheme adds ice-phase processes to the P B L scheme. There is no supercooled water and there is immediate melting of snow below the freezing levels. A cloud-radiation scheme accounts for longwave and shortwave interactions with explicit cloud and clear-air. A five-layer soil model allows substrate temperature to be predicted in layers. 6.6 MM5-SCM Implementation In the analyses of wind, temperature and dew point at pressure levels, the M M 5 -S C M L I T T L E _ R module allows an anisotropic ROI. The ROI are elongated in the direction of the flow and curved along the streamlines, resulting in a banana shape. This scheme results in an elliptical ROI shape under straight-flow conditions. Unless otherwise specified, the ROI is set here to be 10 grid points along the direction of flow. The M M 5 -S C M number of iterations will be a maximum of 10, the amount allowed by L I T T L E R . Justification for these settings is given in Section 9.5. 50 After S C M - D A , the model data is checked for hydrostatic balance, but there is no mass-wind balancing. The S C M - D A may introduce spurious gravity-wave modes. 6.7 MM5-3DVAR Implementation Barker et al. (2003) provides a detailed description of the M M 5 - 3 D V A R implementation. M M 5 - 3 D V A R may be run from both a "cold-start" mode, where the background field is taken from another model or it can be run in "cycling" mode where the background field is a short-range M M 5 forecast or a newly improved M M 5 - 3 D V A R analysis. A n observation preprocessor prepares observations for ingest into M M 5 - 3 D V A R and estimates the error for each observation. The NCAR-adopted errors are shown in Table 6.1, and are used for this study unless otherwise noted. N C A R PR< 3VIDED E R R O R E S T I M A T E S ( Pa = Pascal > Mean sea level -Pa Height m Pressure Pa Wind m/sec Temp degK Relative Humidity % surface 200 6.0 100 1.1 2 10 100 kPa n/a 5.0 100 1.1 1 10 85 kPa n/a 5.4 100 1.1 1 10 70 kPa n/a 6.0 100 1.4 1 10 50 kPa n/a 9.4 100 2.0 1 10 40kPa n/a 11.6 100 2.8 1 10 Table 6.1. The errors are based on those used in the FNMOC NOGAPS model and the ECMWF model. M M 5 - 3 D V A R uses an incremental approach. It assumes that the first-guess field (in this case, a 12 hr forecast from a previous forecast) is balanced. The minimization and balancing is carried out on the increments (the procedure is described in Barker et al. 2003, pp 18-19), so any errors introduced by the minimization algorithm are limited to the increments and do not affect the full analysis field. Mass-wind balancing is achieved by computing geostrophically and cyclostropically balanced perturbation increments from the 51 wind-analysis increments. A l l increments are then combined with the background field and output as the new analysis. This is important to remember when interpreting the results. In MM5-3DVAR, the wind increments determine the mass increments. To estimate climatological background error covariances, the N C A R " N M C method" produces monthly-averaged forecast differences between 24 hr and 12 hr forecasts valid at the same time. These background errors are computed for a variety of resolutions. In reality, errors in the background field are synoptically dependent and vary from day-to-day depending on the current weather situation. Operationally, the use of climatology is realistic; background error statistics cannot be calculated for each D A cycle. The M M 5 -3DVAR has a number of variables that allow tuning of the N M C climatological estimates. Scalings are used within a recursive filter to approximate a Gaussian correlation between the first-guess, observations and the updated analysis. The Gaussian correlation may be expressed as exp(-(r/As)2), where "s" is the NMC-method estimate of the background error scale-length; " A " is the empirical scaling factor available in M M 5 -3DVAR for user tuning; "r" is the distance between the observation and grid points. These parameters determine the range of influence of the data. The control variables used by MM5-3DVAR are streamfunction, velocity potential, unbalanced pressure and specific humidity. A scaling parameter is available for each of the control variables. Further discussion is included in Section 9.5. Surface observations must be within a 100 m of the model surface or the data will be rejected. Upper-air increments must be within 5 times the observation-error limit (Table 6.1) or the observation will be rejected. Linearized balance constraints such as hydrostatic, geostrophic and cyclostrophic are continuously applied to the mass and wind fields during 52 the minimization. The horizontal arid vertical-error components are applied via isotropic recursive filters and empirical orthogonal functions, respectively. The increment vector is calculated for all grid components. A n "outer-loop" link permits the recalculation of the increments using the new analysis as an improved background. In this way, observations previously rejected may be accepted in a subsequent outer-loop run. After the cost-function is calculated, minimization algorithms are applied. M M 5 -3 D V A R allows a choice between a quasi-Newton minimization algorithm or conjugate-gradient algorithms. Since the 3 D V A R procedure influences the entire domain, the lateral boundary conditions must be modified to reflect differences between the background forecast and the new analysis. Post-processing includes checks to ensure that the output data is within physical limits. 6.8 Initial And Lateral Boundary Data The E T A model domain is shown in Figure 6.1. The M M 5 domain is superimposed for comparison. The E T A model has 22 km horizontal resolution and 50 levels. The boundary conditions to the E T A model are provided by the N C E P Global Forecast System (GFS) model and are updated every three hours. The greatest vertical resolution is in the boundary layer below 85 kPa with 24 layers. Each E T A forecast is initialized by the E T A Data Assimilation System (EDAS), N C E P ' s implementation of a fully cycled F D D A (Section 2.3). At the end of a three-hour run, the N C E P 3 D V A R is used to create an updated analysis. E D A S - 3 D V A R assimilates the complete set of observations (e.g. Table 3.1) within 1.5 hours of the analysis time 53 presented to it, including a large amount of radiance data (Section 3.8). Error minimization algorithms arrive at the optimal updated objective analysis (Staudenmaier 1996). Details are available at Rogers et al. (a Web site). Figure 6.1. The ETA 22 km domain is the larger area bordered by the curved solid line. The MM5 45 km domain is bordered by the inside rectangular area. The ETA104 product domain is bordered by the inside dashed line. Each day N C E P conducts four separate E D A S / E T A production runs. When finished, E T A data is packaged in standard GRJB format and made available to users at the N C E P ftp site1. The E T A 104 data product is downloaded routinely at U B C and becomes available about two hours after its valid time. Unfortunately, the E T A 104 data is provided at much lower resolution than the resolution used by the model. Attributes of ETA104 are: Nested Grid Model (NGM) super C grid number 104; 147x110 grid points; a polar ftp://tg%.nws.noaa.gov/SL.us008001/ST.opnl/MT.eta_CY.00/RD.20030219/PT.grid_DF.grl/£h.0000_tl.p ^ .g rbgrd 1 54 stereographic grid centered at 105W; horizontal grid point distances are 90.8 km at 60 degrees N latitude; vertical resolution is approximately 5 kPa. The M M 5 runs must be provided with initial conditions throughout the domain, and boundary conditions for each time step of the integration period. The ETA104 data are decoded and converted to the M M 5 format. E T A 104 includes forecast data at three-hour intervals starting at the initial time to the maximum of 84 hours. The horizontal resolution of the E T A 104 data is only about Vi the resolution of the M M 5 model used. The vertical resolution is about Vi of the M M 5 model at the lower altitudes and is about the same at the higher altitudes. To provide data at all the 160x200x42 M M 5 grid points, interpolation is required. 6.9 Visualization After the M M 5 has produced the forecast files, they are converted to G r A D S (Grid Analysis and Display System) format for post-processing. G r A D S is an interactive desktop tool used for manipulation and visualization of gridded data. G r A D S uses a 4-dimensional data environment consisting of longitude, latitude, vertical level, and time. Library or user-defined functions allow for the data manipulation and visualization. 6.10 Summary N C E P E D A S is a state-of-the-art D A system with the ability to efficiently assimilate all in-situ data including satellite radiance data. A lower resolution version of the E T A output is made available to users as the ETA104 product. In this study, the ETA104 data is used by M M 5 for initial and boundary conditions. M M 5 has both a S C M 55 and 3 D V A R D A module. Table 6.2 includes a summary of the M M 5 setup and the attributes of the E T A data. MODEL AND DATA 1 PRODUCT ATI rRIBUTES MODEL ETA Model ETA104 Data Product MM5 Model Horizontal resolution 22 km 91 km 45 km Number of levels 50 about 20 42 Vertical resolution 0.6 km lower 1.5km higher about 5 kPa 0.25 km lower 2.5 km higher Time step 60sec N / A . 30 sec Domain Figure 6.1 Figure 6.1 Figures 3.1, 6.1 Table 6.2. The attributes of the ETA model, the ETA104 data product and the MM5 model implementation. 56 Chapter? Experimental Methodology 7.1 Foreword The quality of the M M 5 forecasts is related in part to the quality of the E T A 104 data products that are used to initialize and bound them. It is assumed here that the E T A analysis is of high quality given the available data and 3 D V A R - D A (Section 6.8). Based on the preceding assumption, it is further assumed that data extracted from the ETA104 OOhr data are a close approximation to in-situ data that would be obtained if an observing system (e.g. RBS) had been present and produced the same result per observation. This critical assumption forms the basis for the OSSE procedure. The terms "analysis" and "OOhr" will be used interchangeably when referring to the first data set in the ETA104 data product, regardless of the valid time (00Z or 12Z) of the analysis. Hereafter, a reference to E T A data will mean data extracted from the ETA104 data product and converted to M M 5 format. 7.2 Basic Experiments Some of the experiments do not require data assimilation. The REF12 and REF00 forecasts are initialized using, respectively, the E T A 12Z and 00Z analysis data. The 00Z dataset is always valid 12 hours following the 12Z dataset. Besides being used to provide initial and boundary condition data, the E T A forecasts are used for comparison with the corresponding M M 5 results. The E T A forecasts initialized at 12Z and 00Z are named E T A 1 2 and E T A 0 0 respectively. Various experiments will compare the results of the M M 5 and E T A forecasts. R E F X X and E T A X X mean that the X X is a placeholder for 57 either 00 or 12 or both. Certain experiments will require additional extensions to the names and will be defined later. 7.3 OSSE Methodology The M M 5 modeling system is used for both the D A and N W P , so the O S S E procedure used here is an identical-twin (Section 5.9). The control or "reference-atmosphere" is the REF12 forecast that excludes any new data. The identical twin using the assimilated data will be referred to as an "OSSE atmosphere". For an O S S E atmosphere run 12 hrs after initial time, the REF12 forecast is "enhanced" with extracted virtual data from the new E T A 00Z analysis. "After data addition" (ADA) or "after data denial" (ADD), data impact is measured as the difference in day-1 to 3 evolutions of the R E F X X and O S S E atmospheres (Graham et al. 2001) from the (VER) verification atmosphere. 12Z-00HR 00Z-12HR 24 36 48 ISO 72 12Z-84HR START REF12 FO RECAST END REF12 . FORECAST FROM 12Z INITIAL DATA —w-EX TRACT DATA FR 0M 00Z AN ALYSIS INS E RT INTfl 1? HOUR FnRFHAST END OSSE » FORECAST START RE FDD FORECAST FROM 00Z INITIAL DATA END REFDO „ FORECAST COMPARE wrm T ANALYSIS LUM PARE WITH VERIFYING ANALYSES Figure 7.1. The OSSE timeline. The control or "reference-atmosphere" is the REF12 forecast that excludes any new data. The OSSE forecast has new data assimilated at 00Z. After data addition, data impact is measured as the difference in one-to-three-day evolution of the REFXX and OSSE atmospheres from the verification atmosphere. The virtual data that are added in the OSSEs have two configurations: virtual radiosonde soundings (vRAOB), and virtual R B S soundings (hereafter v R O B for virtual 58 rocketsonde soundings). Each v R A O B is a single vertical profile with all the E T A model levels included (surface to 16 km). Normally, an evenly spaced array of many v R O A B s is assimilated to fill the case-study areas, where each v R A O B is horizontally spaced 225 km (i.e., at every 5th M M 5 grid point). This horizontal resolution of 225 km is the approximate finest horizontal resolution of the existing radiosonde network over North America. The average horizontal radiosonde resolution is 315 km. vROBs are similar to vRAOBs , except that the vertical profiles are only from the surface to 2, 4, 6 or 8 km. Unless otherwise specified, the v R O B vertical profiles will span the surface to 6 km. The number of v R O B sites is small and relatively widely spaced, corresponding to proposed locations for the R B S buoys, compared with the v R A O B s (Section 7.5). A few specialized experiments deviate from the above convention. They will be discussed as required (Chapter 9). 7.4 Reference and Benchmark Atmospheres REF12 - Reference Atmosphere: No virtual sounding data is added to this atmosphere, initialized from the E T A 12Z analysis and bounded by the E T A 12Z forecast data. REFOO - Reference Atmosphere: No virtual sounding data is added to this atmosphere, initialized from the OOZ E T A analysis and bounded by the E T A OOZ forecast data. Relative to the 12 hr forecast from 12Z valid at the same time, this atmosphere has the benefit of the complete 12Z to OOZ E D A S value. From the previous 12Z through to OOZ, all the data types shown in Table 3.1 and large amounts of radiance data have been assimilated through four E D A S cycles. The results of the REFOO run form the "ideal" 59 benchmark results for the OSSEs, and are used as a screening tool for the Pacific analyses. When the REFOO verifies better than the previous R E F 12 atmosphere, it is assumed that the OOZ analysis (12 hrs later than the 12Z analysis) has been updated by E D A S in a proper manner. BEN-Benchmark Atmosphere: v R A O B s are added densely and evenly (at every 5th grid point) within the large bolded rectangular area shown in Figure 7.2. This a60tmosphere establishes "practical" benchmark results for the OSSEs (as distinct from "ideal" benchmark results available from the REFOO forecasts). For the purposes of the OSSEs (when the analysis over the Pacific Ocean is reasonable) the forecasts from the B E N atmosphere are assumed to be the best that can be expected from an ideal radiosonde observational network within the same domain. Descriptions of some experiments require extension of the B E N label (Section 9.13). The results of the R E F 12, B E N and REFOO forecasts are used for comparison, screening and bounding the M M 5 O S S E atmosphere results, which should fall somewhere between the REF12 and B E N results; the REFOO provides the ideal results. Verification takes place only over the North American continent where the actual observational network is dense and represents the current ideal. A n issue with this O S S E procedure is the uncertain Pacific analysis available within the OOZ E T A data. There is only limited in-situ data available to update the Pacific analysis (Table 3.1). Fortunately at times, the data happens to be available in critical areas and has significant positive impact. When the forecasts from the B E N and REFOO atmospheres provide a substantial (verified) improvement over the forecast initialized earlier (i.e., the REF12 atmosphere), this is a strong indication that E D A S has corrected the 60 Pacific analysis in a proper manner within the areas that later impact the verification area. vROBs from the same source (a subset of the B E N vRAOBs) can then be used with some confidence. At other times there is little new data, and a questionable first-guess field negatively influences the analysis. If the B E N and REFOO atmospheres deteriorate the forecast, then it is assumed that E D A S has not corrected the Pacific analysis properly in the area that later impacts the verification area. The examples of forecast deterioration can help document the negative impact of the Pacific data-void. 140W 130W 120W H O W 100W Figure 7.2. The heavy rectangle outlines the B E N region, which for this OSSE is filled with a uniformly distributed, high-density array of virtual radiosonde soundings (vRAOBs) to 16 km altitude (over both continent and ocean). For the PACvoid experiments, virtual sounding data are excluded from the mostly-oceanic region west of the idealized west coast of N . America, and south of the idealized US-Mexican border. For W N A void experiments, the simulated west coast of N . America is shifted further east, with sounding data excluded everywhere west and south of that shifted coastline. 7.5 OSSE Atmospheres Eight O S S E working atmospheres (PACvoid, P A C _ R B S , PACvoid_RBS, P A C T A R , WNAvoid , W N A _ R B S , W N A v o i d R B S , and A R C ) are described below. 61 These O S S E atmospheres are divided into three groups, based on where the vROBs are added: P A C for the Pacific Ocean; W N A for western North America, and A R C for the arctic. 7.5.1 The OSSE - PAC Atmospheres P A C v o i d - Pacific Data-void Atmosphere: For this simulation, an idealized coastline (interface between data-void and data-rich regions) is located approximately parallel to the west coast of North America (Figure 7.2). Data added east of the interface are exclusively v R A O B s , while no soundings are added west of the interface. 140W 130W 120W now 100W Figure 7 .3 . The PAC_RBS atmosphere consists of only low-altitude virtual rocketsonde soundings (vROBs) at the locations indicated, which are at the intersection of every 5 ° parallel and meridian. Subsets include the eastern Pacific (ErnPACRBS) and central Pacific (CenPAC_RBS). There are no other soundings anywhere else in the whole domain. 62 PAC_RBS - Pacific Rocketsonde Buoy System Atmosphere: vROBs are added at potential RBS locations in the Pacific (Figure 7.3) with no virtual soundings (neither vROBs nor vRAOBs) anywhere else. Three, Six, Nine, and Twelve-buoy configurations within an eastern Pacific array (ErnPAC_RBS) are shown. The SixBuoy array includes the ThreeBuoy locations, the NineBuoy array includes the SixBuoy locations, etc. A similar TwelveBuoy array can cover the central Pacific (CenPAC_RBS) area. The AllBuoy array includes both the E r n P A C R B S and the CenPAC_RBS arrays. 180W 17QW 160W 120W 80W 70W 6QW M O W 1 3 0 W 120W H O W 100W Figure 7.4. The PAC_TAR atmosphere consists of only low-altitude virtual rocketsonde soundings (vROBs) at the locations indicated, which are at the intersection of every 5° parallel and meridian. When referenced the ErnPAC_RBSs will be preceded by an E. The CenPACRBSs will be preceded with a C. PACvoid_RBS Atmosphere: The vROBs from P A C _ R B S over the ocean and the v R A O B s from PACvoid overland (as shown in Figures 7.2 & 7.3) are merged into atmosphere P A C v o i d R B S . This atmosphere is used to support various operational scenarios; it would be similar to a future real radiosonde network of dense sounding 63 coverage over land and sparser, lower-altitude, coverage over oceans. vROBs added in the Pacific data void are assumed to be from one of the RBS arrays described above. P A C T A R - Targeted-RBS Atmosphere: A specialized set of OSSEs will study the possibility of R B S real-time management through targeting. vROBs from R B S sites are handled individually. The ErnPAC_RBS array is extended further south. The R B S sites of the E r n P A C R B S and C e n P A C R B S arrays will be labeled E l to E15 and C l to C12, respectively. These configurations are shown in Figure 7.4. Some experiments require an extension to the names listed above and will be discussed with the General Findings (Chapter 9). 7.5.2 The OSSE - WNA Atmospheres To provide a regime of consistently reliable E T A data initial values from which to extract more accurate vROBs, the simulated west coast of North America (i.e., the data-void interface) is shifted 20-25 degrees longitude east of the PACvoid interface. For these OSSEs, analogous atmospheres to the P A C group (introduced above) are developed. This type of experiment is a hybrid of OSSEs and OSEs; real radiosonde soundings are denied from a region, while vROBs (of more accuracy than the P A C _ R B S vROBs) are inserted at selected locations. WNAvoid Atmosphere: A data void is created that spans both western North American and the Pacific (Figure 7.2). Data added east of the new interface are always a dense coverage of vRAOBs . WNA_RBS Atmosphere: This is an RBS-like situation over western North America, where data added west of the new interface are the vROBs with a maximum height of 8 km. This RBS configuration (Figure 7.5) allows the use of consistently reliable 64 data (from E T A analyses over western North America) to act as vROBs. Adding all vROBs between the PACvoid interface and the WNAvoid interface (Figure 7.2) will be considered a W N A a l l atmosphere. No soundings are added anywhere else in the domain. 140W 130W 120W H O W 100W Figure 7.5. The WNA_RBS atmosphere consists of only low-altitude virtual rocketsonde soundings at the locations indicated, which are at the intersection of every 5° parallel and meridian. There are no other soundings elsewhere in the domain. W N A v o i d _ R B S Atmosphere: vROBs are added west of the shifted interface, and v R A O B s are added east. This atmosphere is analogous to the P A C v o i d R B S atmosphere, but with consistently reliable data available to insert as vROBs. 65 7.5.3 The OSSE ARC - Arctic Atmosphere 140w 130w 120w l l O w lOOw Figure 7.6. The ARC_RBS atmosphere consists of 6 km apogee vROBs assimilated at the locations indicated. The A R C atmosphere is shown in Figure 7.6 and is represented by a linear array of 6 RBS-equivalent sites through Alaska, Yukon, Northwest Territories and Nunavut. Ideally an arctic R B S array should be on the south shore of the Arctic Ocean. For this domain, the array shown in Figure 7.6 site was chosen in order to remain reasonably away from the northern (upper) domain boundary, to avoid lateral boundary errors. 7.6 Verification atmospheres V E R Atmosphere: E T A analyses are used for verification. Quantitative verifications are made only over North America in the VERwest and VEReast subdomains (Figure 7.7). 66 140W 130W 120W HOW lOOW Figure 7.7. OSSE verification domains over the data-rich continent are divided into VERwest and VEReast, which are placed just east of the PACvoid shoreline and the WNAvoid-shifted shoreline, respectively. The verifications over VERwest and VEReast will represent the NWP results over WNA and ENA respectively. 7.7 Groups Gradual sequences of O S S E changes can be studied together as a group. Examples include a gradual increase in area covered by vROBs, or gradual changes in v R O B maximum altitude. The P A C Group examines the effect of increasing sounding locations added at OOZ, starting with zero v R O B in REF12, and proceeding through several variations of P A C _ R B S , PACvoid, several variations of PACvoid_RBS, to the B E N atmospheres. The W N A Group also examines the effect of increased v R O B coverage, but for the shifted coastline. The R E F 12, several variations of W N A _ R B S , WNAvoid , several variations of W N A v o i d R B S , and B E N atmospheres form another group. 67 The Data-void Group: The R E F 12, WNAvoid, PACvoid, B E N , and REFOO atmospheres represent atmospheres with increasing amounts of data available as v R A O B s , and compare the effect of forcing larger data-void areas across the domain (Figure 7.2). 7.8 Impact metrics The difference between these OSSE results and the verifying analysis valid at the same time is compared to the corresponding difference of the R E F X X atmosphere parallel run from the same verifying analysis. The arithmetic difference (DIF), root-mean-square-difference (RMSD) from V E R , and spatial-correlation (COR) with V E R are used to assess the impact of the virtual soundings. Parameters used for verification are mean-sea-level pressure (MSLP), 70 kPa geopotential heights (HT70) and 70 kPa Wind speeds (WS70). It is assumed that trends in these statistical indicators measure forecast improvements or degradations. A perfect analysis or forecast will provide a R M S D score of 0 and a C O R score of 1. When the forecasts deviate from V E R , the R M S D values increase bounded by the physical limits of the parameter. Concurrently C O R will decrease as the fields become out of phase, lowering to minus 1.0 if the fields are totally out of phase. R M S D metrics measure the anomalies between fields. C O R metrics measure the correlation between the full fields. R M S D statistics produce a greater variation and seem better suited to support the analysis of the results (that can be considered anomalies from a reference field). Surface pressure fields have a greater variation than the upper-air fields. R M S D M S L P statistics produce a greater variation in values but are prone to errors in mountainous regions (like W N A ) due to the reduction to sea-level pressure requirement. The free atmosphere HT70 is less prone to errors and measures the sensitive region of the free atmosphere and is the approximate mid-point of the RBS profiles. 68 Absolute N W P error comparisons will be quantified in percentage values and will be used when dealing with the reference atmospheres. Using an example with R E F l 2 and REFOO, an example error comparison would be RMSD REFU-RMSD REFOO %Difference = •100% (7.1) RMSD _REF\2 Some N W P error will always be present, so O S S E evaluations will be made in relative terms. Unless otherwise noted, the B E N results will provide the standard. The relative O S S E results will be measured by a G A I N index, the amount of improvement from REF12 relative to B E N . If the B E N atmosphere has not provided a proper forecast, then REFOO will be substituted as a higher standard. Using an example with REF12, a ThreeBuoy O S S E , and B E N , the relative gain would be measured as " RMSD _ REFl 2 - RMSD _ ThreeBuoy' GAIN = (7.2) RMSD _ REFl 2 - RMSD _BEN Verification of the M S L P and HT70 involves the comparison of two scalar quantities and is straightforward. Verification of the wind field is more complicated. The error between two vectors is itself a vector and cannot be quantified with one number. Wind speed is a scalar and will be assessed. Unfortunately this does not present a total picture, for wind error is affected by errors of direction as well as errors of speed. It will be assumed that over the verification area, the average wind direction error is zero. Wind speed verification will be used only when winds are the specific parameter under investigation. 69 7.9 Summary The O S S E procedure described above has the advantage of a well-controlled environment. Model-error and numerical-error biases are factored out with the identical twin comparisons. Error growth rate of the O S S E atmospheres remains constant, so the forecast differences result from variations in the initial conditions. Use of B E N and REFOO atmospheres allow Pacific-analysis screening and calibration of the forecast results. The improvement of the forecast statistics relative to the reference (REF12) atmosphere provides the most useful information. 70 Chapter 8 OSSE Case-study Results 8.1 Research Objectives The primary objective of this research is to support the feasibility, design, development, deployment, operation and cost-effectiveness of the R B S system. A l l other issues flow from these goals. Operational issues include Pacific RBS location, buoy density, size, seasons, launch times, launch tilt and the desirability of real-time management. Data issues include profile heights, sensors, and error tolerances. Impact issues include: the standalone impact and the RBS impact within an operational environment. Deployment issues include possible phased strategies. Preparatory requirements for the study included the analysis of MM5-NWP accuracy and a determination of the possible upper limit of R B S N W P improvement. These experiments can be called "basic experiments" since they did not need MM5-DA. Other requirements included the examination of the tools used to carry out the research: the appropriateness of the MM5, its D A schemes and the O S S E procedure. D A issues included the sensitivity and responsiveness of the MM5 results to the S C M and 3DVAR schemes, the appropriate S C M ROIs, the 3DVAR error covariance scaling parameters, the effects of 3DVAR observation-error levels, and the effect of quality-control criteria. Other issues included the effects of the ETA-driven initial-data resolution, the boundary conditions, and the validity of the perturbation linear-growth assumption. These experiments can be termed "setup experiments". 71 Subsequent issues involved possible redistribution strategies for the current observing network leading to funding options for the RBS. Further issues involved the comparison of R B S impact with other potential T H O R P E X instrumentation. 8.2 The Study Approach The study started in December 2001 and ended in August 2004. For convenience, the year was divided into two periods, winter and summer. Winter started in October and continued through to April of the following year; summer was May to September. It was assumed that winter encompassed the synoptically active period; summer encompassed the quiet period. It is recognized that this arbitrary division does not always hold true. During the winter of 2001/02, most of the research effort was placed on the development of the O S S E and supporting activities (described above). The winter of 2001/02 was a near-normal winter over W N A . OSSEs performed on a few high-impact cases captured during the winter of 2001/02 along with a few cyclone summer cases for 2002 were initially analyzed. This period will be termed the Test Period. During the Test Period, the forecasts were a maximum of 60 hrs, 48 hrs A D A . The D A procedure was the S C M . The verification parameter was M S L P . After the end of the Test Period, the Study Period was started in August 2002. The forecast duration was extended to 84 hours with the hope of better discerning certain trends. Unfortunately, the winters of 2002/03 and 2003/04 were E l Nino 1 winters. The winter of 2002/03 had a strong E l Nino signal. A ridge of high pressure persisted close to the west coast of North America (See Figure 3.3b for an example case). Most of the 1 El Nino is a disruption of the ocean-atmosphere system in the tropical Pacific having important consequences for weather around the globe. 72 Pacific storm trajectories turned northward to Alaska, Yukon and northern B C , and deflected away from the VERwest area. Many E N A winter storms had their origin in the North American arctic; their trajectories were directed over VEReast (Figure 3.3b). The impact of the flow from the arctic areas (polar data void) was significant during this period. The same pattern (although weaker) set up and persisted during the winter of 2003/042. The O S S E cases for the winter period for 2002/03 and 2003/04 were assimilated with the S C M and 3 D V A R respectively. For comparison, a few O S S E cases (winter 2003) were assimilated with both. The results for the Test Period and the Study Period cases will be presented separately, with the Test Period results leading and the Study Period results following (Chapters 8 & 9). The winter Test Period results were for cyclone cases only and will be used as leading indicators of the effects to be determined. The verification metric used for the Test Period cases is M S L P ; the verification metric used for the Study Period cases is mostly HT70 but sometimes WS70 when winds are the specific issue. Study-Period cases include active and null 3 cases; the statistics are more robust, but forecast-skill effects are less dramatic. 8.3 Cases Overview The computer runs needed to support this research were carried out with "research priority', less priority than the U B C "operational runs". The dates and number of runs were based on the availability of resources, both human and computational. The U B C computer resources that were available are described in Appendix B. The non-availability of continuous resources resulted in many gaps in the data sets. As a general rule, the basic-2 A similar pattern occurred during the winter of 2004/05 until the beginning of March 2005. 3 The term "null cases" should be interpreted as very low impact synoptic situations. 73 experiment reference cases were run real-time every day when possible. Setup experiment cases were completed while preparing for the OSSEs. Cases with some meteorological significance were chosen for the OSSEs. Specific cases will be referenced with the naming convention "daymonthyear" where month and year are abbreviated. For example, 12Oct04 is the case for the 12 October 2004. The REF12 forecast would have started at 12Z on 11 October 2004. The REFOO forecast will have started on OOZ on 12 October 2004. O S S E forecasts have D A assimilation on the 12 hr R E F 12 forecast valid at OOZ, the start time for the REFOO forecast. The case dates are listed in Appendix C. 8.4 OSSE Case Examples To illustrate the O S S E procedure, three cases will be analyzed in detail; the first two occurred during the Test Period, the third during the Study Period. For the sake of documentation brevity, no other detailed descriptions are included. Where graphs can clearly display the information, they will be the display vehicles of choice; where graphical displays would be too cluttered, tables will be used. For the Test Period cases, M S L P (kPa) forecast impact over W N A and E N A are presented. The REF12 forecasts are for 60 hrs; the O S S E forecasts for 48 hrs A D A and REFOO forecasts for 48 hrs. For the Study Period cases, HT70 forecast impacts were determined. The REF12 forecasts are for 84 hours, the O S S E forecasts for 72 hrs A D A and the REFOO forecast for 72 hrs. Most comparisons are restricted to the 60 hr (48 hr A D A ) forecast period or less, for reasons that will become apparent. 74 8.4.1 Test Period WNA Case: From 12 Dec @12Z to 15 Dec @00Z (2001) On 12 December 2001 an intense cyclone approached the B C southwest coast. The low deepened to 97.8 kPa as it came ashore, and maintained its intensity. Northwest winds behind the low reached storm force before subsiding on 14 December, and caused widespread damage and power outages to thousands of people. Commercial flights and ferry services were cancelled, schools were closed and wind-blown debris was everywhere (Chan 2002). Before using this case, we must confirm that the new 00Z analysis (from which we will extract data to create virtual soundings) is more accurate than the 12 hr earlier analysis starting the REF12 forecast. Table 8.1 shows that both B E N and REFOO score better (lower R M S D , higher COR) than R E F l 2 over W N A . This case passes the Pacific analysis update test. Valid Time Dec12 12Z Dec13 00Z Dec1312Z Dec14 00Z Dec1412Z Dec15 00Z WNA RMSD after data addition OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA REF12start at12Z 0.0000 0.1349 0.2457 0.4003 0.4333 0.4662 BEN start at 12Z 0.0000 0.0113 0.2075 0.2794 0.2858 0.2921 REFOO start at 00Z 0.0000 0.2321 0.3396 0.3060 0.2724 WNA COR after data addition OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA REF12start at 12Z 1.0000 0.9845 0.9655 0.9595 0.8777 0.7959 BEN start at 12Z 1.0000 0.9999 0.9726 0.9800 0.9487 0.9174 REFOO start at 00Z 1.0000 0.9740 0.9769 0.9546 . 0.9322 Table 8.1. Quality of the Pacific analyses for the 13Dec01 case, using root mean square difference (RMSD) and correlation coefficient (COR) as the metrics, for forecasts at various hours after data addition (ADA). Verification is for WNA in the VERwest region. Rows are for atmospheres with different amounts of soundings available (see text). The REFOO forecast starting at 00Z verifies better than the REF12 forecast starting at 12Z (12 hrs before). The BEN OSSE atmosphere, with the virtual radiosonde soundings added at 00Z also verifies better. From these statistics, it is concluded that the 00Z Pacific analysis was updated in a proper manner. 75 Figure 8.1a. Mean sea-level pressure (MSLP dark isobars every 0.4 kPa) from the REF12 atmosphere for the intense mid-latitude cyclone of 14 December 2001. Overlaid is the difference (DIF = VER-REF12) between the verification atmosphere and the reference atmosphere. DIF contours every 0.1 kPa are thin dashed for positive and dotted for negative differences, (a) Valid 00Z 13 Dec showing the 12 hr forecast. A 99.8 kPa storm near 50N 140 W is moving into a long wave trough positioned off the BC coast, with MSLP values that are already 0.5 kPa too high. Figure 8.1b. Valid 00Z 14 Dec showing the 36 h forecast. Forecast cyclone movement is too slow. DIF values (-0.9 kPa over BC, and +0.9 kPa west of Oregon) indicate that the actual low center has moved onshore. For REF12, Figures 8.1 shows both M S L P and pressure difference (DIF) between the verification and reference atmospheres for the 12, 36, and 60 h forecasts. At the start of 76 this period, the 99.8 kPa cyclone near 50N H O W was moving into a long-wave trough off the B C coast. At OOZ 13 December, forecast MSLPs west of B C were already 0.5 kPa too high. Figure 8.1c. Valid OOZ 15 Dec showing the 60 hr forecast. This forecast is quite poor, with MSLP ridges and-troughs out of phase as indicated by the large positive and negative DIF values over BC/Washington and the NE Pacific, respectively. Figure 8.2a. Effect of adding ErnPAC_RBS ThreeBuoy vROBs to the otherwise Pacific data-void forecast for the 13Dec01 cyclone. Plotted are MSLP (dark isobars every 0.4 kPa) and the MSLP difference (DIF, thin lines every 0.1 kPa dashed for positive and dotted for negative) between the PAC_RBS atmosphere and the REF12 atmosphere. Valid OOZ 13 Dec at 12 hr into the forecast, at the time of RBS data addition. DIF fields show that the cyclone center is better captured, with more accurate (0.7 kPa lower) central pressure and better location (further west). 77 Figure 8.2b. Valid OOZ 15 Dec at 60 hr into the forecast (48 hr after vROB data addition), showing only the MSLP DIF field for clarity. The improvements due to the DA of vROBs over the eastern Pacific have moved over the continent, and indicate that the forecast better captures a deeper low that moves east faster. Meanwhile, near zero DIF values over the ocean now indicate that data-void effects from further upwind have moved over the NE Pacific. Figure 8.2c. Close-up of the western N. America verification area valid OOZ 15 Dec (60 hr forecast, 48 hr ADA), but showing the improvement in RMSD associated with the addition of the ThreeBuoy vROB data. Plotted is the difference between (VER - REF12) RMSD minus (VER-PACRBS) RMSD, where positive values (dashed contours) indicate RMSD improvement in the OSSE forecast compared to the REF12, and negative (dotted contours) indicate degradation. Evident is a region of improvement over BC, WA and ID, and a dipole region of improvement and degradation across the Rocky Mountain crest in the northeastern portion of the figure. 78 For an O S S E identical run, vROBs for ThreeBuoy P A C _ R B S (Figure 7.3) are added at 13 December at OOZ. Compared to the R E F 12 forecast, the approaching surface low has deepened 0.5 kPa (Figure 8.2a), which is in the proper sense. M S L P difference (DIF) in Figure 8.2b shows that the 60 hr (48 h A D A ) P A C _ R B S forecast verifies better than R E F 12 (Figure 8.1c) in the VERwest area. A close up is shown in Figure 8.2c. Adding more vROBs reduces R M S D forecast error, as shown in Figure 8.3 and Table 8.2 for the P A C R B S group. Compare these R S B O S S E results to opposite extremes of an unfilled data void (REF 12, top curve) and a completely filled data void (soundings to 16 km altitude over all land and ocean areas; B E N , bottom curve). 0.50 00 h ADA 12 h ADA 24 h ADA 36 h ADA 48 h ADA R U N P E R I O D A F T E R D A T A A D D I T I O N ( A D A ) Figure 8.3. Growth of forecast errors with time after data addition (ADA) of various numbers of RBS sounding locations, for the 12 Dec storm. Forecast error metric is root mean square difference (RMSD) between the OSSE forecasts and corresponding VER analyses, valid at the same time, and verified in the VERwest region. Initialization is from 12Z on 12 Dec, with data addition at OOZ on 13 Dec. Smaller RMSD is better. 79 Figure 8.4a shows the end result (48 hr A D A ) for each group member. If B E N is a benchmark result from a balanced observing network, then the SixBuoy array provided a G A I N of 0.65 (Figure 8.4b) for this case, while the ThreeBuoy array provided the most G A I N per buoy (Figure 8.4c). GROUP PAC RMSD RESULTS - Starting at 13 Dec @ 01 DZ • OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA REF12 0.1349 0.2457 0.4003 0.4333 0.4662 ThreeBuoy 0.1349 0.2310 0.3581 0.3729 0.3876 SixBuoy 0.1349 0.2287 0.3452 0.3495 0.3538 NineBuoy 0.1349 0.2265 0.3402 0.3570 0.3737 TwelveBuoy 0.1255 0.2389 0.3236 0.3376 0.3516 CenPAC Array 0.1349 0.2430 0.3707 0.4117 0.4526 AllBuoy 0.114 0.2384 0.3067 0.3133 0.3198 BEN 0.0113 0.2075 0.2794 0.2858 0.2921 Table 8.2. The 13Dec01case. MSLP RMSD is shown as the forecast error for the various PAC Group configurations. 0.50 -i Figure 8.4a. Changes in 60 hr forecast quality of MSLP measured over the continental VERwest region, for different numbers and arrangements of sounding buoys upstream over the Pacific, for the Dec 13 storm. Various metrics of forecast quality valid at 00Z on 15 Dec (48 hr ADA) are used: (a) RMSD of the OSSE forecasts vs. the REFl2 forecasts valid at the same time, where smaller RMSD is better. 80 The CenPAC R B S array (Figure 7.3) captured a non-growing feature for this case, resulting in minimal forecast improvement downstream over VERwest. The C O R results shown in Figure 8.5 support these conclusions. Figure 8.4b. Same as (a), but normalized to give the relative gain in accuracy, where GAIN = 0.0 for forecasts started from the REF12 background-state atmosphere of no soundings added, and GAIN =1.0 for the idealized scenario of a dense array of added soundings covering all of the Pacific (BEN). Larger GAIN is better. Fi gure 8.4c. Relative GAIN per sounding buoy, where the data from (b) is divided by the number of buoys. Larger GAIN-per-buoy is better. 81 1.00 0.95 PS O 0.90 3 0.85 0.80 0.75 —•— REF12 *- ThreeBuoy —A— SixBuoy —X— NineBuoy —*— TwelveBuoy CenPAC Array —+— AHBuoy BEN 00 h ADA 12 h ADA 24 h ADA 36 h ADA 48 h ADA Figure 8.5. The effect of increasing numbers of sounding buoys on M S L P forecast quality, for different forecast durations for the 13Dec01 case. C O R is the spatial correlation of the OSSE forecast with the corresponding verification analysis (VER) valid at the same time. The forecast initialization time was 12 Dec at 12Z, and data addition occurred on 13 Dec at OOZ. All values for 36 hr after data addition (ADA) are interpolated, because the windstorm caused a power outage and a computer failure resulting in the missing verification data. Larger C O R is better, with C O R = 1 representing a perfect forecast. The PACvoid RBS experiments better simulate an operational OOZ analysis update by adding both vROBs over ocean ( P A C R B S ) and vRAOBs over North America (PACvoid, see Figure 7.3) to form a PACvoidRBS array. Figures 8.6 show MSLP forecast results 12 hr and 48 hr A D A , with superimposed difference isopleths (DIF) between ThreeBuoy P A C v o i d R B S and PACvoid. The ThreeBuoy atmosphere has made most of the relative gains. Figure 8.7 has RSMD values 48 hr A D A for each PAC_RBS and P A C v o i d R B S array scenario. 82 Figure 8.6. Forecasts of MSLP (solid lines every 0.4 kPa) for the PACvoid_RSB atmosphere, where vROBs up to the mid troposphere were assimilated at the ThreeBuoy ErnPAC_RBS sites, and vRAOBs at every fifth grid point (i.e., 225 km horizontal spacing) up to 16 km altitude were retained over land. Dashed and dotted contours represent the positive and negative difference (DIF) between the REF12 and the ThreeBuoy PACvoidRBS atmosphere in 0.1 kPa units, (a) Valid at 00Z on 13 Dec 2001, (12 hr forecast, 0 hr ADA). Comparison with the REF12 atmosphere of Figure 8.1a shows improvements, with a weaker ridge over BC and a sharper trough along the coast, (b) Valid at 00Z on 15 Dec 2001 (60 h forecast, 48 hr ADA). Comparison with the 60 hr REF12 forecast (Figure 8.1c) shows improvements due to RBS data addition; namely, the low over Alberta is located further east, and the ridge over eastern BC is stronger. 83 EFFECT OF ADDING VIRTUAL RADIOSONDE DATA 0.50 o! g 0.45 K H 0.40 > Q 0.35 I 0.30 0.25 Figure 8.7. Impact of adding/denying vRAOBs over land to a scenario where vROBs would already exist over the Pacific. Forecast errors (RMSD, ordinate) of MSLP over VERwest are plotted for various numbers of sounding buoys (abscissa), with all results for 48 hr ADA in the 13Dec01 storm. The PACRBS curve has only sparse vROBs over the Pacific (with no soundings over land), while the PACvoid_RBS curve is for the sparse RBS soundings over the Pacific plus vRAOBs spaced 225 km apart over land. For the VERwest region, RBS soundings were more important than continental soundings for this storm. Smaller RMSD is better. 8.4.2 Test Period ENA Case: From 20 Feb @12Z to 23 Feb @00Z (2002) In this event, a cyclone over the northeast Pacific moved towards a ridge over BC. This cyclone brought heavy rain to the west coast, but filled rapidly after landfall, before moving over the VEReast area. 7.4). Similar to the W N A case, RBS-equivalents are added west of the shifted interface, and virtual radiosonde soundings are added east. These OSSEs take advantage of the reliable data between the idealized W N A data-void interface and the real west coast of North America, and is equivalent to a data-denial OSSE west of the new data-void For this OSSE the data-void interface is moved 20-25 degrees eastward (Figure interface. 84 0.50 0 0 h A D A 12 h A D A 24 h A D A 36 h A D A FORECAST PERIOD Figure 8.8. Increase in forecast errors with time similar to Figure 8.4, but for the 20Feb02 case of Figure 8.9, with a virtual coastline shifted east (Figure 7.2) vROBs are added at the W N A R B S locations of Figure 7.5. Results are verified in the VEReast domain of Figure 7.7. Smaller RMSD is better. Results show that the effects of a poor Pacific analysis update have moved over the VEReast area. Results show the forecast over VEReast improving after adding W N A _ R B S vROBs, but only until 24-36 h A D A (Figure 8.8). Afterward, the forecast deteriorates as the effects of a poor Pacific analysis moves into the verification area. Figure 8.9 shows the 12 hr WNA_RJ3S ThreeBuoy forecast of M S L P , and its difference from REF12 (which has no additional soundings). The REF12 forecast filled the low too slowly. Initial improvement in the W N A R B S simulation due to the three additional vROBs later deteriorated as the influence from the real Pacific data-void moved over the verification area. Unfortunately, the subsequent Pacific update resulted in a worse REFOO 36 hr forecast. The situation was similar to the series-of-busts cases described by Doyle 85 (2001) and Hacker et al. (2003). The R M S D results (Figure 8.8) show substantial improvement over R E F l 2 , until deteriorating 36 h A D A , suggesting that the E T A analysis over the Pacific was not updated properly. Figure 8.9. OSSE result for the 20Feb02 storm, but where the virtual west coast of North America is shifted 20+° eastward. ThreeBuoy vROBs are added at the WNA_RBS locations of Figure 7.5, with vRAOBs retained covering the remainder of the continent east of the shifted coastline (Figure 7.2). Shown are MSLP (solid contours) for a 12 hr forecast valid at 00Z on 21 Feb 2002. Superimposed are difference DIF contours (WNARBS - REFl 2) of the OSSE forecast from the verifying analysis. Although positive DIF values (dashed) appear where the RBS soundings were added, negative values (dotted) west of the BC coast corresponds to deepening of the approaching low, which is a change in the wrong sense. 8.4.3 Study Period WNA Case: From 15 Oct @12Z to 19 Oct @00Z (2003) The Pacific winter westerly jet routinely steers weather systems to the northwest coast of North America, resulting in rain and associated weather. Sometimes the airflow will be unusually warm and saturated, resulting in extreme rainfall. Before the advent of routine satellite images (early 1980s), the reasons for these extreme events remained a mystery to researchers and operational forecasters alike. Examination of the meteorological charts would bring little insight into the nature of the responsible 86 mechanisms. The only exception could have been the examination of R A O B s taken on the W N A coast; the mid to high-level airmass potential temperature during the extreme rainfall event would be unusually high. The problem with the R A O B indicator was the timing. B y the time the R A O B s were available, the extreme event was well underway. Upstream advance indicators (RAOBs) did not exist (Section 3.1). B y the mid 1980s, forecast centers were being supplied with satellite images from geostationary satellites. The US National Oceanographic and Atmospheric Administration ( N O A A ) Geostationary Operational Environmental Satellites (GOES) provided coverage over North America and the eastern Pacific. The Japanese Geostationary Meteorological Satellite (GMS) provided coverage over the western Pacific. The views from the geostationary satellites showed that that the airflow causing the extreme events had a tropical origin, nicknamed "Pineapple Express" by the media. Sometimes the airmass for the extreme event was the remnants of tropical storms and/or hurricanes that were caught up in the western Pacific westerlies. Sometimes hurricanes and tropical storms produced the connection between the tropical easterlies and mid-latitude westerlies. Portions of the following description were based on a presentation made by McCollor (2004). In mid-October, 2003, a quasi-stationary high-pressure cell west of Hawaii resulted in a connection between the inter-tropical convergence zone and the mid-latitude westerly jet (Figure 8.10) in the western Pacific region. After 13 October 2003 the normal Pacific westerly zonal flow carried copious amounts of moisture. The leading edge of the tropical moisture reached the W N A coast on 16 October, turning a normal westerly flow into an extreme event producing heavy rainfalls. 87 Figure 8.10. GOES water-vapour image composite for the Pacific. The white regions indicate large moisture content in the mid to upper troposphere. The images shows the connection between the intertropical convergence zone at about 30N 150E and the mid-latitude westerly flow at about 45N. This image was produced by NOAA and posted on their web site. On Wednesday morning, 15 October 2003, the Pacific Weather Centre 4 issued a forecast of "rain tonight" and a heavy rainfall warning for southwestern British Columbia for the next day (Thursday) providing a 24 hr lead-time. The 24 hr amounts included with the warnings were "35 to 45 mm". From Portland, Oregon to Port Hardy, B C , normal west-coast rain began in the evening as forecast. Near mid-day Thursday, the tropical component of the flow reached the coast, and hourly rainfall rates skyrocketed, an event that was not well forecast. Thursday October 16th was the heaviest rain day over the B C south coast (Table 8.3). Hourly rainfall rates reached as high as 18 to 24 rnm/hr in the B C 4 The Pacific region office of the Meteorological Service of Canada, Environment Canada located in Vancouver, BC. 88 Lower Mainland. Twenty-four hour rainfall totals included 86 mm to 141 mm in the population centers and 185 mm to 200 mm at some of B C Hydro's river-basin-based Data Collection Platforms, two to four times the amount forecast. Daily Rainfall (mm) WED 15 Oct THU Oct 16 FRI Oct 17 SAT Oct 18 SUN Oct 19 MON Oct 20 TUE Oct 21 7 DAY TOTAL BC Location Port Renfrew 86 199 106 67 55 52 4 568 Squamish 17 102 137 79 51 86 5 476 Cowichan Forestry Station 62 118 99 44 36 51 4 414 Roberts Creek 40 68 112 70 37 16 3 345 UBC Research Forest, Maple Ridge 44 146 60 19 11 52 1 333 Port Alberni 8 77 67 123 1 50 0 326 Malahat 16 160 36 1 17 82 6 317 Agassiz 4 99 87 0 17 69 31 308 Sheringham Point Lighthouse 14 140 47 5 24 53 11 293 Tofino 60 72 90 43 17 5 M 287 Hope 3 72 102 0 16 64 28 283 Victoria Airport 13 136 33 0 15 73 8 279 Pitt Meadows 12 137 74 0 15 22 15 275 Abbotsford 12 94 74 0 12 51 25 267 White Rock 12 119 57 0 7 30 15 239 West Vancouver 3 98 88 5 9 28 1 231 Whistler 0 77 71 42 15 16 ? 222 Vancouver Airport 13 85 56 1 6 25 4 190 Nanaimo M 54 34 10 25 13 TR 137 Powell River 23 27 42 26 14 1 1 133 Campbell River 32 22 28 17 26 2 1 129 Comox 14 37 16 33 0 21 TR 121 Table 8.3 Rainfall data, 15-21 October 2003. The data was provided by Environment Canada. On Friday October 17, the heavy rain band drifted northward and continued across central Vancouver Island, and southwestern B C . Since this was very much a mid-level atmospheric event, the Coast Mountain range was ineffective in its typical role as a Pacific moisture barrier. On Saturday October 18 th, the heavy rain moved across the B C interior regions to eastern B C . This heavy rainfall event in southern B C ended up as the biggest 89 recorded rainstorm as measured by rainfall storm totals, and the widespread extent of heavy rain and floodwaters. In the following five days, from Sunday the 19th through Thursday the 23rd, a series of three distinct Pacific frontal systems, each carrying remnant tropical moisture, crossed southern B C bringing periods of moderate to heavy rain. People were evacuated from their homes; road and rail lines were washed out; bridges were lost (Figure 8.11) and motorists lost their lives. Figure 8.11. Rutherford Creek between the villages of Pemberton and Whistler, BC Floodwaters washed out the bridge on October 18th near 4 a.m. Three cars drove off the bridge into the river before the washout was noticed. Lives were lost. Photo courtesy of McCollor (2004). The main question here is whether the RBS system would have been able to improve the rain accumulation forecasts and tropical rainstorm timing. The critical forecast was the onset of the rain event that happened on 15 October. The R E F l 2 forecast started 14 October 12Z would have had the start of precipitation within the tomorrow (second day) 90 forecast. The TwelveBuoy P A C J R S B forecast successfully brought the precipitation in sooner than the R E F 12 forecast (Figures 8.13a & 8.13b). B y 17 October OOZ, the TwelveBuoy forecast accumulations were more realistic, up 50% more than the amounts from REF12 (Figure 8.12c) and were still up to 20% higher by 17 October 12Z (Figure 8.12d). 127U/ 126.5W 126UI 125.5UI 125W 12+.5W 12+W 123.511/ ' 123UJ 122.5W 122U/ 121.511/ 12111/ 120.511/ 12DU/ 91 1 • I 1*1 I I ^ ILA^ n -~ • | , I ^ < ^ ^ ^ f _ > j 127W 126.5UI 12BU1 125.5W 125UI 12+.5W 12+UJ. 123.5W 123W 122.5W 122W 121,5W I21W 12Q.5U/ 12DW Figure 8.12. (c) 60 hr rainfall accumulations valid 17 Oct 00Z. (d) 72 hr rainfall accumulations valid 17 Oct 12Z. The solid lines are the rainfall accumulations for the REFl2 forecast starting at 15 Oct 00Z. The dashed lines are the added rainfall accumulations predicted by the TwelveBuoy P A C R S B forecast starting at 15 Oct 00Z. The TwelveBuoy RBS forecast advanced the rain faster and resulted in higher early accumulations closer to the actual event. The intensifying westerly jet did not make a significant penetration into the VERwest area until 16 October 2003. The significant events took place in the mid to upper levels of the troposphere, so it is appropriate that the HT70 field be used for verification. Figure 8.13 shows the M M 5 12 hr HT70 kPa forecast valid 16 October 2003 at 00Z. The dashed lines are the DIF from the V E R field, the next analysis. The dashed lines mean that 92 at HT70 kPa the heights over the Pacific and W N A were forecast as much as 30 to 50 m too high, suggesting that the low-pressure area near 5 ON 145W was deeper and the westerly stream to the south was more intense. Table 8.4 shows that every one of the ErnPAC_RJ3S arrays was successful in lowering the 70 kPa forecast heights over W N A 36-48 hrs A D A . Figure 8.13. The dark lines are the 70 kPa 12 hr forecast height contours valid 16 October OOZ by the REF 12 forecast started at 15 October 12Z. The dashed lines are the DIF between VER and the REF 12 forecast. The negative values show that the forecast was 20 to 50 m too high in the vicinity of the low near 50N 145W and over WNA. VER was the ETA 00Z analysis that was probably analyzing heights too high in the eastern Pacific since EDAS was slow to react to actual events. Still the heights were not low enough. The primary reason was probably because the vROBs taken from the E T A OOZ analysis were not accurate enough. The F D D A E D A S cycle was not able to capture the extreme situation and make accurate updates. A real 93 P A C _ R B S array would have been able to better capture the intensity of the westerly jet and discern the tropical airmass with its high potential temperatures. GROUP PAC RMSD Octl6 Octl6 Octl7 Octl7 Octl8 Octl8 Octl9 OOZ 12Z OOZ 12Z OOZ 12Z OOZ OOhr 12hr 24hr 36hr 48hr 60hr 72hr ADA ADA ADA ADA ADA ADA ADA REF12 17.9 19.0 25.0 25.7 19.3 27.8 23.4 ThreeBuoy 19.9 17.3 24.2 19.1 15.7 25.5 23.7 SixBuoy 19.3 16.0 21.5 17.0 15.2 23.6 21.1 NineBuoy 19.3 15.4 21.6 16.7 14.6 22.6 20.0 TwelveBuoy 19.4 15.1 21.6 17.0 14.1 21.9 20.3 CenPAC Array 20.5 17.8 29.3 23.1 16.6 26.4 26.0 AllBuoy 19.6 15.9 25.3 20.0 14.8 23.9 22.1 BEN 7.3 15.0 26.3 18.7 12.8 23.2 31.3 REFOO 0.0 9.7 15.5 16.5 19.1 19.1 19.3 GROUP PAC GAIN REF12 0.00 0.00 0.00 0.00 0.00 0.00 ThreeBuoy 0.42 0.08* 0.95 0.56 0.50 -0.07* SixBuoy 0.75 0.37* 1.26 0.63 0.90 0.56* NineBuoy 0.91 0.36* 1.29 0.72 1.13 0.83* TwelveBuoy 0.99 0.36* 1.26 0.80 1.29 0.76* CenPAC Array 0.31 -0.45* 0:38 0.41 0.31 -0.63* AllBuoy 0.78 -0.03* 0.81 0.70 0.84 0.32* BEN 1.00 -0.14* 1.00 1.00 1.00 -1.93* Table 8.4. The REF12 forecast was started 15 October 12Z. The PAC Group vROBs were added on 16 October 00Z, 12 hrs later. The rainstorm moved onshore between 16 October 12Z and 17 October OOZ. The PAC Group vROBs were assimilated with MM5-3DVAR. The 72 hr ADA results are erratic. *At times, adding virtual data makes the forecast results worse for certain time periods. An explanation is offered in Section 9.6.4. When BEN RMSD is greater than REF 12 RMSD, REFOO is substituted as a higher standard. Even though the atmosphere never repeats itself exactly, another extreme rainfall episodic event occurred over the same area on 16-19 January 2005. A similar link between the tropics and the westerly zonal flow was established streaming warm tropical moisture-laden air towards Washington State and British Columbia. Floods and evacuations occurred, mudslides destroyed homes, roads were washed out, extreme avalanche hazards 94 closed highways, and people died. The parallels to the 17-21 October 2003 episode were uncanny. 95 Chapter 9 OSSE General Findings 9.1 Foreword Appendix C lists the dates used for the Test-Period and Study-Period OSSEs. A large number of statistics have been compiled, but only the highlights will be presented. R M S D statistics have a greater range than C O R statistics, and are preferred. In the discussions, a reference to " R M S D results" should be interpreted as term synonymous with "forecast error". To keep the document short, the C O R statistics will not be presented. W N A Group results were used for further confirmation of the P A C Group results and will not be presented, except in special circumstances. The number of cases used for each set of experiments varies greatly. The Test Period focuses on M S L P R M S D results for a limited number of high impact winter events, similar to the Section 8.4.1/8.4.2 cases. The Test Period results were used to refine the experimental design and will be presented as such. Mostly, Study-Period summaries feature averaged HT70 results from a wide spectrum of cases, resulting in subdued effects when compared to Test-Period results. When winds are a specific issue, WS70 R M S D metrics will be used. When a reference is made to R M S D trends relative to a perfect forecast ( R M S D = 0.0), a % figure will be used (e.g. 23%). When a reference is made to R M S D trends relative to a reference or benchmark forecast, a G A I N figure will be used (e.g. 0.23). Occasionally, the G A I N calculations are erratic. Minus GAINs will be allowed to show 96 negative results, but only to a limit -1.0; lower GAINs will be flagged with a <-1.0. GAINs greater than 1.0 will similarly be flagged with a >1.0. Even though the Study Period N W P forecasts run to 84 hrs, the results up to a 60 hr reference forecast or a 48 hr A D A OSSE forecast will be considered the reliable forecast length. This will make the Study-Period results comparable to the Test-Period results. The results of the 60-72 hr (48-60 hr A D A ) forecast period are viewed with caution. The 84 hr (72 A D A ) forecast results are sometimes erratic and are not viewed with confidence. However, they are still presented for academic reasons including the need to determine the cause of the poor forecast skill. Other reasons for the 60 hr threshold will become obvious. The winter season spans two portions of successive calendar years starting from the beginning of October to the end of April of the following year. For example, the first Study-Period winter spans October 2002 to the end of April 2003. This winter period will be referred to as winter 02/03. The first Study-Period summer spans May to September of 2003 and will be referred to as summer 03. When the date is number coded, the format is Y M M D D H H , where Y is the single digit year1, M M is the month, D D is the day and H H is the hour (either 00 or 12). If the prefix (Em or Cen) is omitted when referring to the R S B arrays of Figure 7.3, it should be assumed that the array belongs to the E r n P A C _ R B S array. For example SixBuoy P A C R B S refers to the SixBuoy RBS array within the E r n P A C R B S sector. The prefixes will be maintained when both the ErnPAC_RBS and the CenPAC_RBS arrays are being discussed. 1 A l l years are 200Y. The "200" is left out to save space in the table cells. 97 9.2 EDAS 9.2.1 EDAS Failures For several reasons, it is expected that the forecast from the OOZ initial field will verify better than the previous 12Z forecast valid at the same time. The E T A OOZ initial field has had 12 more hours of updated information and the benefit of four more E D A S cycles. If the previous REF12 forecast verifies 5% better than the same valid time REFOO, then REFOO is considered to be an analysis and forecast failure. The 5% threshold is arbitrarily chosen to set the criteria above the noise level. Forecast failures occurred in 15% of the cases tested. Naturally, the percentage of forecast failures would vary with the threshold setting. Examples are shown in Table 9.1. The same effect is seen with the E T A 1 2 versus the ETAOO forecasts. Intuition would suggest that the primary reason for the intermittent VERwest forecast failures are the Pacific data void and the inability of E D A S to consistently mitigate the effects of the Pacific data deficiency. Most of the time the failure occurrences are intermittent. At times a series of failures occur. A n E T A 3 hr forecast is used as the first guess field for the E D A S 3 D V A R - D A . When the first-guess field is bad, it is difficult for the E D A S to catch up to the real events (Sections 3.4, 8.4.3); real data are thrown out, details are smoothed over, and a series of busts can occur. Such a series is shown in Table 9.1. The next forecast is worse or not much better than the preceding one; E D A S (and the subsequent M M 5 forecasts) have not caught up to the real events. The thicker solid arrows follow the same time group. Even with the benefit of successive E D A S updates, the successively shorter-length forecasts do not get better. 98 A S E R I E S O F D A I L Y R E F E R E N C E R U N S - H T 7 0 R M S D (m) VERwest 4020700 4020712 4020800 4020812 4020900 4020912 4021000 REF12 15.5 15.8 31.2 4"0^8\ 29.8 35.6 ^ REFOO 0.0 13.2 27.2 ^ 42.7 "~ 2 4 ^ 4 3 ^ 4020800 4020812 4020900 4020912 4021000 402*012 4021100 REF12 19.1 17.7 24.8 39.6 58.3 23.9 REFOO 0.0 16.5 35.8 ^ 26,0 34.8 39.6 4020900 4020912 4021000 4021012 4021100 4021112 4021200 REFl 2 26.4 19.2 35.0 27.8 26.9 19.5 28.7 ^ REFOO 0.0 19.0 34.3 28.0 25,8 22.4 32$ 4021000 4021012 4021100 4021112 4021200 4821212 4021300 REF12 20.1 22.5 18.2 20.5 20.8 34.5 46.3 REFOO 0.0 16.3 17.6 18.6 34.9 40.8 4021100 4021112 4021200 ^ 4flll212 4021300 4021312 4021400 REFl 2 15.3 15.0 2 7 . 8 ^ 32.4 34.5 42.3 40.5 REFOO 0.0 18.1 32*3 37.7 39.0 35.4 23.7 Table 9.1. REFl2 and REFOO HT70 RMSD results every 12 hrs of the forecast duration. Lower RMSD is better. The REF12 results start with the 12 hr forecast. The same valid time REFOO results start with the initial time. The 07-11 February 2004 forecast failures showing a "series of busts". The heavier solid arrows follow the same time group. REFOO is worse than the REFl2 even though REFl2 is 12 hrs longer. Heavy downward sloping lines show that successively shorter-length forecasts do not get better. The shorter downward sloping lines compares REF12 and REFOO for the same forecast length. REF12 is often markedly worse than REFOO. An interesting observation was noted when viewing the synoptic charts; the WNA atmospheric flow was shifting from a zonal flow on 07 February 2004 to a meridional flow with a major change occurring on 10 February 2004. 9.2.2 EDAS Value E D A S - 3 D V A R is able to assimilate the data shown in Table 3.1 as well as a large number of satellite radiances. Between 12Z and 00Z, E D A S has four complete D A cycles. To estimate the added cumulative value of the observations and the F D D A within the 12 hr period, the results of REFOO forecasts were compared with forecasts started from the same valid time E T A 12 12 hr forecasts (EFG, as E T A First Guess). The E T A 12 12 hr forecast would be analogous to the first-guess field. The REFOO OOhr field would be equivalent to the fully updated analysis valid at the same time. The results for 143 cases are shown in 99 Table 9.2. The averaged REFOO HT70 R M S D improvements over E F G at 36 to 60 hrs are between 10-15% for both VERwest and VEReast. Table 9.2 also includes the results of ETA00 compared with ETA12 . After a 12 hr F D D A with four complete cycles, E D A S is able to improve the VERwest forecast averages by about 20%. If 15-20% represents the maximum averaged improvement that can be achieved by a powerful F D D A system like E D A S , then it should form the upper bound on what can be expected from the M M 5 OSSEs. EDAS VALUE - HT70 RMSD (m) VERwest MM5 - 143 cases 12hr 24hr 36hr 48hr 60hr 72hr EFG (starting with the ETA12 12 hr forecast) 16.0 19.5 26.7 28.5 35.6 38.0 REFOO 16.4 17.3 23.2 24.5 26.9 29.8 % Improvement of EDAS update -2% 12% 13% 14% 24% 22% VEReast MM5 - 143 cases EFG 15.3 19.2 26.2 28.7 33.8 37.3 REFOO 12.9 16.6 22.2 25.4 28.6 30.7 % Improvement of EDAS update 16% 13% 15% 12% 15% 18% VERwest ETA - 143 cases ETA12 12.8 16.1 20.2 23.1 28.3 30.0 ETA00 12.3 12.9 15.9 18.3 22.7 25.8 % Improvement of EDAS Update 3% 20% 21% 21% 20% 14% VEReast ETA - 143 cases ETA12 12.3 15.7 20.7 23.5 30.2 35.9 ETA00 10.2 13.4 18.3 21.5 27.5 29.5 % Improvement of EDAS update 17% 15% 12% 8% 9% 18% Table 9.2. ETA First Guess (EFG) is the MM5 forecast starting with the ETA 12 hr forecast. REFOO is the forecast from the subsequent ETA OOZ EDAS update. A comparison estimates the average value that EDAS has provided. The ETA12 forecast also carries on from its own 12 hr forecast. The average improvement for the same 143 cases is shown. The EFG VERwest improvement is comparable to the EFG VEReast improvement. The ETA 12 VERwest improvement is higher than the ETA 12 VEReast improvement. Lower RMSD represents a better forecast. For each comparison (REFOO vs E F G , E T A 0 0 vs ETA12), the model error remains the same, the weather regime remains the same and the effects of the boundary conditions 100 remains the same. The only difference is the initial conditions. The absolute M M 5 O S S E forecast improvements will be scaled by this 20% objective. Various intensive targeting experiments like the W S R (Section 3.10) have shown forecast improvement results at about 10-20% (Toth 2004). 9.3 RBS NWP Support Test Period results suggested that 6 km profiles would deliver most of the R B S benefit (discussed later). Based on this preliminary result, an R B S engineering constraint of 200 6-km rocketsondes per buoy payload was established (Chapter 4). A research objective is to design an operational program able to maximize the utility of these resources. 9.3.1 Synoptic Hour support With a capacity of 200 rocketsondes, the R B S can routinely provide only one rocketsonde launch per day for part of the year. It is logical that the launches should coincide with one of the main synoptic hours (OOZ or 12Z). It is necessary to determine which of the two main synoptic times is optimal. On average, the same-forecast-length R E F 12 R M S D results are worse than the REFOO. Examples are shown in Table 9.1. The shorter downward lines cover two successive forecasts of the same length. Averaged winter Test-Period results are shown in Table 9.3. The REFOO 48 hr forecasts are 12-19% better than REF12 48 hr forecasts. The more varied case Study-Period results are shown in Table 9.4. Even though many null cases are included in the average, the VERwest OOZ forecasts still provide better results than the 12Z forecasts. The VERwest 6-12% REFOO to REF12 difference approaches half 101 the 15-20% improvement possible (established in Section 9.2.2). Though subdued, the same effect is seen with the VEReast results. Published research on this VERwest "diurnal effect" seems to be lacking, so several theories will be advanced. The effect could be the result of a diurnal varying accuracy of the Pacific initial conditions; this theory would explain why the effect is seen mostly with the VERwest results. The merchant ships providing observations have a skeleton staff on overnight duty; there are more Pacific ship reports at OOZ than at 12Z .. It is easier to generate GOES cloud-drift winds from the visible imagery than from the IR, so the Pacific daylight OOZ scenes provide more SATOB observations (Section 3.1) than nighttime 12Z. Another reason may be the greater convective activity at OOZ; the physics contained in the OOZ observations may be more complete than at 12Z. T E S T PERIOD M S L P RMSD (kPa) VERwest - 5 cases 24hr 36hr 48hr 60hr REF12 0.214 0.288 0.349 0.400 REFOO better than REF12 +5% +12% +12% REFOO 0.204 0.256 0.312 VERwest average 0.209 0.272 0.331 0.400 VEReast - 5 cases 24hr 36hr 48hr 60hr REF12 0.176 0.235 0.303 0.349 REFOO better than REF12 +4% +9% +19% REFOO 0.170 0.215 0.255 VEReast average 0.173 0.225 0.279 0.349 PACIFIC D A T AVOID P E N A L T Y VERwest RMSD compared to VEReast RMSD +21% +21% +18% +15% Table 9.3. VERwest and VEReast RMSD averages the 5 winter 01/02 cases. REFOO verifies better than REF 12 suggesting a diurnal effect. The VEReast forecasts verify better than the VERwest forecasts, indicating the effects of the Pacific data-void. 2 This was a personal observation while working at the Pacific Weather Centre, Environment Canada, Vancouver 1975-1996. 102 It can be concluded that the Pacific 12Z analyses requires the most help, and would benefit most from R B S soundings. N W P forecasts starting from 12Z initial fields are usually produced too late to help day-1 VERwest forecasts. The 12Z VEReast local time is later. If the objective is to enhance day-1 forecasts, then the preferred launch time is 00Z. If the objective is to enhance day-2 forecasts, then the proximity to the day-2 morning and the poorer verification scores suggest that the R B S launch times should be at 12Z. STUDY PERIOD - HT70 RMSD (m) VERwest - 375 cases 12hr 24hr 36hr 48hr 60hr 72hr 84hr REF12 15.9 20.5 24.9 29.0 31.9 36.6 38.4 REFOO 16.4 19.4 23.3 25.7 29.9 32.8 REFOO better than REF12 +5% +7% +12% +6% +11% VERwest Average 16.2 20.0 24 A 27.4 30.9 34.7 38.4 VEReast - 375 cases 12hr 24hr 36hr 48hr 60hr 72hr 84hr REF12 13.3 17.4 21.5 28.0 33.2 41.2 45.4 REFOO 13.8 16.4 21.7 26.1 33.3 38.7 REFOO better than REF12 +6% -1% +7% +0% +6% VEReast Average 13.5 16.9 21.6 27.1 33.2 40.0 45.4 PACIFIC DATA VOID PENALTY VERwest RMSD compared to VEReast RMSD +20% +18% +12% +1% -7% -13% -15% NWP FORECAST ERROR Averaged VEReast and VERwest Forecast Error 14.8 18.4 22.9 27.2 32.1 37.3 41.9 Table 9.4. The Study Period VERwest HT70 RMSD results show that the forecasts started at 00Z are better than the ones started at 12Z by 6 to 12%, which is almost half of the forecast improvement possible. Over VEReast, the effect is less pronounced. VERwest starts with a nearly 20% penalty over VEReast due to the Pacific data void. The penalty reverses itself after the 48 hr forecast period due to other influences (weather regime?). 9.3.2 Summer Period Denial For the Test-Period VERwest winter 01/02 cases, M S L P R M S D of about 0.2 kPa for 24 hr forecasts doubles to about 0.4 kPa for 60 hr forecasts (Table 9.5). The summer 02 R M S D results are less than half the winter values, showing that weaker synoptic systems in 103 summer are easier to forecast, even with lack of data in the Pacific data void. The Study-Period winter HT70 RMSD results are 18-25% higher than summer. The differences are not as large as the Test-Period differences for the following reasons: HT70 is used as a metric rather than MSLP; the averages are from a larger spectrum of cases, not just high impact events; the WNA Study Period winters were relatively quiet. Still, the difference is comparable to the maximum benefit that EDAS (and possibly the RBS virtual-data OSSEs) can routinely provide. If the winter NWP forecast error could be reduced to the summertime level by observations from the RBS and other THORPEX systems, it would be a significant achievement. TEST PERIOD - SUMMER vs WINTER - MSL RMSD (kPa) VERwest 24hr 36hr |48hr 60hr WINTER REFXX - 5 cases 0.209 0.272 0.331 0.400 SUMMER-REFXX - 4 cases 0.114 0.130 0.129 0.144 % DIFFERENCE 83% 109% 156% 178% STUDY PERIOD - SUMMER vs WINTEI * - HT70 RMSD (m] 1 VERwest 24hr 36hr 48hr 60hr 72hr 84hr WINTER REFXX - 305 cases 21.1 25.5 28.8 32.3 35.7 39.8 SUMMER REFXX - 68 cases 16.9 20.8 24.4 27.3 31.0 33.5 % DIFFERENCE +25% +22% +18% +18% +15% +19% Table 9.5. For the Test Period VERwest winter cases, M S L P R M S D of about 0.2 kPa for 24 hr forecasts increases to about 0.4 kPa for 60 hr forecasts. The 60 hr R M S D summer values are less than half of the winter values, showing that weaker synoptic systems in summer are easier to forecast, even with the Pacific data void lack of data. The Study Period R M S D values for the summer cases are 18-25% lower than winter cases. The reasons for the use of the two verification parameters is explained in Section 9.1. Tables 9.4 and 9.5 provide evidence that a RBS operation during summer may not be cost-effective; the negative impact of the Pacific data-void is lower. An one-launch per day, seven-month operation centered on the winter storm season would reduce the required number of launches to about 200, meeting the RBS engineering constraint. 104 9.3.3 WNA Quiet Periods 180W 170W 160W 120W 80W 70W 60W 140W 130W 12QW HOW 100W Figure 9.1. HT70 valid OOZ, 11 February 2003, the middle of the quiet period 2-16 February 2003. The locations of the RBS arrays are shown. The arrow indicates that the HT70 flow over ENA is primarily from the arctic. W N A frequently experiences quiet periods during winter such as during 2-16 February 2003; a large ridge persisted over the eastern Pacific (Figure 9.1). The western ErnPAC_RBS and C e n P A C R B S vROBs may have improved the forecasts over Alaska but the benefit would have been dispersed before reaching VERwest or VEReast. The flow over W N A was from the northwest parallel to the coasts so the Pacific data-void was effectively two-to-three days away. The polar data void would have had more of an effect over VEReast and probably over VERwest as well. Table 9.6 shows the VERwest HT70 RMSD during this quiet time was much better than the winter average (including the quiet 105 periods), and tending towards the maximum benefit that the R S B array could provide (winter vs summer). The Study-Period winters had many quiet periods 3. QUIET TIME, 2-16 FEBRUARY 2003 - HT70 R MSD (m) VERwest 24hr 36hr 48hr 60hr 72hr 84hr WINTER REFXX - 305 cases 21.1 25.5 28.8 32.3 35.7 39.8 SUMMER REFXX - 68 cases 16.9 20.8 24.4 27.3 31.0 33.5 QUIET TIME REXXX - 15 cases 17.4 21.2 25.7 29.4 32.4 34.5 Table 9.6. VERwest HT70 RMSD results during this quiet time were much better than the winter average (including the quiet periods), and tending towards the summer values. 9.4 Regional NWP Forecast Errors 9.4.1 Sources of Forecast Errors It is recognized that every N W P model is imperfect (and always wi l l be). The sources o f N W P forecast errors (Section 5.7) are the effects of the initial conditions, model physics and boundary conditions. To determine the sources of error, a N W P model is often tested against various synoptic situations, taxing the strengths and exposing the weaknesses. Figures 3.3 show two different weather regimes; the weather regime shown in Figure 3.3a suggests much activity over both VERwest and VEReast; the weather regime over Figures 3.3b and 9.1 suggests more activity over VEReast than VERwest. Figure 9.2 shows the Study Period averaged HT70 R M S D growth for VEReast and VERwest individually and collectively. The highest growth rate occurs in the first 12 hrs (Table 9.7). After the first 12 hrs, the error growth rates are approximately linear (Figure 9.2). The error growth over VEReast is higher than that over VERwest. The reasons for the VEReast higher model error growth could be: (1) after 60 hrs all areas of North America are probably affected by upstream data voids eliminating the VEReast initial 3 Winter 04/05 is experiencing a similar number of quiet periods. 106 advantage; VEReast has passed from the beneficiary of good upstream data to the victim of less quality data resulting in a faster deterioration of forecast quality; (2) the effects of poorer modeling of the cross-mountain W N A flow reaches VEReast; Figures 9.7 and 9.9 show averaged 12 hr forecast error fields; (3) the high incidence of blocking ridges over or close to west coast of North America (Figures 3.3b, 9.1) results in many quiet periods over VERwest; (4) the VEReast verification area is too close to the eastern domain boundary and is affected by spurious boundary effects (Section 5.8). MODEL ERROR GROWTH 45.0 40.0 ? 35.0 Q W 5 30.0 a. CL 25.0 o 20.0 15.0 10.0 -•—VERwest REFXX ••—VEReast REFXX VEReast & VERwest REFXX 12hr 24hr 36hr 48hr 60hr 72hr 84hr Forecast Period Figure 9.2. The Study Period averaged M M 5 error growth (after 12 hrs) corresponding to results shown in Table 9.4. After the first 12 hrs, the error growth rates are approximately linear. The VEReast error growth has a larger magnitude than the VERwest error growth. The average is over 373 cases. Smaller RMSD is better. Figure 9.9 shows 12hr HT70 forecast errors; the mountainous areas have large magnitudes. It will be shown (Section 9.8) that the errors can propagate in the prevailing flow. The effect of cross-mountain flow was suggested by determining forecast error for a regime with lower cross-mountain flow. The period of 8-22 February 2004 was a period of 107 lower cross-mountain flow (similar to the synoptic situation of Figure 9.1). The averaged results are shown in Table 9.7. The 8-22 February 2004 VEReast HT70 R M S D error is lower but comparable to the general case suggesting that cross-mountain flow is not the larger forecast-error source. During this period, the arctic data void may have had the stronger negative influence on the forecast results. This analysis leaves the effect of the data voids, the weather regime and the eastern boundary as possible sources of the greater VEReast forecast-error growth rates. 9.4.2 Initialization & Model Attributes Because of its better initialization and superior attributes (Section 6.8), the E T A model should deliver better results than the M M 5 . Comparisons were made between the M M 5 and the E T A forecast results (Table 9.7). 60 hr VERwest and VEReast E T A forecasts were respectively about 25% and 19% better than the M M 5 forecasts. A long with greater model error, the coarser spatial resolution of the M M 5 initial data (Table 6.1) probably produced much of the E T A - t o - M M 5 forecast degradation. This finding suggests a deficiency in the mesoscale modeling effort. The same effect should be true of all mesoscale models using the same and similar data products. Regional-center computing power has increased exponentially in recent years (Section 2.6). Commumty models such as the M M 5 can be driven with increasing model resolution (Section 2.7). Unfortunately, Table 9.7 results suggest that model resolution alone wi l l not improve regional-model N W P i f N W P forecasts are initiated and bounded by lower-resolution data products. 108 Table 9.7 shows that much of the forecast error growth occurs in the first 12 hrs, the period where errors due to initial conditions dominate. The first 12hr M M 5 forecast error growth rate is 12-20% higher than the E T A rate, probably because of poorer initialization. The increasing community-model resolution must be accompanied by greater value-added initial conditions. Benefits produced by new observations reaching N C E P (RBS data) may be lost to mesoscale modeling in the data-product resolution-reduction procedure. Better and faster communications systems are needed to deliver enhanced data-products (ETA104 plus) for the community mesoscale modeling effort (a hoped for future event). ETA vs MM5 - HT70 RMSD (m) 12hr 24hr 36hr 48hr 60hr 72hr 84hr VERwest ETAXX - 143 cases 11.7 13.1 16.0 19.7 23.7 28.3 32.2 VERwest REFXX - 143 cases 14.0 16.5 20.7 25.1 29.7 35.7 39.9 VEReast ETAXX - 143 cases 10.6 13.3 17.2 20.8 24.9 29:0 35.4 VEReast REFXX - 143 cases 12.4 15.8 20.9 25.4 29.8 35.0 42.3 MM5 PENALTY RELATIVE TO ETA VERwest REFXX vs ETAXX +20% +26% +29% +28% +25% +26% +24% VEReast REFXX vs ETAXX +15% +19% +21% +22% +19% +21% +19% ETA vs MM5 - HT70 RMSD ERROR GROWTH RATES (m/hr) FORECAST PERIOD 0-12hr 12-24hr 24-36hr 36-48hr 48-60hr 60-72hr 72-84hr VERwest ETAXX 1.0 0.10 0.23 0.30 0.25 0.23 0.36 VERwest REFXX 1.2 0.17 0.33 0.38 0.33 0.34 0.42 VEReast ETAXX .88 0.19 0.34 0.35 0.34 0.40 0.54 VEReast REFXX 1.0 0.25 0.42 0.43 0.39 0.52 0.64 ERROR GROWTH FOR A L L FLOWS vs SMALLER CROSS MOUNTAIN FLOW - HT70 RMSD (m) 12hr 24hr 36hr 48hr 60hr 72hr 84hr VEReast REFXX (general) - 143 cases 12.4 15.8 20.9 25.4 29.8 35.0 42.3 VEReast REFXX (flow from arctic) - 17 cases 14.3 16.4 19.6 22.6 27.8 30.8 39.3 Table 9.7. A comparison of the forecast skill of ETA and MM5 shows that ETA forecast skill is much higher than the MM5. An estimate of the error growth rates shows that the VEReast error growth is greater than the VERwest error growth. Error growth rates for the averaged VEReast synoptic situation and one period (6-22 January 2004) where there was small cross-mountain flow. There was little difference between the RMSD results. 109 After the first 12 hrs, the M M 5 continues to have a greater error growth rate, probably due to lower model resolution that was used here for M M 5 . 9.4.3 Effect of the Boundary Conditions To determine the potential effects of the boundary conditions, the M M 5 lateral boundary-condition tendencies were updated at different frequencies; every 3 hrs, 6 hrs and 12 hrs. The results are shown in Table 9.8. As the boundary-tendency temporal resolution decreased, the VERwest forecasts did not deteriorate appreciably until 60 hrs into the forecast, probably the time it took for the boundary effects to be propagated to VERwest . The effect over VEReast was felt much sooner. These findings suggest that the domain boundary can have some effect on the results, especially within the longer forecast period and during active weather regimes. L A T E R A L BOUND AY CONDITION RMSD (m) VERwest REFXX - 3 cases 12hr 24hr 36hr 48hr 60hr 72hr 84hr 3hr update 19.5 22.2 26.0 28.7 32.6 34.1 46.1 6hr update 20.0 22.9 25.7 28.5 33.4 34.8 45.8 12hr update 18.8 22.5 26.4 31.1 39.1 38.6 44.4 VEReast REFXX - 3 cases 12hr 24hr 36hr 48hr 60hr 72hr 84hr 3hr update 14.5 18.7 26.7 34.0 38.5 39:7 38.5 6hr update 14.9 19.2 27.5 34.5 40.9 43.8 39.2 12hr update 14.7 21.0 31.1 38.3 45.1 47.9 43.0 PENALTY PAID BY 12HR UPI >ATE vs 3HR UPDAT1 VERwest -4% +1% +2% +8% +20% +13% -4% VEReast 2% +.12% +17% +13% +17% +21% +12% Table 9.8. The lateral boundary tendency updates are varied from every 3 hrs to every 12 hrs. VEReast forecasts are affected more than VERwest forecasts. Since the atmospheric flow is mostly zonal (Figure 3.3a), VEReast forecasts are less l ikely to be affected by the western boundary but more likely to be affected by the eastern boundaries. VEReast is much closer to the M M 5 domain eastern boundary than 110 VERwest. During a typical meridional synoptic situation (Figures 3.3b, 9.1), VEReast forecasts are more likely to be affected by the northern and eastern boundaries. It is possible that some of the larger VEReast error growth (Figure 9.2) could be the result o f distortion caused by the eastern M M 5 boundary. Comparing VEReast E T A X X forecasts to VEReast M M 5 X X forecasts tested this hypothesis. The E T A domain boundary is further east (Figure 6.1), so there should be little or no E T A X X boundary distortion over VEReast. Table 9.7 lists the R M S D growth depicted in Figure 9.3. Both the E T A and M M 5 VEReast error growth is higher than the corresponding VERwest error growth. This result suggests that the eastern M M 5 boundary conditions are not the major cause of the M M 5 VEReast error growth. E T A vs MM5 F O R E C A S T S 45.0 + 50.0 - H —•—VERwest ETAXX —•—VERwest REFXX —A—VEReast ETAXX - * — VEReast REFXX 5.0 0.0 12hr 24hr 36hr 48hr 60hr 72hr 84hr Forecast Period Figure 9.3. A comparison of the ETA vs VERwest forecasts. This sample was taken during the winter of 02/03. Except for the initial periods, the VEReast forecasts were worse than the VERwest forecasts. The VEReast error growth rates were higher. This is a reflection of the respective weather regimes (El Nino winters over VERwest). The averages are over 143 cases. Smaller RMSD is better. Ill The remaining reasons for the differing VERwest vs VEReast error growth rates (Section 9.4.1) are the effects of the data voids and the respective weather regimes. 9.5 Data Assimilation Initialization quality is a function of data availability and D A . We have seen that E D A S - D A produces failures. It is reasonable to expect M M 5 - D A failures as well. The effectiveness of the O S S E procedure and the results of this research are very much dependent on the M M 5 - D A routines, S C M (Section 6.6) and 3 D V A R (Section 6.7). Section 9.2 results have suggested that the maximum averaged HT70 forecast improvement expected from a 12 hr D A is about 20%. In order to reach that level of improvement the M M 5 D A routines need to be optimized. 9.5.1 SCM Sensitivity Studies SCM SENSITIVITY FOR VERwest OSSEs - HT70 RMSD (m) 6 cases OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA 5 grid point ROI 7.38 16.58 22.39 28.87 . 29.83 35.55 47.69 10 grid point ROI 7.38 16.83 22.32 29.21 29.82 35.67 47.39 15 grid point ROI 7.37 17.24 23.79 30.56 30.44 36.47 47.32 SCM SENSITIVITY ] FOR THI E 23 JAN03 PAC RSB THREE BUOY OSSE 23Jan03 case 3012300 3012312 3012400 3012412 3012500 3012512 3012600 REF12 17.7 41.8 25.9 23.3 33.0 32.4 35.0 5 grid point ROI 17.7 42.4 28.3 24.6 34.2 33.1 35.0 10 rid point ROI 17.7 44.3 32.6 28.8 39.7 37.5. 37.7 15 grid point ROI 17.7 47.4 37.4 33.9 45.0 42.8 41.9 20 grid point ROI 17.7 50.4 40.9 37.5 47.3 43.1 42.0 BEN 1.58 35.0 36.8 33.4 37.5 35.4 37.6 REFOO 0.0 29.3 36.2 33.0 35.4 38.7 36.2 Table 9.9. Analysis of SCM sensitivity to the 5, 10, 15 ROI. The average is over 6 cases. The RMSD does not increase until the ROI is more than 10. The 23Jan03 case shows that when bad Pacific vROB data is used, the RMSD can increase rapidly with increasing ROI. 112 The most important M M 5 - S C M parameters are the number of iterations and the grid point ROI of each o f the iterations. The maximum number o f iterations allowed by S C M - M M 5 is 10. Where the maximum ROI was 10 grid points or less, after each iteration the ROI was stepped down one grid point. For sensitivity studies where the maximum ROI was more than 10, the higher numbered ROIs were spaced and the lower numbered ROIs were concentrated4. The maximum ROI was varied from 3 to 20 grid points for a subset o f the six cases studied. A l l cases were treated ROIs of 5, 10 and 15 grid points. The averaged results seemed to be insensitive to the ROI within the 5-15 grid point range (Table 9.9) but do show a shallow R M S D minimum at about the 10 grid-point maximum ROI. Therefore, a ROI of 10 grid points with 10 iterations wi l l be used for the OSSEs. This insensitivity is not always the case. On 23 January 2003 a strong low was approaching the B C coast (Figure 9.4a). ThreeBuoy vROBs with various ROI were assimilated with the 12 hr R E F 12 forecast. This case did not pass the Pacific analysis screening test. Both B E N and REFOO had worse R M S D than R E F 12 for 24-26 January 2003 (Table 9.9). The 72hr A D A HT70 DIF with ROI 10 grid points is shown in Figure 9.4a. The positive and negative areas of DIF are areas of improvement and deterioration respectively. The ROI strongly influenced the HT70 VERwest R M S D (Figure 9.4b) as the impact passed through the VERwest area (12hr A D A ) deteriorating the forecast results. Assimilation of real data not compatible with the first-guess field would l ikely show the same effect, the forecast would deteriorate rather than improve. Management of bad data was a concern throughout this study as it is in an operational environment. This effect is discussed further in Section 9.6.4. 4 The sequences ate shown in Appendix C, Section 9.5.1 SCM-DA Tuning. 113 S C M SENSITIVITY for 23 J A N U A R Y 2003 T H R E E _ B U O Y F O R E C A S T S 60.0 10.0 \ J> • • 0.0 OOhr 12hr 24hr 36hr 48hr 60hr 72hr ADA ADA ADA ADA ADA ADA ADA Forecast Period Figure 9.4. (a) For the 23Jan03 case, the 72 hr ThreeBuoy HT70 RMSD DIF with a SCM-DA ROI of 10. There was hardly any impact left over VERwest. Most of the positive benefit went to Alaska, the negative effect went to the Hudson Bay area. The arrow shows the approaching direction of the cyclone with the head indicating the position at OOhr ADA. (b) Varying the ROI from 5 to 20 grid points resulted in appreciable VERwest HT70 RMSD scatter. For the M M 5 - 3 D V A R scheme, the parameter values to be determined are: the number of recursive passes; the background and observation error-covariance scale-lengths, and the cost-function minimization exit criteria (Section 6.7). The M M 5 - 3 D V A R developer (NCAR) has provided a climatological estimate of the background errors based on the N M C method (Section 6.7). The N M C method uses the monthly mean differences between 24 hr and 12 hr forecasts valid at the same time. 9.5.2 3DVAR Sensitivity Studies 114 The OSSEs assimilate data into 12 hr REF12 forecasts. Figure 9.2 shows that N W P forecast error increases approximately linearly with forecast period (after the first 12 hrs). Since the N C A R climatological error is for a 12-24 hr forecast period, it w i l l be assumed that the background errors (appropriate for the O S S E 12 hr forecast serving as a first-guess field) are approximately half the N M C estimate. As a direct test of the sensitivity, the empirical scale-factor " A " 5 was varied from 1.0 to 3.0. R M S D forecast results showed little difference (not shown). The forecast results are not sensitive to the background error scalings. A n " A " value of 2.0 was adopted as the standard. Barker et al. (2003) have carried out experiments using a model with the same horizontal grid spacing. They set their observation scale length to be 0.25. A t the same time, Barker 6 suggested some experimentation to determine the correct settings. With the background scale length set to 2.0, the observation scale length was varied and the influence plotted for the 15Oct03 case. The results are shown in Figure 9.5. The influence of the single sounding located at 50N 145W increased rapidly with increasing scale length. For an observation scale length of 0.25, the effect of the observation at 50N 150W was felt at the Queen Charlotte Islands location, 30 grid points to the east with a maximum located about 12 grid points to the east. This was probably due to the M M 5 - 3 D V A R emphasis on the wind increment. The wind increment is calculated first; then the mass field is brought into balance with it. The results for larger scale lengths are greater height increments maximizing at increasing distance from the observation. As the scale length increased, the influence increased in order to maintain the same gradient. 5 "A" is the dimensionless scale factor defined in Section 6.7. 6 This Barker suggestion was made by email communication. 115 Obviously, a correct scale length is critical. The value of 0.25 seems excessive. The pattern for 0.18 seems more reasonable. Forecasts for 15Oct03 with various observation variance scalings are shown in Table 9.10. Based on these results, a setting of 0.20 (between 0.18 and 0.25) was adopted. The ratio of observation to background scale length is about 0.10 and is consistent with the value used by Mi l ler and Benjamin (1992) and Deng and Stull (2004). These researchers used a ratio of 0.08 within their D A schemes. M M 5 - 3 D V A R D A of data close to the V E R area can have a negative effect on the OOhr A D A R M S D result (Table 9.10) for the reasons explained above. When this happens, the OSSEs with the NineBuoy vROBs are preferred to the TwelveBuoy vROBs . The v R O B locations are further away from VERwest (Figure 7.2). M M 5 - 3 D V A R uses four control variables: streamfunction, potential velocity, unbalanced pressure and specific humidity. The experiments described above used the same covariance scaling for all the control variables. The results seem to be insensitive to the background error scaling but much more sensitive to the observation scaling. For the 16Oct03 case, the observation-error scalings for the mass and wind control variables are varied. Temperature and humidity are taken together as the mass variables (TTD). Wind is indicated by W N D . The three v R O B modes used are the regular observations including both W N D and T T D and then W N D and T T D alternately denied, Scale lengths are alternately set at 0.20 and 0.25. The B E N v R A O B S continue to include both W N D s and TTDs at the same scale length setting as the vROBs. The results are shown in Table 9.11. 116 (a) obs scale = 0.10 (b) obs scale = 0.18 (c) obs scale = 0,25 (d) obs scale = 0.40 'e) obs scale = 0.50 H 9 / / / / / * . \ \ * m.i v H ') ' r , \ \ \ V ' -V M ' i \ * \ \ V (f) obs scale = 0.60 Figure 9.5. 3DVAR HT70 innovations for the 15Oct03 case from a single RBS sounding located at 50N 150W. The observation location is marked with an X . The units are in m. The background error scaling parameter is set at 2.0. The observation scaling parameter from frames a to f is increased. The influence of the single observation increases dramatically as the scale length increases. NINE BUOY PAC RSB OSSE HT70 1 RMSD (m) 15Oct03 Case 3101500 3101512 3101600 3101612 3101700 3101712 3101800 REF12 19.7 26.2 44.6 46.8 58.1 54.1 26.0 Scale Length = 0.10 19.5 26.2 44.8 47.1 57.7 52.9 25.1 Scale Length = 0.18 19.5 26.4 45.9 45.5 54.5 51.4 26.4 Scale Length = 0.25 20.5 27.6 47.3 44.0 51.6 49.7 27.1 Scale Length = 0.32 36.2 38.6 62.2 55.8 59.3 51.0 26.6 Scale Length = 0.40 36.7 38.7 64.4 . 57.4 61.1 53.0 26.3 REFOO 0.00 27.5 32.8 29.1 38.5 29.7 16.9 Table 9.10. 3DVAR sensitivity to the observation scale length for the NineBuoy 15Oct03 OSSEs The optimum result seems to be a scale length between 0.18 and 0.25. It was set to 0.20. PARAMETER SCALING SENSITIVITY- 16OCT03 3101600 3101612 3101700 3101712 3101800 3101812 3101900 REF12 2.22 3.82 4.49 3.85 3.27 4.27 3.57 WND, scaIe=0.20,No TTD 2.13 3.40 3.89 3.43 3.51 4.04 3.10 TTD, scale=0.20,No WND 2.15 3.44 3.83 3.42 3.49 4.22 3.02 WND, TTD scale =0.20 2.13 3.37 3.82 3.37 3.57 4.06 3.07 BEN, scale= 0.20 1.05 3.39 3.67 3.41 4.19 4.08 3.95 REFOO; scale = 0.20 0.0 3.17 3.42 3.11 4.11 3.33 2.47 WND, scale=0.25,No TTD 2.18 3.69 4.20 3.57 3.40 4.14 3.85 TTD, scale=0.25,No WND 2.13 3.39 3.81 3.43 3.54 4.25 3.16 WND, TTD scale =0.25 2.07 3.32 3.66 3.36 3.61 4.05 3.23 BEN, scale=0.25 1.04 3.42 3.70 3.29 4.41 4.13 3.95 GAIN OVER REF12 - GAIN IS RELA1 riVE TO BEN WND, scale=0.20,No TTD 0.08 0.97 0.74 0.95 ** >1.00 0.52* TTD, scale=0.20,No WND 0.07 0.87 0.81 0.98 ** 0.24 0.61* WND, TTD scale =0.20 0.08 >1.00 0.82 >1.00 ** >1.00 0.56* WND, scale=0.25No TTD 0.04 0.31 0.37 0.51 ** 0.96 -0.33* TTD, scale=0.25,No WND 0.08 >1.00 0.87 0.76 ** 0.10 0.44* WND, TTD scale =0.25 0.13 >1.00 >1.00 0.87 ** >1.00 0.69* *REF00 used as a higher standard. **Both BEN and REFOO scored worse than REF12. Table 9.11. MM5-3DVAR WS70 RMSD results for the 16Oct03 case. Temperature and humidity (TTD) are taken together as the mass variables. Wind is indicated by WND. The three vROB modes used are the regular observations including both WND and TTD and then WND and TTD alternately denied. Scale lengths are alternately set at 0.20 and 0.25. The BEN vRAOBS continue to include both WND and TTD at the same scale length setting as the vROBs. The results showed that the WS70 deteriorated with increasing WND scale length. 118 With the scale length increased, the contribution from the mass observations improved the results slightly while the contribution from the wind observations deteriorated the results. Different scale-length settings are required for the mass and wind variables. To generalize the finding, more experimentation is needed. For M M 5 - 3 D V A R to provide a reasonable result, the requirement for balance in the observation field is obvious. For the 15Oct03-"obs=0.25" example (Figure 9.5), i f an observation near the Queen Charlottes were available to the D A , a more reasonable result west of 50N 145W would probably have resulted. M M 5 - 3 D V A R offers two choices for cost function minimization; a quasi-Newton minimization and the preconditioned conjugate gradient minimization schemes. Both were tried for the 03Jan04 case. There was little difference between the results (not shown). Ideally, the cost function is minimized when the gradient is zero; practically, the cost function was considered minimized when the cost-function gradient is less than 1%. Experimentation with stricter criteria yielded little discemable differences. The purpose of the recursive passes is to approximate a Gaussian distribution for both the background and the observation error. The number o f recursive passes was set to 10. Experimentation with a larger number yielded little discernible differences. 9.5.3 SCM/3DVAR Comparison COMPARISON OF SCM-DA AND 3DVAR-DA - HT70 RMSD(m) NineBuoy OSSE - 3 cases OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REF12 14.8 23.4 31.9 36.0 42.5 44.6 51.5 . SCM-DA 11.7 22.9 29.3 30.6 34.1 35.4 45.0 3DVAR-DA 14.3 23.6 30.5 35.3 40.4 40.8 47.3 Table 9.12. The HT70 RMSD forecast comparison between the results of SCM-DA and 3DVAR-DA for 3 cases. The SCM-DA scored better. 119 For the NineBuoy P A C _ R S B array, the results of M M 5 - S O M and M M 5 - 3 D V A R were averaged for the 26Oct02, 16Nov02, and 16Dec02 cases. The results are shown in Table 9.12. M M 5 - S C M consistently scored better. A n obvious interpretation could be that the observational error of M M 5 - S C M data is zero while the observational error of the M M 5 - 3 D V A R data is appreciable, resulting in the lower impact of the M M 5 - 3 D V A R results. These results cannot be generalized for two important reasons. The S C M - D A assumes zero observational error, an unreasonable assumption. Many of the observation types that t h e . M M 5 - 3 D V A R can assimilate (e.g. radiance data) are not done here. This comparison with M M 5 - S C M is limited to the effect of these three vROBs only. A n assimilation of vROBs simultaneous with all the other observations (Table 3.1) should produce a different result7. However, the comparison does show that S C M - D A can provide reasonable results. Furthermore, the S C M - D A scheme has enough advantages to justify its use. The range of data impact is easy to envision. Quality control criteria are easy to implement. The calibration exercise is straightforward. Computer resources are saved. Because of its simplicity, it can be a good tool for starting an O S S E design. S C M - D A results could be treated as preliminarily leading to the application of the more sophisticated 3 D V A R - D A scheme. 7 These types of experiments should be conducted as a followup to this research. 120 9.5.4 Quality Control Another D A issue that should be considered is the success of the quality-control algorithms. M M 5 - S C M allows the user to turn off the quality control. The data are accepted no matter how hideous the values are. Since the virtual sounding data are taken from a subsequent 12 hr later analysis produced by E D A S , the data should to be physically close to the R E F l 2 12 hr forecast values. However cases do occur where E D A S resolves some feature that was missed 12 hrs previous and substantial differences do occur (another symptom of the Pacific data void). During these situations data are rejected. It was noticed that most of the rejected observations are the wind field components. The wind magnitude may not vary much between first-guess and observation but the change in direction due to a misplaced cyclone may make the wind components quite different. SCM QUAL1 LTY CONTROL - 7 NOV02 - HT70 RIV1 [SD (m) OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA R E F l 2 8.4 16.2 24.4 35.5 38.6 43.0 34.92 Q C - O N 8.4 17.5 23.0 30.4 37.5 41.6 33.0 QC - OFF 8.4 25.8 25.1 28.1 40.8 45.9 35.9 BEN 0.7 13.2 20.6 31.1 39.5 45.1 35.1 Table 9.13. The effect of SCM quality control on HT70 RMSD for the 7Nov02 case. For most periods, the quality control has lowered the RMSD. The RMSD is measured in m. Smaller RMSD is better. Experiments with M M 5 - S C M O N / O F F quality control did not produce much R M S D difference (not shown) with three (out of four) of the cases examined. Table 9.13 shows the results for 7Nov02 SixBuoy v R O B D A with and without quality control. The v R O B s were taken from the analysis valid 12 hrs later and presumably more accurate than the first-guess field. The Q C - O F F acceptance of several of the wind components made the forecast worse. This effect is indicative of a real life problem with D A ; the acceptance of 121 real and accurate data does not always lead to a better forecast; the data has to be compatible with the background field into which it is being assimilated. The M M 5 - 3 D V A R scheme has a different approach to quality control. There are two checks with the surface pressure: (1) i f the difference of the model surface height minus the observed height is greater than 100 m, the pressure observation is discarded; (2) i f the innovation of surface pressure is 5 times the observation error, then the data are discarded. The latter is also true of the upper-air data. Currently, these checks are hard-coded. Most of the rejections for the cases studied here (see Appendix C) were surface pressure (especially in the mountainous terrain) and the wind observations. Wind observations were rejected for the same reason as with the M M 5 - S C M scheme. The magnitude of the wind components can be a highly variable function of cyclone position. For example, for the 16Oct03 B E N 3 D V A R - D A , 27,133 specific humidity observations were accepted; but only 27,025 temperature observations, 26,783 v-wind component observations and 26,498 u-wind component observations. The wind observations suffered an averaged 2.5% rejection rate. This figure becomes even more important when it is seen that the rejections occur in active areas (cyclones). M M 5 -3 D V A R emphasizes the wind increments; they are computed first and the mass increments are brought into balance with them. Without accurate wind observations in active areas, it is unlikely that the mass field wi l l be updated properly. 122 9.6 Validity of the OSSE Procedure 9.6.1 The Study Period Unfortunately for this study, the most active winter over VERwest occurred during the Test Period (La Nina episode). A few exceptions aside, the VERwest winters of the Study Period were relatively quiet but more severe over VEReast (El Nino episodes). The larger model-error growth for VEReast vs VERwest (Figure 9.2) was shown to l ikely be the result of a more active weather regime (Section 9.4). This Study Period aberration is a concern. Generalization of the results may lead to some indeterminate and possibly wrong conclusions. The number of cases captured during the Test Period do not provide for robust statistics. However, experience8 suggests that the Test Period cases may be more representative of the normal VERwest vs VEReast situation. It is recognized that weather patterns are constantly changing (global warming), but the shift after winter 01/02 seems to be too extreme to be permanent. If the shift is permanent it would mean that W N A would have warmer winters and E N A would simultaneously have colder winters. If climatology was available for comparisons, the Test Period results and Study Period results may respectively lie on either side of the average. 9.6.2 Basic Screening Test Section 7.3 introduced a basic acceptance requirement for the core R B S OSSEs; namely, B E N and REFOO VERwest forecasts must verify better than the VERwest R E F l 2 forecasts. If this test is positive then a subset of data assimilated with B E N , which is a subset of data assimilated with REFOO, can be used with some confidence. Ideally, R E F l 2, 8 The author has experience as an Environment Canada Meteorologist for 26 years. 123 the O S S E atmospheres, B E N and REFOO delivers increasingly better results. Unfortunately, this is not always so. Such cases should not be used for OSSEs. The E D A S value results of (Section 9.2) have suggested that the maximum averaged REFOO improvement over R E F l 2 that can be expected is 20%. In order to carry out meteorologically significant OSSEs, an REFOO improvement of 15%+ (after 36-48 hrs) was required, placing the chosen cases in the upper categories of possible E D A S improvement. 9.6.3 NWP Error Linear Growth Figure 9.2 shows that after the first 12 hrs, the forecast error increases approximately linearly through the entire forecast period. Implicit in the O S S E procedure is the linear growth assumption (Section 5.10); the perturbation caused by the added data w i l l travel along the same basic state trajectory as the R E F l 2 forecast with which it is assimilated. Ideally, more data added from the 00Z data set wi l l move the forecast R E F l 2 results to the B E N results and then to the REFOO results. If the B E N results stay within the R E F l 2 to REFOO bounds, then it is assumed that perturbation linear growth prevails for that period. For small perturbations and for short forecast periods, this assumption is probably reasonable. When assimilating more data and integrating past a certain time period, a threshold must exist where non-linear interactions tend toward increasing N W P chaos. In order to validate the O S S E procedure, this threshold should be determined. The averaged B E N results are compared to R E F l 2 and REFOO. The results are shown in Table 9.14. The REFOO R M S D is taken as the standard. The B E N G A I N results 124 are measured relative to REF12 to REFOO R M S D improvement. For the averaged M M 5 -S C M (42) cases, linear growth was suggested for all forecast periods A D A . For the averaged M M 5 - 3 D V A R (29) cases linear growth may have stopped at 60 hrs A D A . If the linear-growth assumption is true for the large amount of data added by B E N , it wi l l be assumed true for the vROBs (a subset of the B E N data). A n examination of the individual cases shows that forecast-error growth patterns are case dependent. MM5-SCM LINEAR GROWTH - HT70 F JVISD (m) VERwest - 42 cases Nhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REF12 24.6 38.0 44.8 53.8 64.3 83.7 90.5 BEN 3.3 32.1 39.5 47.2 56.5 73.2 81.4 REFOO jO.O 31.2 38.3 44.6 51.3 63.8 73.1 GAIN - BEN OVER REF12 RELATIVE TO REFOO 0.86 0.81 0.72 0.60 0.53 0.52 VEReast - 42 cases OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REF12 22.7 31.3 44.8 61.3 66.5 80.5 91.1 BEN 9.3 28.4 40.7 56.2 61.8 75.6 86.2 REFOO 0.0 26.7 37.4 49.2 52.6 67.3 75.2 GAIN BEN OVER REF12 RELATIVE TO REFOO 0.63 0.56 0.42 0.34 0.37 0.31 MM5-3DVAR LINEAR GROWTH- HT7C RMSD (m) VERwest - 29 cases OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REF12 19.5 23.8 27.2 29.5 32.2 36.2 38.1 BEN 13.3 20.0 23.5 25.5 28.0 32.0 37.4 REFOO 0.0 19.8 22.9 24.3 26.5 29.0 31.3 GAIN - BEN OVER R E F l 2 RELATIVE TO REFOO 0.95 0.85 0.76 0.74 0.58 0.10 VEReast -29 cases OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REF12 14.2 17.3 21.6 26.8 32.9 39.0 37.1 BEN 8.6 16.6 19.4 23.9 29.0 34.8 40.9 REFOO 0.0 14.2 17.8 22.4 27.0 32.7 32.9 GAIN BEN OVER REF12 RELATIVE TO REFOO 0.20 0.58 0.65 0.67 0.67 -0.88 Table 9.14. The BEN SCM-DA (42 cases) and 3DVAR-DA (29 cases) are averaged and compared to the same case REF12 and REFOO. For MM5-3DVAR linear growth cannot be assumed after 60 hrs ADA. 125 Interestingly, the S C M - D A B E N G A I N deterioration is gradual; the 3 D V A R - D A G A I N collapses after 60 hr A D A . 9.6.4 The Uneven Pacific Analyses In certain cases, D A of the B E N data wi l l improve the forecast while D A of a certain subset would deteriorate the forecast. Alternately D A of B E N wi l l deteriorate the forecast while a certain subset wi l l improve the forecast. This is due to the uneven quality of the Pacific analysis. This forecast degradation due to poor areas of the Pacific analyses depends on the location of both v R O B insertion and verification. For the 20Feb02 case (Table 9.15), adding data from the Three and SixBuoy P A C R B S deteriorates the 36 hr VERwest forecast. Including data from the Nine and TwelveBuoy arrays mitigates the detrimental effect of the Three and SixBuoy data. Ignoring the E r n P A C R B S TwelveBuoy data and adding only the C e n P A C R B S TwelveBuoy data improves the 60 hr forecast. TEST PERIOD -20Feb02CASl E - MSLP RMSD (kPa) VERwest OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA REFl 2 0.126 0.222 0.296 0.318 0.516 ThreeBuoy 0.126 0.207 0.303 0.359 0.565 SixBuoy 0.127 0.191 0.299 0.374 0.536 NineBuoy 0.127 0.198 0.264 0.323 0.517 Eastern Array 0.115 0.201 0.268 0.356 0.539 Central Array 0.127 0.222 0.285 0.312 0.453 AHBuoy 0.115 0.202 0.265 0.350 0.551 BEN 0.008 0.176 0.234 0.310 0.466 REFOO 0.006 0.170 0.260 0.330 0.482 Table 9.15. VERwest results for the 20Feb.02 case, showing RMSD values after the addition of vROBs for various atmospheres (rows) at various forecast times (columns). Adding "irrelevant data" from the ThreeBuoy and SixBuoy arrays deteriorates the forecast. Adding the NineBuoy vROBs mitigates the effects of the bad data and improves the forecast for some time periods. The best result is by not adding data from the ErnPACRSB array, and adding data from only the CenPACRSB array. This suggests that the degree of forecast degradation caused by the Pacific analysis is uneven, and depends on the juxtaposition of initial data area, verification area, and 126 synoptic situation. The same is true for all the O S S E cases. If v R O B data are inserted at bad locations, then the result may be negative even though other data within the same data set would have contributed positively. This is further justification for a targeted or adaptive R B S observation strategy. 9.7 Impact of the Data Voids 9.7.1 Pacific Data-Void Test Period results (Table 9.3) show that VERwest forecasts suffered a 21% penalty relative to VEReast in the first 36 hours. A 12-20% penalty was seen with the Study Period results (Table 9.4). The penalty decreased with the increasing forecast period and reversed. The analysis of Section 9.4 suggested two main reasons. Firstly, the VEReast forecast skil l decreased toward the already low skil l VERwest because the benefit of the North American data moved out of the VEReast area. Secondly, for the Study Period the VEReast error growth was larger due to the more active weather regime. In order to determine the effects of the data-voids only, the effects of the weather regime must be isolated. OSSEs can quantify the effect of a data void on a data-rich area like VEReast. v R A O B s are added to the R E F l 2 atmosphere starting with the subdomain east of W N A void, and then extending the area westward to include the subdomain east of PACvo id , and finally covering the whole B E N area of Figure 7.2. The E N A weather regime remains the same, only the data-void area changes. Ideally B E N results should improve over P A C v o i d results, which should improve over W N A results. Since the B E N domain is much larger than the PACvo id domain, the results should follow suit. Unfortunately, it was found that this was not the case, probably because of the data-voids. 127 TEST PERIOD - IMPACT OF DATA VOIDS - MSLP RMSD (kPa) VEReast - 13Dec01 OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA REF12 0.085 0.118 0.237 0.297 0.357 WNAvoid 0.062 0.078 0.176 0.240 0.303 PACvoid 0.043 0.069 0.128 0.198 0.267 BEN 0.043 0.070 0.125 0.210 0.295 REFOO 0.0 0.116 0.140 0.184 0.228 GAIN OVER REFl 2 - RELA1 riVE TO REFOO WNAvoid 0.27 >1.00 0.63 0.50 0.42 PACvoid 0.49 >1.00 >1.00 0.88 0.70 BEN 0.49 >1.00 >1.00 >1.00 0.48 % IMPROVEMENT OVER RELATIVE TO A PERFEC1 REFl 2 r FORECAST (RMSD = 0.0) WNAvoid 27% 34% 26% 19% 15% PACvoid 49% 42% 46% 33% 25% BEN 49% 41% 47% 29% 17% REFOO 2% 41% 38% 36% Table 9.16. VEReast RMSD values for the 13Dec01 case for different atmospheres (rows) and different forecast durations (columns). A comparison of WNAvoid with PACvoid results show a deterioration of about 35% in GAIN during the 24 and 36 hr ADA forecast period. Including the WNAall vRAOBs to fill PACvoid results in a 20% improvement in the forecast. Including the BEN area results in a loss. This result suggests that a data-void can cause a 20% loss in forecast accuracy and as much as a 0.35 loss in GAIN. Table 9.16 shows the M S L P R M S D results over VEReast for the Test Period 13Dec01 case. Including the W N A a l l v R A O B S over W N A (and fi l l ing the P A C v o i d area) provides a consistent improvement of about 10%, as early as 12 hr A D A (Table 9.16) into the forecast. Adding the B E N v R A O B s to the PACvo id atmosphere causes deterioration in the longer periods. Figure 9.6 shows the averaged VEReast results for the Test Period winter cases. The 13Dec01 W N A v o i d forecast had greater VEReast R M S D than the P A C v o i d forecast. The G A I N deterioration is as much as 0.35 in the 24 hr A D A forecast (Table 9.16). This finding suggests that the addition of real R B S data west of North America 128 could improve, in selective cases of the most energetically growing storms, the forecast quality by 35%. Tables 9.17 shows the results for the more general Study Period cases. Both the M M 5 - S C M and the M M 5 - 3 D V A R G A I N results shows VEReast improvement when W N A a l l v R A O B S were included with WNAvo id to f i l l PACvo id . There was a lesser improvement when v R A O B S were added to fill B E N . The forecasts improved as more data are added upstream from VEReast but the effect is more subdued than the Test Period results. 0.40 1— 0.00 R E F l 2 WNAvoid PACvoid BEN REFOO Increasing Data Added by Extending the-Domain West Figure 9.6. Variation of forecast errors (ordinate) with increasing area coverage toward the west of RBS buoys (abscissa), for different forecast durations (plotted curves). Error metric is MSLP RMSD over VERwest averaged over all winter cases. Results show that VEReast MSLP forecast errors are reduced as more virtual sounding data are added to the 00Z forecast, but only through PACvoid along the abscissa. Further extension of RBS coverage westward does not improve forecast quality, and even reduces it for the 60 hr forecast. Smaller RMSD is better. 129 STUDY PERIOD- IMPACT OF DAT A-VOIDS - HT70 RMSD (m) MM5-SCM RESULTS VEReast OOhr 12hr 24hr 36hr 48hr 60hr 72hr ADA ADA ADA ADA ADA ADA ADA REF12 11.6 15.9 21.1 30.6 33.0 40.7 44.9 WNAvoid 5.0 14.5 18.9 28.0 31.5 39.7 44.3 PACvoid 5.1 14.6 18.7 27.4 30.3 38.4 43.0 BEN 4.5 14.4 18.7 27.0 29.8 37.4 42.1 REFOO 0.0 13.2 17.0 23.9 25.8 32.9 37.7 GAIN OVER REF 2 - RELATIVE TO REFOO WNAvoid |0.57 0.52 0.54 0.39 0.21 0.12 0.08 PACvoid 0.56 0.50 0.58 0.48 0.37 0.29 0.27 BEN |o.61 0.55 0.58 0.53 0.45 0.42 0.39 % IMPROVEMENT OVER REF12 RELATIVE TO A PERFECT FORECAST (RMSD WNAvoid 57% 9% 10% 8% |5% 2% 1% PACvoid 56% 8% 11% 10% 8% 6% 4% BEN 61% 9% 11% 12% 10% 8% 6% REFOO 100% 17% 19% 22% |22% 19% 16% MM5-3DVAR RESULTS VEReast OOhr 12hr 24hr 36hr 48hr 60hr 72hr ADA ADA ADA ADA ADA ADA ADA REF12 13.4 19.6 23.4 32.6 40.4 50.3 47.6 WNAvoid 8.1 18 20.4 29.0 37.0 46.5 46.6 PACvoid 8.4 17.8 19.1 26.8 34.1 45.3 46.3 BEN 8.7 18.1 18.8 25.8 32.1 43.1 46.9 REFOO 0.0 15.4 16.6 23.8 29.5 38.5 41.3 GAIN OVER REF12 - RELATIVE TO REFOO WNAvoid |0.39 |0.40 0.44 0.41 J0.31 0.32 0.17 PACvoid 0.37 0.45 0.64 0.66 0.58 0.42 0.20 BEN |0.35 |0.35 0.67 0.77 |0.76 0.61 0.11 % IMPROVEMENT OVER REF12 RELATIVE TO A PERFECT FORECAST (RMSD = 0) WNAvoid 40% 8% 13% 11% 8% 8% 2% PACvoid 37% 9% 18% 18% 16% 10% 3% BEN 35% 8% 20% 21% 21% 14% 1% REFOO 21% 29% 27% 27% 23% 13% Table 9.17. The winter 02/03 MM5-SCM and the winter 03/04 MM5-3DVAR VEReast results for the data-void area experiments. Relative to the WNAvoid atmosphere results, the PACvoid atmosphere makes most of the GAIN and % improvement. The BEN atmosphere does not improve the forecasts as much. 130 9.7.2 Polar Data-void When the flow has a large northerly component the area affected by the polar data-void can extend very far south to the highly populated areas of E N A . The Study Period winter 02/03 often had a large-amplitude ridge over or close to the west coast of North America resulting in a northwest flow over VEReast. A n example case is shown in Figure 9.1. To see how Arctic data-denial could affect the M M 5 results, the B E N area (Figure 7.2) providing v R A O B s was rolled south 10 degrees of latitude to create a BEN-smal l area (BENs). The period from 6-20 January 2004 was chosen due to the prevailing flow from the Arctic to over VEReast. The results are shown in Table 9.18. Since the f low was mostly from the Arctic areas to VEReast, little impact was seen over VERwest (as expected). Intuition would suggest that the B E N to B E N s VEReast R M S D results would deteriorate. VEReast results did show some deterioration from B E N to B E N s in the 48-72 hr A D A period, but were erratic for the 12-36 hr periods. INFLUENCE OF A POLAR DATA-VOID - HT70 RMSD (m) VERwest OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REF12 20.2 24.0 25.2 29.5 31.4 34.8 36.6 BENs 14.9 21.2 22.3 26.0 27.4 31.3 35.4 BEN 14.9 20.8 22.2 26.2 27.5 31.4 35.2 REFOO 0.0 20.0 22.3 25.1 26.5 29.5 32.6 VEReast REF12 14.7 15.5 20.4 22.9 28.5 32.3 40.3. BENs 11.1 15.9 21.1 23.7 28.1 29.9 36.7 BEN 8.6 15.7 20.1 22.7 27.1 29.1 35.0 REFOO 0.0 13.2 18.9 21.5 25.7 29.1 35.7 Table 9.18. The Arctic data-denial eliminates the vRAOBS from the top 10° of BEN to become BENs. The results are an average over 16 cases from 06-21 January 2004. Except for the longest forecast period, the results are indeterminate. 131 One interpretation of these results is consistent with a data-void negative impact hypothesis. Because of the lack of new data, the E T A OOhr analyses in the Arctic areas were heavily influenced by the 12hr forecast first-guess fields and remain very close to the first-guess. The small difference between B E N s and B E N results should be expected. 9.8 Pacific RBS Operations 9.8.1 Location Optimum locations for R B S buoys are in areas where the N W P model routinely misforecasts growing meteorological structures, and provides unreliable first-guess fields for the next update. Initial field perturbations in these areas usually produce the most ensemble forecast scatter. To support this hypothesis, meteorological "error structure" fields are produced with mapped 12 hr REF12 R M S D . Table 9.4 shows R M S D of about 15 m in 12 hrs forecast duration (almost half of the total growth) rising to 32 m in 60 hrs. This clearly illustrates that the forecast error in the first 12 hrs has a large impact on the final result, accounting for almost half of the total forecast error growth. For the Test Period, Figure 9.7 shows averaged R M S D over all four high-impact winter cases that have good Pacific analyses for both 12 and 24 hr forecasts. The fields are smoothed because of the high-variability of M S L P R M S D over the mountainous areas (Section 9.4). The error-structure maxima have somewhat north-south orientations. One maximum is just off the Pacific coast, another near 145W, and another near 155W (Figure 9.7a). Figure 9.7b shows that the pattern propagates eastward in the prevailing flow. These Test Period results suggest that 00Z vROBs added at the locations sketched in Figure 9.8 would l ikely have improved the forecasts considerably (see S ection 9.11). 132 Figure 9.7. MSLP forecast-error magnitudes (kPa) averaged over the four high-impact winter storms (13Dec01, 18Feb02, 16Mar02, and 12Apr02). (a) MSLP RMSD of 12 hr forecasts valid at OOZ. The subjectively-drawn heavy lines highlight maximum errors, and indicate that RBS data could have been added there to maximum advantage in forecasting these winter 2001-2002 storms, (b) RMSD of 24 hr forecasts valid at 12Z. Xs are added at the error-maxima locations from the previous 12 h, with arrows toward the new maxima showing how the uncorrected errors propagated for these cases. 133 Figure. 9.8. Potential RBS sites (Xs) optimized to reduce the maximum forecast errors shown in Figures 9.7. Compare these potential RBS networks with the existing, denser radiosonde sounding network over land (Figure 3.2a). Figure 9.9 shows the Study Period results. Figure 9.9 shows the HT70 REF12 forecast errors averaged over 254 winter cases. The HT70 R M S D variation is less than the M S L P R M S D variation, so less smoothing was required. The pattern is less pronounced since the field is averaged over a spectrum of cases, including null cases. However, it is similar to that of Figure 9.7 in many respects. The isopleths are elongated in the north-south direction. The maximum just west of 145W is a prominent feature. Another maximum occurs near 160W. hi those areas, the Study Period analysis confirms the Test Period preferred locations. The Test Period maximum just off of the coasts is not evident in Figure 9.9. It is l ikely that the west-coast maximum emerges only for high-impact synoptic situations, when the E D A S Pacific analysis procedure may have difficulty at the Pacific data-void interface. 134 Figure 9.9. HT70 12 hr forecast errors (m) averaged over 254 Study Period winter cases. No virtual soundings have been incorporated into these runs. The maximum near 50N 145W is the outstanding Pacific feature. Another maximum occurs near 50N 160W, both agreeing with the high-impact analysis of Figure 9.7. The near-coast maximum is missing suggesting that it may only appear with high-impact cases. The analysis shows the high and variable forecast errors caused by the mountain effects. Another feature is the high and variable 12 hr forecast error over and alongside the mountain areas. It is possible that the errors just east of the Rocky Mountains are caused by poor model resolution. M M 5 propagates the error with the prevailing flow (Figure 9.8b) probably contributing to greater forecast error growth over E N A during critical periods. The analysis of Section 9.4.1 suggests that the cross-mountain effects are not a heavy contributor to VEReast forecast error growth when compared to the weather-regime effects. Perhaps this is due to insufficient modeling of the mountain effects. It is l ikely that these error structures shift from year to year with changes in weather patterns (El Nino), model development and analysis procedures. 135 9.8.2 Buoy Array Size 0.44 Figure 9.10. Reduction of MSLP RMSD forecast errors (ordinate) over VERwest with increasing number of sounding buoys over the Pacific (abscisa), averaged over the 2 high-impact storms (13Dec01 and 12Apr02) known to have good analyses. For both the 36 and 48 ADA forecasts, if only a reasonably small number of RBS sites is economically feasible, then most of the benefit was achieved with the SixBuoy array Smaller RMSD is better. The ideal deployment scenario for the R B S is for coverage over the entire Pacific with horizontal spacing similar to the North American radiosonde network. However, incremental improvements from an enlarging R B S array could be associated with diminishing returns (Figure 8.4c). The degree of improvement is case dependent on the number and the relative location of the buoy array, the storm track and the lead-time. The 13Dec01 ThreeBuoy O S S E had quite a noticeable positive effect with only three v R O B s (Section 8.4.1). However, the R B S array was positioned right under a deepening low and was probably located in the best strategic position possible for that case. 136 SCM-DA - PAC GROUP - HT70 RMSD (m) VERwest -16 cases OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REF12 12.4 17.4 22.8 27.8 33.1 41.6 46.0 ThreeBuoy 12.5 17.4 22.5 27.2 32.1 40.1 44.5 SixBuoy 12.5 17.6 21.9 26.0 31.6 39.7 44.2 NineBuoy 12.5 17.3 21.4 25.4 31.2 39.9 44.1 ErnPAC TwelveBuoy 12.3 16.6 20.7 24.6 30.5 39.0 43.9 CenPAC_TwelveBuoy 12.5 17.9 22.1 26.9 32.4 39.9 43.8 AllBuoy (24 vROBs) 12.3 17.6 21.9 25.9 30.4 37.5 40.1 BEN (690 vRAOBs) 1.0 14.4 18.6 21.7 26.2 33.9 39.1 REFOO 0.0 14.3 18.0 20.4 24.5 31.0 37.6 GAIN OVER REF12 - GAIN IS RELATIVE TO BEN ThreeBuoy -0.01 -0.01 0.07 0.10 0.15 0.20 0.23 SixBuoy -0.01 -0.06 0.22 0.29 0.22 0.25 0.27 NineBuoy -0.01 0.05 0.33 0.39 0.27 0.23 0.28 ErnPAC_TwelveBuoy 0.01 0.28 0.50 0.52 0.37 0.34 0.31 CenPAC_TwelveBuoy -0.01 -0.14 0.16 0.14 0.10 0.22 0.32 AllBuoy (24 vROBs) 0.01 -0.07 0.22 0.31 0.38 0.54 0.86 % IMPROVEMENT OVER REF12 ThreeBuoy -1% 0% 1% 2% 3% 4% 3% SixBuoy -1% -1% 4% 6% 5% 5% 4% NineBuoy -1% 1% 6% 9% 6% 4% 4% ErnPAC_TwelveBuoy 1% 5% 9% 12% 8% 6% 5% CenPACTwelveBuoy -1% -3% 3% 3% 2% 4% 5% AllBuoy (24 vROBs) 1% -1% 4% 7% 8% 10% 13% BEN (690 vRAOBs) 92% 17% 18% 22% 21% 19% 15% REFOO 100% 18% 21% 27% 26% 25% 18% Table 9.19a. Winter 02/03 results PAC Group OSSE results with MM5-SCM DA. Study Period PAC Group OSSE results. Most of the GAIN is made by the smaller arrays. Larger arrays deliver more GAIN but with diminishing returns. The TwelveBuoy array was available to improve the forecast results by half of what is possible. Pailleux et al. (1998) reported on OSEs using North-Atlantic A S A P data. On occasions one or two soundings were sufficient to have a large impact. Routinely, 10 soundings were needed to make a "clear positive impact". 137 SCM-DA OSSE RESULTS —•—ThreeBuoy OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA Figure 9.1 la. The SCM-DA HT70 RMSD results for winter 02/03 are averaged over 16 cases. This graph corresponds to the results shown on Table 9.19a. The Atlantic is much smaller than the Pacific, a greater density of conventional buoy and aircraft data already exists over the Atlantic, and the North-American observing network allows models to create relatively good first-guess fields downstream over the Atlantic. If existing northeast Pacific analyses are poorer than Atlantic analyses, then perhaps routine "clear positive impact" might be reached with fewer then 10 additional soundings. 138 GAIN PER BUOY 48HRS ADA 0.06 n Figure 9.11b. The SCM-DA GAIN per buoy after 48 hrs ADA. Most of the GAIN is made with the SixBuoy Array. The hypothesis was tested for the Test Period high-impact cases. The averaged R M S D results for the 13Dec01 and 12Apr02 cases are shown in Figure 9.10. For a reasonably small number of buoys (0-12), most of the R M S D improvements are made with a SixBuoy array. The 48hr A D A R M S D reduced 15-20%, almost equivalent to the improvement goal of 20% (Section 9.2.2). Figure 9.8 shows one possible SixBuoy deployment, in a cross configuration (labeled 6C6k) at the optimum location. Tables 9.19 show the same data for the Study Period. Table 9.19a has the winter 02/03 S C M - D A results; Table 9.19b has the winter 03/04 3 D V A R - D A results. Figures 9.1 l a and 9.12a, respectively, graph the same results. The B E N and REFOO forecasts show that the average improvement is 20-25%, higher than the Section 9.2 estimate of 15-20%. These cases were deliberately chosen (Section 9.6.2) for their forecast success. Both sets show that increasing the buoy-array size improves the results, however, most of the G A I N -139 per-buoy is made by the smaller arrays (Figures 9.1 lb & 9.12b). The G A I N improvement for a 6 to 12 buoy system ranged from 0.30 to 0.60. 3DVAR-DA - PAC GROUP - HT70 RMSD (m) VERwest OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REF12 18.6 24.9 34.2 30.9 32.2 36.8 38.9 ThreeBuoy 18.7 23.7 32.8 28.7 31.2 34.4 41.7 SixBuoy 18.8 23.4 32.4 27.9 30.3 34.2 38.8 NineBuoy 18.6 22.9 32.0 27.8 30.2 33.7 41.1 ErnPAC TwelveBuoy 18.5 22.4 31.2 26.3 29.3 34.0 42.0 CenPACTwelveBuoy 18.7 23.5 33.2 28.6 30.4 34.3 40.8 AllBuoy 18.5 22.0 31.0 26.5 28.6 33.4 41.3 BEN 11.2 18.7 27.9 24.1 24.3 28.3 39.1 REFOO 0.0 20.1 26.7 23.1 24.8 25.3 28.5 GAIN OVER REF12 - GAII> I IS RELATIVE TO B l EN ThreeBuoy -0.01 0.19 0.21 0.33 0.14 0.28 10.52 SixBuoy -0.02 0.25 0.28 0.44 0.24 0.31 10.74 NineBuoy 0.00 0.31 0.35 0.45 0.25 0.36 8.44 ErnPACTwelveBuoy 0.01 0.40 0.47 0.67 0.38 0.32 11.55 CenPAC_TweIveBuoy -0.01 0.22 0.16 0.34 0.23 0.29 7.19 AllBuoy 0.02 0.47 0.50 0.64 0.46 0.40 9.20 % IMPROVEMENT OVER REF12 RELATIVE TO A PERFECT FORECAST (RMSD 0.0) ThreeBuoy -1% 5% 4% 7% 3% 7% -7% SixBuoy -1% 6% 5% 10% 6% 7% 0% NineBuoy 0% 8% 6% 10% 6% 8% -6% ErnPAC_TwelveBuoy 1% 10% 9% 15% 9% 8% -8% CenPAC TwelveBuoy -1% 6% 3% 7% 6% 7% -5% AllBuoy 1% 12% 9% 14% 11% 9% -6% BEN 40% 25% 18% 22% 25% 23% -1% REFOO 19% 22% 25% 23% 31% 27% Table 9.19b. Winter 03/04 PAC Group OSSE results with MM5 3DVAR-DA. Most of the GAIN is made by the smaller arrays. Larger arrays deliver more GAIN but with diminishing returns. The forecast improvement for 6 to 12 RSBs are about half of what is possible. 140 3 D V A R - D A RSB A R R A Y SIZE 45.0 OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA Figure 9.12a. The 3DVAR-DA HT70 RMSD results for winter 03/04 are averaged over 9 cases. This graph corresponds to the results shown in Table 9.19b. GAIN PER BUOY 48HRS ADA 0.05 i Figure 9.12b. The 3DVAR-DA GAIN per buoy after 48 hrs ADA. Most of the GAIN is made with the Three and SixBuoy Array. 141 The findings show that three v R O B S (Section 8.4.1) can have large positive impact; six v R O B s can have clear positive impact; 12 vROBs may be required for consistent results. The Pailleux et al. (1998) Atlantic findings are confirmed or bettered for the analogous North Pacific R B S scenario. Since Pailleux used (real) A S A P soundings, the O S S E findings have some real-world support. 9.9 RBS Data Issues 9.9.1 Profile height For affecting downstream forecast quality, the most sensitive layers of the atmosphere are near the 3 km level (Doerenbecher et al. 2001, Bergot and Doerenbecher 2002). This level must be spanned by a rocketsonde up to some optimum higher level. During the data assimilation process, the influence of mid-tropospheric data can be carried to higher model levels through the D A model vertical-structure functions. Future (real) R B S sounding profiles can be merged with the remainder of the available data (Section 3.1) to yield the best overall D A improvement. For the 13Dec01 high-impact case, ThreeBuoy P A C _ R B S OSSEs with different sounding heights were performed (Figure 9.13a). Figure 9.13b shows the M S L P R M S D 48 hrs A D A . Figure 9.13c shows the G A I N . Smaller improvements accrue with increasing altitude past 4 km, with maximum G A I N of near 0.50 reached at 6 km. These considerations suggest an optimum maximum altitude of 6 km, although 4 km may be adequate. For the winter 02/03 OSSEs using S C M - D A and the winter 03/04 OSSEs using 3 D V A R - D A , Tables 9.20a and 9.20b show the more general Study-Period results. The Study-Period ThreeBuoy results did not show much impact, so the TwelveBuoy 142 atmosphere was used for the winter 03/04 cases (Table 9.20c). GA INs of 0.60 made by the 6 km profiles are shown in Figures 9.14c and 9.14d. Figure 9.13a. Forecast error growth (MSLP RMSD along the ordinate) in the VERwest area as a function of RBS sounding altitude (multiple curves), for various forecast durations (abscissa). These OSSEs are for the 13Dec01 storm, and they all use the ErnPACRSB ThreeBuoy atmosphere of Figure 7.3. Smaller RMSD is better. The BEN curve is shown for comparison, which would fill the whole Pacific region of Figure 7.2 with about 100 vRAOBs with soundings up to 16 km. 13DEC01- RMSD AFTER 48HR ADA 0.50 Figure 9.13b. For the high-impact 13Dec01 storm, results indicate that three buoys launching rocketsondes to only 4 or 6 km would have eliminated near 20% of the forecast error and would have achieved a GAIN near 0.50, compared to the much more extensive and expensive soundings of BEN. 143 13DEC01- GAIN Figure 9.13c. GAIN from increasing the RBS sounding altitude for a ThreeBuoy configuration for the 13Dec01 storm. The 48 hr ADA results of Figure 9.13b, normalized to show the gain (RMSD error reduction) normalized between 0.0 GAIN (highest error, associated with no vRAOBS of REF12) and maximum GAIN of 1.0 (least error, BEN scenario of hundreds of vRAOBs). The most GAIN for the least altitude was achieved with soundings up to 4 to 6 km. SCM-DA PAC THREE BUOY OSSES- HT70 RMSD (m) VERwest OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REF12 -16 cases 12.4 17.4 22.8 27.8 33.1 41.6 46.0 Profile to 2 km 12.5 17.2 22.6 27.6 32.7 41.2 45.6 Profile to 4 km 12.5 17.4 22.7 27.5 32.5 40.7 45.0 Profile to 6 km 12.5 17.5 22.8 27.5 32.4 40.6 44.9 Profile to 8 km 12.5 17.4 22.5 27.2 32.1 40.1 44.5 BEN (to 16 km) 1.0 14.4 18.6 21.7 26.2 33.9 39.1 REFOO 0.0 14.3 18.0 20.4 24.5 31.0 37.6 GAIN OVER REF12 - GAIN IS RELATIVE TO BEN Profile to 2 km 0.06 0.06 0.04 0.05 0.06 0.06 Profile to 4 km 0.00 0.03 0.06 0.08 0.12 0.16 Profile to 6 km -0.04 0.01 0.05 0.09 0.14 0.16 Profile to 8 km -0.01 0.07 0.10 0.15 0.20 0.23 Table 9.20a. Winter 02/03 MM5-SCM ThreeBuoy OSSE results for 19 cases. Compared to the 13Dec01 case (Figures 9.13), there is much smaller GAIN. 144 THREE BUOY SCM-DA - VARIABLE HEIGHT PROFILES 00 hr 12 hr 24 hr 36hr A D A 48hr A D A 60hr A D A 72hr A D A A D A A D A A D A Figure 9.14a. The ThreeBuoy results of Table 9.20a. For the 19 cases analyzed, there is little GAIN. 3DVAR-DA PAC T H R E E B U O Y OSSES- HT70 RMSD (m) VERwest - 3 cases OOhr A D A 12hr A D A 24hr A D A 36hr A D A 48hr A D A 60hr A D A 72hr A D A REF12 17.6 25.5 41.5 35.3 38.1 42.8 44.1 Profile to 2 km 18.1 25.4 41.8 33.2 37.8 41.3 40.9 Profile to 4 km 18.1 25.5 41.9 33.8 38.1 41.1 40.8 Profile to 6 km 18.2 25.5 41.9 33.6 37.8 40.9 40.7 Profile to 8 km 18.2 25.4 41.6 33.3 37.4 40.5 40.4 B E N (to 16 km) 9.6 20.6 30.5 24.3 25.7 27.8 33.7 REFOO 0.0 23.0 29.0 26.5 29.6 30.6 27.0 GAIN OVER REFl 2 - GAIN IS RELATIVE TO BEN Profile to 2 km -0.05 0.03 -0.03 0.19 0.02 0.10 0.31 Profile to 4 km -0.06 0.01 -0.04 0.14 0.00 0.11 0.32 Profile to 6 km -0.07 0.01 -0.04 0.15 0.03 0.13 0.33 Profile to 8 km -0.07 0.02 -0.02 0.18 0.06 0.15 0.36 Table 9.20b. Winter 03/04 MM5-3DVAR ThreeBuoy OSSE results. There was little GAIN for the 8 cases analyzed. 145 THREE BUOY 3DVAR-DA - VARIABLE HEIGHTS 50 -i 45 OOhr A D A 12 hr A D A 24 hr A D A 36hr A D A 48hr A D A 60hr A D A 72hr A D A Figure 9.14b. The ThreeBuoy results of Table 9.20b are shown. Three was little GAIN for the 3 cases analyzed. 3 D V A R - D A T W E L V E B U O Y O S S E S - H T 7 0 R M S D R E S U L T S VERwest - 3 cases OOhr A D A 12hr A D A 24hr A D A 36hr A D A 48hr A D A 60hr A D A 72hr A D A REF12 20.2 27.2 32.3 33.8 34.3 39.7 34.1 Profiles to 2 km 20.6 25.0 28.5 29.5 30.8 37.2 37.7 Profiles to 4 km 20.4 24.6 27.9 28.9 30.3 37.3 38.4 Profiles to 6 km 20.2 24.6 27.6 27.8 29.9 37.9 39.1 Profiles to 8 km 20.1 24.2 27.1 26.8 29.4 37.9 39.1 B E N (to 16 km) 11.3 20.6 23.5 23.6 23.3 31.3 38.9 REFOO 0.0 20.9 23.9 22.7 25.5 24.6 26.0 G A I N O V E R R E F ] 12 - GA] IN IS R E L A T ] [VE T O B E N Profiles to 2 km -0.04 0.34 044 0.42 0.32 0.30 0.75 Profiles to 4 km -0.02 0.39 0.51 0.48 0.37 0.29 0.90 Profiles to 6 km -0.01 0.39 0.54 0.59 0.40 0.22 >1.00 Profiles to 8 km 0.01 0.45 0.59 0.69 |0.45 0.21 >1.00 Table 9.20c. Variable height TwelveBuoy vROB OSSEs assimilated with MM5-3DVAR. Most of the benefit was delivered with 6 km profiles. 146 T W E L V E B U O Y 3DVAR-DA - V A R I A B L E HEIGHTS Figure 9.14c. The TwelveBuoy results of Table 9.20c are shown. The results became erratic at the end of the forecast. GAIN - 48HRS ADA Figure 9.14d. The TwelveBuoy 3DVAR-DA GAIN 48 hrs ADA. A GAIN of near 0.60 was achieved by 6 km profiles (3 cases). A GAIN of 0.70 was made by 8 km profiles. 147 9.9.2 Meteorological Variables R B S developers have the option of deploying pressure-temperature-humidity (PTH) rocketsondes or GPS rocketsondes that provide P T H and wind data. The GPS rocketsondes are about twice as expensive as the P T H rocketsondes. SCM-DA SIXBUOY PAC RSE I OSSE - WS7 ORMSI ) (m/sec VERwest- 2 cases OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REF12 2.50 3.41 3.87 3.67 3.47 4.92 6.96 TTD, no WND 2.50 3.35 3.70 3.56 3.43 4.95 7.04 WND, no TTD 2.48 3.22 3.64 3.52 3.40 4.56 6.58 WND and TTD 2.49 3.10 3.40 3.39 3.38 5.01 7.27 BEN 0.30 1.70 2.36 2.80 3.24 4.64 6.48 REFOO 0.00 1.69 2.36 2.77 3.02 3.87 5.75 GAIN OVER REF12 - GAIN IS RELATIVE TO BEI TTD, no WND 0.00 0.04 0.12 0.12 0.16 -0.10 -0.15 WND, no TTD 0.01 0.11 0.15 0.17 0.29 1.00+ 0.79 WND and TTD 0.00 0.18 0.31 0.33 0.40 -0.33 -0.63 VEReast OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REF12 1.88 2.17 3.03 3.61 3.71 3.84 4.16 TTD, no WND 1.88 2.17 3.03 3.60 3.62 3.68 3.94 WND, no TTD 1.87 2.17 3.07 3.60 3.62 3.61 3.85 WND and TTD 1.87 2.17 3.05 3.58 3.54 3.57 3.71 BEN 0.94 1.66 2.44 3.59 3.48 3.74 3.45 REFOO 0.00 1.49 2.59 3.26 3.36 3.78 3.56 GAIN OVER REF12 - GAIN IS RELATIVE TO BEI V TTD, no WND 0.00 0.00 0.00 0.74 0.41 1.00+ 0.31 WND, no TTD 0.01 -0.01 -0.06 0.76 0.39 1.00+ 0.44 WND and TTD 0.00 0.00 -0.03 1.00+ 0.76 1.00+ 0.64 Table 9.21a. The impact of the wind (WND) and mass (TTD for temperature and humidity) observations on the OSSE results. The observations for the 20Nov02 and 24Mar03 cases were assimilated with MM5-SCM. 148 3DVAR-DA SIXBUOY PAC RSB OSSE - WS70 RMSD (m/sec) VERwest OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REF12 2.37 3.74 3.79 5.04 5.45 5.41 5.30 TTD, no WND 2.32 3.56 3.63 4.73 5.09 5.17 5.28 WND, no TTD 2.34 3.48 3.51 4.81 5.04 5.11 5.28 WND and TTD 2.35 3.43 3.43 4.75 4.94 5.04 5.19 BEN 1.27 2.76 3.31 3.92 4.36 4.47 4.87 REFOO 0.00 2.52 3.15 3.96 4.11 3.98 4.21 GAIN OVER REF12 - GAIN IS RELATIVE TO BEN TTD, no WND 0.05 0.18 0.33 0.27 0.33 0.26 0.04 WND, no TTD 0.03 0.26 0.59 0.21 0.37 0.32 0.04 WND and TTD 0.02 0.32 0.75 0.26 0.47 0.39 0.25 VEReast OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REF12 2.08 3.06 3.65 4.23 4.79 6.12 6.66 TTD, no WND 2.06 2.99 3.56 4.10 4.50 5.69 6.31 WND, no TTD 2.06 3.00 3.57 4.13 4.46 5.69 6.38 WND and TTD 2.06 2.99 3.56 4.10 4.44 5.63 6.28 BEN 2.06 2.76 3.18 3.81 4.13 5.48 5.87 REFOO 0.00 2.30 2.85 3.32 3.73 4.81 5.55 GAIN OVER REF12 - GAIN IS RELATIVE TO BEN TTD, no WND 1.32 0.24 0.18 0.31 3.44 0.67 0.44 WND, no TTD 1.30 0.22 0.16 0.24 ( 3.51 0.67 0.36 WND and TTD 1.37 0.24 0.18 0.31 ( 3.53 0.76 0.48 Table 9.21b. The impact of the wind (WND) and mass (TTD for temperature and humidity) observations on the OSSE results. The observations for the six cases were assimilated with MM5-3DVAR. To determine the relative impact of the observations that each could provide, OSSEs were carried out for the SixBuoy P A C _ R B S array, but alternatively denied the mass and wind parameters. Tables 9.21a and 9.21b show the WS70 R M S D results for M M 5 - S C M and M M 5 - 3 D V A R respectively. Pressure is always observed in order to determine the height. The three modes used were the regular OSSES including both W N D and T T D and then W N D and TTD alternately denied. For S C M - D A , there is substantial G A I N when W N D observations are added to TTD. The 3 D V A R - D A results are not as 149 pronounced as the S C M - D A results. M M 5 - 3 D V A R places a heavy reliance on the wind increments (Barker et al. 2003); greater W N D than T T D impact was expected. VERwest results show improvement when W N D is added to TTD, but VEReast results suggest that W N D and T T D work equally well. These findings suggest that R B S observations should include pressure, temperature, humidity, and wind. 9.9.3 Observation-error Sensitivity The N C E P errors are shown in Table 6.2. In 3 D V A R - D A , the relative weight given to the observations vs. the first-guess is a function of their error variances (Section 6.7). For the NineBuoy P A C R B S 15Oct03 case, the observation-error variances were halved and then doubled. The results are shown in Table 9.22. The results do not appear to be sensitive to the error levels. More experimentation is required before this finding can be generalized. ERROR SENSITIVITY - NINEBUOY PAC RSB HT70 RMSD (m) - 15OCT03 VERwest 3101500 3101512 3101600 3101612 3101700 3101712 3101800 REF12 19.7 26.2 44.7 46.9 58.2 54.1 26.1 Half Error 20.6 27.1 47.3 46.1 54.4 51.1 25.6 Full Error 20.5 27.6 47.5 44.5 52.3 50.1 27.0 Double Error 20.4 27.4 47.4 45.8 54.6 51.3 26.2 REFOO 0.0 27.5 32.8 29.1 38.5 29.7 16.9 Table 9.22. The 15Oct03 VERwest HT70 RMSD results for various observation-error magnitudes. For this case the results do not seem sensitive to the observation errors. 9.10 RBS Targeting The positive impact of new data can be substantial when the new data projects onto a growing meteorological structure not captured properly by the first-guess field. This occurred with the 13Dec01 ThreeBuoy P A C R B S OSSE (Figures 8.2). The impact can be far less i f the data projects itself onto a weakening meteorological structure, even i f the 150 first-guess field has not captured the feature properly. A n example is the C e n P A C _ R B S vROBs included in the 13Dec01 case. If the R E F 12 12 hr forecast is already good, then the O S S E impact of inserted soundings wi l l l ikely be low, even for a high-impact event. If a predetermination can be made where the growing meteorological structures would be located (such as is being currently done using adjoint and other methods at some of the larger numerical-forecast centers) , then Targeted R B S launches could be limited to those R B S s already in that area. The same would be true of the current observing system and of the other new T H O R P E X observing systems (dropsondes, aerosondes, smart balloons, etc). Furthermore, i f a predetermination can be made on the quality of the first-guess field, then the number of R B S launches can be modified accordingly. If the first-guess field is poor then all the RBSs should launch; i f the first-guess field is good, R B S rocketsonde resources can be saved. The 15Oct03 and 16Oct03 (heavy rain cases of Section 8.4.3) are treated with the Targeted R B S simulations. For the 15Oct03 synoptic situation, the history of the M S L P surface-trough positions is shown in Figure 9.15a along with the M M 5 12 hr forecast. The forecast aids discussed in Section 2.4 could be used to support or alter the anticipated position of the low-pressure trough. The G O E S image sequence to 14 October 23:30Z (just before 15 October OOZ) shows that the frontal system digging the trough further south and not dispersing as it encounters the ridge further upstream (as forecast). The system cloud area is sketched in Figure 9.15 a. 151 180W 170W 160W 120W SOW 70W 60W 140W 130W I20W HOW lOOW Figure 9.15a. DDHH is the day and hour group in October 2003. The low-pressure centre positions are indicated by the ellipses. The low pressure trough positions are indicated by the dashed lines. The 1500 features are the REF12 MM5 12hr forecast positions. The grey line indicates the system cloud taken from a NOAA GOES image. The RBS E2, E7, E l 1, E14 positions are indicated. Profiles are taken at 1500 to 6 km. Figure 9.15b. The impact of the Targeted RBS-E2-E7-E11-E14 profiles is shown. The heavy solid lines are the Targeted RBS reanalysis valid at OOZ on 15 October 2003. The thinner dashed lines are the HT70 (m) difference between the Targeted RBS analysis and the REF 12 12 hr forecast. RBS targeting lowered the heights ahead of the trough which resulted in a more intense system and a better forecast. 152 r r \ ,0 Figure 9.15c. The 1600 features are the REF12 MM5 12hr forecast positions of the surface low and low pressure trough valid 16 Oct at 00Z. The grey line indicates the system cloud taken from a NOAA GOES image valid half hr before. The RBS E5, E6, E7, E10, E13 positions are indicated. Profiles to 6 km are provided at 00Z 16 October 2003. The full compliment of R B S launches is required for the next 12Z synoptic hour (Section 9.3.1), so R B S launches must be limited. Suppose that only RBS-E2-E7-E11-E14 are launched at 00Z on 15 October. These are the RBSs aligned in advance of the frontal complex (Figure 9.15a). The resulting 00Z Targeted-RBS reanalysis lowers the heights over the eastern Pacific (Figure 15b). The analyzed frontal stream is more intense. 24 hrs later, the next weather system in the series is approaching M O W with the low heading northeast (Figure 9.15c). Suppose that RBS-E5-E6-E7-E10-E13 provided profiles at 00Z 16 October 2003 for the 16Oct03 case. After 3 D V A R - D A , the M M 5 forecast results are shown in Table 9.23. The P A C R B S TwelveBuoy results from Section 8.4.3 are included. A critical time during the storm event was from 03101612 to 03101800 153 (Section 8.4.3), a time when the Targeted vROBs would have helped the N W P forecast tremendously. RBS TARGETING - HT70 RMSD (m) VERwest RMSD 3101600 3101612 3101700 3101712 3101800 3101812 3101900 REF12 17.9 19.0 25.0 25.7 19.3 27.9 23.4 RBS-E5-E6-E7-E10-E13 20.3 17.4 24.4 18.9 14.7 24.3 22.2 Twelve Buoy 19.4 15.1 21.6 17.0 14.1 21.9 20.3 BEN 7.3 15.0 26.3 18.8 12.8 23.2 31.3 REFOO 0.0 9.7 15.5 16.5 19.2 19.1 19.3 RBS TARGETING - GAIN OVER REF12 - RELATIVE TO BEN RBS-E5-E6-E7-E10-E13 -0.23 0.40 0.06* 0.99 10.71 0.77 -0.15 TwelveBuoy -0.14 0.98 0.34* >1.00 |o.80 >1.00 -0.39 *REF00 substituted as the benchmark results. % IMPROVEMENT OVER REF12 RBS-E5-E6-E7-E10-E13 -13% 8% 2% 26% 24% 13% 5% Twelve Buoy -8% 21% 14% 34% 27%. 22% 13% BEN 59% 21% -5% 27% 34% 17% -34% REFOO 49% 38% 36% 1% 32% 18% Table 9.23. The RMSD improvements that a RBS targeting strategy can provide are less but comparable to the ErnPACRBS TwelveBuoy benefits. The Targeted R B S results are comparable to the B E N results and provide the major portion of the P A C R S B TwelveBuoy improvement. The R M S D improvement in the critical 48-60 hr A D A period is larger than the 20% objective (Section 9.2.2). R B S targeting has advantages over dropsonde and aerosonde targeting. The R B S targeting procedure does not need a complex decision making process ( E T K F or adjoint modeling). It can be done manually with short-term forecast aids (Section 2.4) and satellite image interpretation. The lead-time for R B S targeting decisions can be very short, an hour or so. The lead-time for dropsonde and aerosonde targeting is 24 hrs or more. The R B S data is synoptic; the dropsonde and aerosonde data are valid over several hours. 154 9.11 RBS Deployment Figure 9.16a. Comparison of RMSD error growth for the two RBS configurations of Figure 9.8 using maximum sounding altitude of 6 km for both. Results are averaged over the storms of 13Dec01, 18Feb02, 16Mar02, and 12Apr02. Smaller RMSD is better. Both RBS arrays have reduced the errors close to the BEN values for the 36 and 48 hr ADA forecast periods. The long-range goal is to have an R B S array (approximately) as shown in Figure 7.3 or 7.6. It is acknowledged that a fully deployed system may be cost prohibitive. It is economically desirable to carry out a phased deployment covering the climatological active areas. This was studied during the Test Period. The error structure analysis of Figure 9.8 suggests that an initial R B S deployment could be the ThreeBuoyTarget located 200 km off the B C and Washington coasts equivalent to RSB-E4-E8-E12 (Figure 7.6), labeled as 3C6k in Figure 9.8. This configuration would protect the coastal radiosondes from D A rejection (Section 3.4) during near-coastal cyclogenesis events (Section 8.4.1). A SixBuoyCross centered at 50N 145W equivalent to RBS-E3-E5-E6-E7-E9-C9 (labeled as 6B6k in Figure 155 9.8) should eliminate some of the "climatological R M S D error" shown in Figures 9.7 and 9.9. Later the TwelveBuoy E r n P A C _ R B S (Figure 7.3) could be deployed finishing with the full-blown array of Figure 7.7, where 6 km apogee rocketsondes would be used. (b) LIMITED RBS DEPLOYMENT 48 HR ADA RMSD 0.44 0.42 o 0.38 8 0.36 | 0.34 0.32 0.30 REFl 2 3C6k 6B6k BEN REFOO Figure 9.16b The average MSLP RMSD 48 hrs ADA for the storms of 13Dec01, 17Feb02, 15Mar02, and llApr02. The profiles are to 6 km. Smaller RMSD is better. The absolute improvement is about 15%. Both RBS arrays have reduced the errors close to the BEN values. The GAIN is close to 0.70. The extra R B S sites shown in Figure 7.7 were added when it became obvious that more coverage to the southwest was needed for the heavy rain situations similar to the one starting 15 October 2003 (Section 8.4.3) and again on 16 January 2005. For the Test Period cases, both configurations provided a G A I N of about 0.70 (Figures 9.16). The absolute R M S D improvement was 15%, a major portion of the 20% objective. The more general Study Period cases show GAINs of up to 0.35 at 36-48 hrs A D A (Table 9.24). 156 SCM-DA PHASED DEPLOYMENT - HT70 RMS] D(m) VERwest 3 cases OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA R E F l 2 12.0 15.6 21.2 26.2 30.7 39.8 45.1 3Tb6k 12.0 15.5 20.3 24.7 30.0 39.2 44.7 6Cb6k 12.0 15.7 19.9 23.4 29.3 38.5 42.6 TwelveBuoy 11.7 15.4 19.9 23.7 27.9 38.9 45.6 BEN 0.8 10.7 14.9 15.8 17.7 24.1 36.7 REFOO 0.0 11.6 16.3 15.0 16.9 22.1 37.8 GAIN OVER REFl 2 - GAIN IS RELATIVE TO BEN 3Tb6k 0.00 |0.02 0.14 0.14 9.06 0.04 0.05 6Cb6k 0.00 -0.03 0.21 0.27 D.ll 0.08 0.30 TwelveBuoy 0.02 |o.04 0.21 0.24 }.22 0.05 -0.05 5DVAR-DA PHASED DEPLOYMEIN T - H l [70 RMSD (mj VERwest - 5 cases OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REF12 17.5 21.7 32.1 29.5 34.0 38.6 36.5 3Tb6k 18.3 21.1 31.4 26.4 31.2 35.3 33.9 6Cb6k 18.3 21.2 31.6 26.6 31.1 34.5 33.3 TwelveBuoy 17.9 20.2 29.5 23.0 27.7 33.2 34.4 BEN 8.5 18.0 26.8 21.1 22.1 25.5 31.8 REFOO 0.0 18.4 24.5 22.5 25.1 25.2 24.5 GAIN OVER REFl 2 - GAIN IS RELATIVE TO BEN 3Tb6k 10.15 0.13 0.37 0.24 0.25 0.56 6Cb6k 0.15 0.10 0.34 0.24 0.31 0.68 TwelveBuoy |o.41 0.50 0.78 0.53 0.42 0.44 Table 9.24. Study Period results for the phased deployment arrays. The winter 02/03 SCM-DA (3 cases) showed GAINs of near 0.25 for the 6Cb6k buoy array. The winter 03/04 3DVAR-DA (5 cases) showed GAINs of near 0.35 for the 36 hr period. 9.12 Arctic Operations A n analysis (similar to Section 9.8.1) of the M M 5 HT70 12hr forecast errors over the Arctic shows that the 12 hr forecast error is comparable to the error over the Pacific data-void (Figure 9.17) in certain areas. 157 3DVAR-DA ARCTIC OPERATIONS- HT70 RMSD (m) VERwest - 7 cases OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REF12 12.7 20.4 22.8 28.3 33.3 40.6 44.5 ARC 6b6k array 12.2 19.7 21.6 27.6 33.1 40.9 46.5 BEN 0.8 17.0 18.6 24.6 29.1 37.7 42.1 REFOO 0.0 16.7 17.5 21.8 25.0 30.9 37.4 VEReast - 7 cases REF12 12.3 15.6 21.6 29.5 28.6 34.5 44.8 ARC 6b6k array 12.3 15.3 20.7 28.4 28.7 35.3 46.5 BEN 4.1 13.6 16.8 24.9 26.3 32.2 38.9 REFOO 0.0 12.3 16.6 22.6 24.8 34.5 36.0 GAIN OVER REF12 - GAI N IS RELAT] [VETt 3 BEN VERwest GAIN ARC 6b6k array 020 0.29 0.20 0.05 -0.12 -0.83 VEReast GAIN ARC 6b6k array 0.20 0.18 0.26 0.0 -0.34 -0.28 Table 9.25. The arctic deployed RBS units provided an averaged (7 cases) VEReast GAIN of about 0.20 up to and including 36 hr ADA. The data became erratic after 48 hrs ADA. Figure 9.17. The Study Period 12hr HT70 RMSD errors (m) over the arctic areas from an average of 220 cases. Over Alaska, Yukon and the Northwest Territories, the 12hr forecast errors are greater than the Pacific data-void forecast errors (Figure 9.9). In a similar manner to the Pacific R B S OSSEs, virtual soundings were added to the 12 hr REF12 forecasts according to the configuration shown in Figure 7.4. Seven cases 158 where the flow was from the north to northwest and the flow was inward into the domain were chosen. The results are shown in Table 9.25. The G A I N during the 12-36 hr A D A period reached above 0.20, suggesting that there is utility to adding R B S sites across northern Canada. Unfortunately, the data became erratic towards the end of the forecast period. 9.13 Observation System Configuration 9.13.1 Extending the Network EFFECT OF ADDING VIRTUAL RADIOSONDE DATA Figure 9.18. Impact of adding/denying virtual radiosonde soundings over land to a scenario where RBS soundings would already exist over the Pacific. Forecast errors (RMSD, ordinate) of MSLP (kPa) over VERwest are plotted for various numbers of sounding buoys (abscissa), with all results for 48 hr ADA for the 13Dec01 storm. The P A C R B S curve has only sparse RBS soundings over the Pacific (with no soundings over land), while the PACvoidRBS curve is for the sparse RBS soundings over the Pacific plus vRAOBS spaced 225 km apart over land. For the VERwest region, vROBS were more important than continental vRAOBs for this storm. For rapidly changing situations, it is the data that is upstream that has the most impact on the forecast over an area; all the VERwest v R A O B s do not help the forecast over 159 VERwest that much. For the Test Period 13Dec01 P A C case, the P A C v o i d _ R B S has less R M S D error than P A C R B S (Figure. 9.18). This is expected because more abundant and reliable data is added to the simulation. However, a surprising outcome is that i f we start with only the SixBuoy R B S observations with no v R A O B s over North America, then the addition of all the other virtual radiosonde data over North America improves the R M S D over VERwest by only 10% or less. Also, the two curves in Figure 9.18 converge as the P A C R B S buoy array size increases. This suggests that the observational network over western North America could be thinned in favor of observations offshore. Future work should investigate this preliminary result more thoroughly. 9.13.2 Balanced Operational Network The existing radiosonde network is unbalanced, with data-rich continents and data-void oceans and polar areas (Section 3.2). One way of balancing the network is to increase the sounding density over the oceans to match that over land. While this ideal scenario is economically unlikely, the merits of a balanced network encourage investigation of an alternative; namely, spreading the existing number of soundings to cover continents and oceans with equal density. Economic arguments against the R B S system could be its high cost and the requirement for new resources. Competition for new resources in a tight fiscal environment can obstruct many good initiatives. The current in-situ observation system is sparse over the Pacific, but possibly redundant, over the lower-mainland United States (Section 3.2.). If the net effect is beneficial, deployment of the R B S at the expense of some of the current North American observational network may provide a funding option. 160 O S S E results show that such reapportionment of existing resources produces better forecast verification than can be obtained with the existing unbalanced network. Namely, i f the total number of sounding sites cannot be increased to include ocean coverage, then a second-best alternative is to sacrifice some continental radiosonde site density, and utilize the saved funds to deploy R B S soundings over the Pacific. A series o f data-denial OSSEs studied the effect of reducing the spatial density of virtual radiosonde soundings. B E N prefixes all these O S S E atmosphere names. BEN05 is the regular benchmark atmosphere with v R A O B S inserts every 5-grid points of the M M 5 domain (Figure 7.2). B E N 0 6 , B E N 0 7 , BEN08 , B E N 0 9 , B E N 10 have 6, 7, 8, 9,10 grid-point spacing respectively. THE16MAR02 CASE-BEN ATMOSPHERE RMSD CHANGES vs VIRTUAL RADIOSONDE SPACING 0.00 -I 1 1 1 , 1 OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA Forecast Period Figure 9.19a. This figure shows the 16Mar02 VEReast MSLP RMSD (kPa) associated with reduced amounts of vRAOB thinning. BEN 10 is the coarsest spacing, and BEN05 is the finest spacing, corresponding to the actual average radiosonde spacing over North America. Smaller RMSD is better. 161 When BEN05 was used for the 16Mar02 case, both the VERwest and VEReast 48 hr A D A forecasts showed good improvement over t he ,REF l2 forecasts. The Pacific analysis was judged to be reasonable. Over VEReast, Figure 9.19a shows R M S D values of 0.3 kPa for REF12 , and 0.2 kPa for BEN10 . The forecast improvement from 0.3 kPa to 0.2 kPa represents the maximum normalized G A I N (Figure 9.19b). The normal v R A O B spacing in the BEN05 atmosphere is every 5 grid points (equal to a nominal distance of 225 km). When the v R A O B spacing was changed to 7 grid points (315 km in BEN07) , the G A I N degraded by 0.05 to 0.10. When the spacing was changed to 10 grid points (450 km in BEN10) , the G A I N degraded by 0.30. BEN07 has a virtual radiosonde density of about 1/2 of B E N 0 5 . Forecast degradation is a non-linear function of decreasing v R A O B density, decreasing slowly at first GAIN 1.00 0.80 0.60 0.40 0.20 0.00 REF12 BEN10 BEN07 BEN05 Figure 9.19b. The 16Mar02 GAIN after 48 hrs ADA for the OSSEs that are shown in Figure 9.19a. BEN 10 is the coarsest spacing, and BEN05 is the finest spacing, corresponding to the actual average radiosonde spacing over North America. Larger GAIN is better. The measured VERwest data-void penalty was shown to be more than 20%. If the O S S E B E N results for VERwest are indicative of those for VEReast, then the radiosonde 162 density can be decreased by 1/2 with only a 0.05-0.10 G A I N penalty over VEReast. SixBuoy R B S OSSEs show GAINs over VERwest of 0.20 and sometimes as high as 0.70. Targeted R B S O S S E show GAINs close to the maximum. S C M - D A N E T W O R K THINN N G - HT70 RMSD (m) - 8 cases VERwest OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REFOO 0.0 16.7 20.4 21.6 23.7 29.1 36.9 BEN05 1.0 16.1 20.7 23.6 26.3 31.6 41.0 BEN06 2.0 16.1 20.7 24.2 26.8 31.9 41.5 BEN07 3.0 16.2 20.6 24.2 27.1 32.2 41.3 BEN08 3.9 16.3 20.3 23.8 26.7 31.9 41.5 BEN09 5.1 16.5 20.1 23.6 26.5 32.1 42.2 BEN10 6.3 16.5 20.3 23.8 27.6 34.4 43.3 REF12 13.3 18.8 20.6 25.4 30.3 37.1 44.7 VEReast OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REFOO 0.0 13.1 17.0 22.8 25.1 35.2 42.2 BEN05 3.5 14.6 18.6 24.9 27.1 37.7 46.5 BEN06 5.3 14.8 18.4 24.4 26.8 37.3 45.7 BEN07 5.1 14.5 18.2 24.3 26.4 37.1 46.3 BEN08 5.7 14.8 18.3 24.5 26.8 37.6 46.4 BEN09 5.5 14.6 18.1 24.3 27.3 38.1 46.3 BEN10 5.6 14.6 18.6 24.6 27.5 38.9 48.6 REF12 11.1 16.2 21.5 28.6 31.2 42.9 53.5 3DVAR-DA N E T W O R K THIN NING- HT70 RM [SD (m) - 3 cases VERwest OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REFOO 0.0 19.1 23.6 23.1 22.7 23.1 31.0 BEN05 13.3 17.4 20.4 24.0 26.2 26.9 35.9 BEN06 12.8 17.0 19.5 24.2 26.9 28.1 37.3 BEN07 13.0 16.9 19.3 24.1 27.3 29.0 37.8 BEN08 13.1 17.5 19.5 24.8 27.5 29.3 38.3 BEN09 13.4 17.4 20.0 24.8 27.9 30.0 38.7 BEN10 13.5 17.3 19.7 25.1 27.5 28.9 37.4 REF12 19.9 24.8 24.9 27.0 31.6 35.6 38.1 VEReast OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REFOO 0.0 15.6 17.9 28.5 39.5 48.8 39.2 BEN05 8.3 18.7 20.5 29.0 38.0 47.0 44.5 BEN06 9.0 19.2 21.3 30.8 40.5 48.7 46.8 BEN07 8.5 19.4 21.4 29.8 38.9 48.6 47.2 BEN08 8.8 19.4 21.6 29.6 38.8 47.9 45.4 BEN09 8.6 19.5 21.7 30.9 40.6 51.0 48.7 163 BEN10 8.6 19.3 21.7 30.5 40.3 50.6 47.0 REF12 13.0 20.1 22.8 34.9 46.6 57.3 49.0 Table 9.26. For the Study Period, 8 cases of SCM-DA (top) and 3 cases of 3DVAR-DA vRAOBs are presented at 5,6,7,8,9,10 grid-point increments. The averaged RMSD results are a slowly varying function of vRAOB spacing. HT70 RMSD 48HR ADA REF00 BEN05 BEN06 BEN07 BEN08 BEN09 BEN10 REF12 Figure 9.20. The results of the Network Thinning OSSEs shown on Table 9.26. The winter 02/03 (8 cases) SCM-DA results and the winter 03/04 (3 cases) results show a slowly varying function of vRAOB spacing. Reallocating radiosonde resources to a RBS to balance the spatial density of the observational network looks like a desirable option for VERwest. Since the data void is further upstream, the short-term penalty over VEReast would be compensated by better forecasts over VEReast in the mid-term. It seems that all regions of North America could benefit. The same experiments were conducted for the Study Period cases. The results are shown in Table 9.26. There seemed to be little RMSD change as the vRAOB density was quartered from every 5 grid points to every 10 grid points. 164 The results suggest that the most loss of skil l from network thinning may happen during storm situations (like the 16Mar02 case of Figure 9.19). During lower-impact synoptic situations (like the Study period winter cases of Table 9.26), there seems to be a smaller dependence on v R A O B spacing. Another operational strategy could be to deal with high-impact events through real-time management. A coarser R A O B regime over non-arctic North America and a coarse R B S regime over the Pacific and arctic North America could capture the basic synoptic situation. Storms anywhere in the forecast domain could be treated with supplementary observations when necessary. Extra radiosondes could be launched over non-arctic North America; dropsondes, aerosondes, and extra rocketsondes could be launched over the Pacif ic; extra rocketsondes could be launched over the arctic. 9.14 The Composite Observing System 9.14.1 Possible Improvements In order to mitigate the (Figure 3.2) mid-atmosphere data-void, the composite observing system must be furthered with a mixture of vertical and horizontal profiles. The technical and practical options should be examined. Even i f technically feasible over the oceans, a B E N density network of R A O B S would be very expensive. Section 3.7 has shown that the most likely systems to deliver new in-situ data are aerosondes, dropsondes and rocketsondes. Afterwards an optimum mix of in-situ data and satellite radiance data from the newly deployed sensors (Morss et al. 2001) could provide the most cost-effective improvement. 165 9.14.2 A Composite Mix Simulation OSSEs can provide information on the relative impact of aircraft reports; surface reports; SSMI and S C A T satellite data, aerosonde flight observations, R B S and dropsonde profile data as wel l as their cumulative impact. The 15Oct03/16Oct03 cases leading up to the heavy rainfall event (Section 8.4.3) were used. The virtual dropsonde areas are located using satellite imagery and short-range forecast aids (Section 2.4) similar to the Targeted R B S exercise of Section 9.10. They are well within the actual dropsonde area (Figure 9.21) serviced by W S R (1999-2003). 1C0E 170E 180 17WU 160W 150UI 1401V 130W 120W 110W Figure 9.21. The actual WSR dropsonde locations (1999-2003) are shown. This graphic is courtesy of James Charbonneau and Jenn Mundy of the UBC Prediction Research Team. The W S R dropsonde missions deployed approximately 20 dropsondes per 4-5 hr flight. 20-25 dropsonde profiles and a flight level of 6 km are used for the dropsonde simulation. Although the actual dropsonde flights were one way along a flight route or along the borders of a flight area, the simulation assumed that the aircraft crisscrossed the 166 area and provides profiles at regular intervals (every 5 grid points) valid at the 00Z synoptic hour. The target dropsonde areas were just ahead of the low-pressure troughs shown in Figures 9.13a and 9.13c. Figure 9.22 shows the aerosonde flight routes at 70-kPa. The heavy dashed line shows a heavy used comrnercial aircraft flight route from east As ia to B C . The flight level is at 25 kPa. laOW 17QW 160W 120W 80W 70W_ 60W 140W 130W 120W H O W 100W Figure 9.22. Aerosonde flight paths and commercial-aircraft flight paths used for the composite mix simulation. These are non-dropsonde flights, so horizontal flight-path data is gathered only at the altitude of the aircraft. Each of the simulation scenarios is explained below. Each vertical profile has observations from the surface to 6 km. Each horizontal flight has an observation every 5 grid points along the flight path. For area-coverage simulations, the observation spacing is every 5 grid points. ErnPACRSB ThreeBuoy, SixBuoy, and TwelveBuoy Atmospheres: These simulations use regular R S B vROBs according to the sampling strategy of Figure 7.3. 167 Targeted RSBs: These simulations use the Targeted R S B vROBs taken in advance of the low-pressure areas and low-pressure troughs (Section 9.10). AEROone: These simulations use single level 70 kPa flight line observations along the line indicated by aerosonde 1 on Figure 9.22. This scenario simulates the contribution from one aerosonde flight. AEROthree. These simulations use the single level 70 kPa flight line observations along the lines aerosonde 1 to 3 shown on Figure 9.22. This scenario simulates the contribution from three aerosonde flights. AIRone: This is the result of adding a series of commercial aircraft weather reports along the 25 kPa flight route shown in Figure 9.22. This scenario simulates the contribution of multiple aircraft reporting along one flight route. AIR_SFC: SHIP, B U O Y , S C A T SSMI and A I R E P data are routinely available over the Pacific. To simulate this environment, observations were added every 5 grid points at 25 kPa and at the surface throughout the B E N area. This is approximately equivalent to the data-sandwich of Figure 3.2. AIR_SFC_TwelveBuoy: To simulate the contribution of v R O B s within the data sandwich, the TwelveBuoy vROBs were assimilated with the A I R S F C data. Dropsonde: To approximate the impact of a dropsonde program, these simulations included 25 virtual-dropsonde profiles in the area just ahead of the low-pressure trough position and under the cloud area. (Figures 9.13a & 9.13c). A 6 km aircraft flight level is assumed. B E N and REFOO: These are the routinely used benchmark atmospheres. 168 THE COMPOSITE OBSERVING SYSTEM - IMPACT OF THE VARIOUS COMPONENTS - HT70 RMSD (m) VERwest OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REF12 18.8 22.6 34.9 36.3 38.7 41.0 24.7 Three Buoy 19.9 21.8 35.3 33.9 37.0 38.9 23.9 Six Buoy 20.4 22.6 35.8 31.2 34.1 37.4 25.0 Twelve Buoy 19.0 19.7 30.2 24.8 29.8 36.9 25.4 4 or 5 Targeted RBS 20.2 21.8 34.0 32.2 35.5 38.5 23.9 AEROone 19.8 22.0 36.0 35.7 37.5 39.4 25.1 AEROthree 19.8 22.1 35.9 34.7 35.3 37.7 24.6 AIRroute 19.0 20.8 34.6 34.0 35.4 38.0 24.4 AIR SFC 10.9 19.0 31.9 32.3 35.6 38.8 24.9 AIRSFCTwelveBuoy 11.7 22.6 35.6 31.3 31.7 32.1 22.4 Dropsondes 10.7 21.5 33.1 • 27.4 31.3 36.1 25.6 BEN 7.4 18.7 30.7 24.5 23.3 27.5 24.3 REFOO 0.0 18.6 24.1 22.8 28.8 24.4 18.1 GAIN OVER REF12 - RELATIVE TO BEN VERwest OOhr ADA 12hr ADA 24hr ADA 36hr ADA 48hr ADA 60hr ADA 72hr ADA REF12 0.00 0.00 0.00 0.00 0.00 0.00 0.00 Three Buoy -0.10 0.20 -0.10 0.20 0.11 0.15 >1.00 Six Buoy -0.14 -0.01 -0.23 0.43 0.30 0.26 -0.72 Twelve Buoy -0.01 0.75 >1.00 0.98 0.58 0.30 <1.00-4 or 5 Targeted RBS -0.12 0.20 0.20 0.34 0.21 0.18 1.99 AEROone -0.09 0.15 -0.28 0.05 0.08 0.12 -0.88 AEROthree -0.09 0.13 -0.26 0.13 0.22 0.24 0.24 AIRroute -0.02 0.46 0.06 0.19 0.22 0.22 0.68 AIR SFC 0.69 0.92 0.70 0.34 0.20 0.17 -0.54 AIR SFC TwelveBuoy 0.78 0.46 0.80 0.70 0.65 0.79 >1.00 Dropsondes -0.16 0.29 0.42 0.75 0.48 0.36 <-1.00 BEN 1.00 1.00 1.00 1.00 1.00 1.00 1.00 Table 9.27. Relative impacts of simulated aerosonde, RBS, aircraft, surface and dropsonde data. The averaged results (for 15Oct03 and 16Oct03) are shown in Table 9.27. It is risky to generalize simulated results based on two cases (Table 9.27). The data indicates that vertical profiles contribute substantially more than the horizontal observations. This finding suggests that the rocketsondes would be more effective than aerosondes. This 169 finding is in agreement with Barker et al. 2003. Starting with the AIR_SFC scenario, then including the TwelveBuoy vROBs increased the GAIN substantially in the critical 36-48 A D A period. One of the advantages of an RBS operation over both an aerosonde and dropsonde operation is the timing of the observations. An RBS has synoptic observations; the others have asynoptic observations. In these simulations, the contribution from aerosondes and dropsondes may be overstated. The simulations assume the data is valid at 00Z. In reality the observations would be spread out over several hours. 4DVAR manages asynoptic observations at the expense of computer resources. E C M W F does run 4 D V A R operationally, but on a global scale with coarse spatial resolution. The adjoint model contains limited and simplified physics (Ruggiero et al. 2002). The average Targeted RBS results did not show great success over VERwest. The impact of the 15Oct03 targeting remained offshore and did not affect the VERwest area much. However, it did succeed in lowering the heights over the eastern Pacific suggesting that the subsequent analyses would have been much closer to reality (See Figure 8.13). The 16Oct03 Targeted-RBS case was successful (Table 9.23). Robust statistics are required. 9.15 Summary Absolute changes in forecast error are referred to in % terms (e.g., 23%). Most results are measured relative to a benchmark forecast, with the benchmark forecast error forming the standard. The results are then normalized by the difference between the reference forecast error and benchmark forecast error (e.g., 0.23). The benchmark results 170 are the M M 5 results starting with the D A from an idealized observing network or the next E D A S update. Weather Regime If proper conclusions are to be drawn from OSSEs, the weather regime must be considered. The weather regime can have a noticeable effect on N W P forecast error growth and the effect can mask the effects of other deficiencies such as the error growth due to representative initial conditions. Winter 01/02 had mostly zonal flow, while winters 02/03 and 03/04 were dominated by meridional flow. For the Study Period, the 60-hour eastern North America forecasts verified worse than the western North America forecasts. because of this meridional flow. The results from the Test Period winters and the Study Period winters would probably lie on the opposite sides of climatology. Data Assimilation For the initialization process, D A is just as important as data availability. Not even the powerful E D A S is able to improve the Pacific analysis consistently, and it often does so unevenly. The failure rate is about 15%. On average, the 12-hour E D A S updates for the non-failed analyses result in a 20% N W P improvement over western North America. This routine improvement is attributed to the (EDAS) 3 D V A R - D A of available in-situ observations supplemented by satellite radiance D A . Since the O S S E virtual data is extracted from the E D A S updated analysis, this averaged 20% improvement level is the most that is expected from the O S S E results. 171 The averaged M M 5 O S S E results are scaled by this 20% improvement goal. To reach this objective, it is mandatory to calibrate the M M 5 - D A schemes. The sensitivity exercise may require a separate study to provide robust information; the results may be specific only for the model used (MM5) and its implementation. The simpler S C M - D A scheme delivered acceptable results used for further O S S E design. With 3 D V A R - D A , the error-level sensitivity for each type of observation is investigated individually. Data-quality control is crucial; the observations must be compatible with the background field, otherwise the forecast skil l may deteriorate. Regional NWP Over western North America, E T A delivers better results than M M 5 by 20-25 %. This is a probably a result of a larger domain, better initialization, better model resolution and possibly better modeling (e.g. the mountain effects). Data forming the initial and boundary conditions should have higher resolution than the E T A 104 data product (used in this study). If mesoscale N W P is going to progress, this requirement is mandatory. The first 12 hours of N W P produces about half of the 60-hour forecast error. After the first 12 hours, forecast-error growth rate is lower and total forecast-error growth is approximately linear. OSSEs The impact of all of the current and proposed in-situ observation systems can be simulated by the O S S E procedure used in this study. To be successful, the O S S E forecast results must remain bounded by the reference and benchmark forecast results suggesting (that relative to those forecasts) the O S S E error 172 growth is linear. The N W P integration time frame for this linear regime is limited. For the M M 5 model configuration used, the reliable forecast period should not extend past 60 hours for no data addition and not past 48 hours after data addition. With caution, a 12 hour extension to 72 hours and 60 hours respectively could be considered. Outside of that time frame, O S S E forecasts may descend into chaos. Data-voids The averaged Pacific data-void penalty paid by western North America winter forecasts is near 20% and could be as much as 35% for high impact events. Pacific data routinely starts to affect eastern North America forecasts 36 hours after initialization. During high-impact events, the Pacific data provides little value for eastern North America forecasts. During routine synoptic situations, the western North America observations have (up to) double the impact of the Pacific data. The same conclusion would apply to western North America forecasts i f a reliable observing system were located upstream. The study of the polar data void is difficult. For verification over eastern North America, the cases would have to be chosen for their appropriate synoptic situations. Not studied here is how the Arctic data-void impacts D A and N W P over Europe. RBS Operations To increase its range of applications and cost-effectiveness, the R B S should support meteorological, oceanographic and seismic missions. The R B S should have a real time management capability. The R B S should routinely support the Pacific 12Z synoptic analyses. The 12Z analyses have a 6-12% penalty compared to the Pacific OOZ analyses; the penalty rises to 173 about 19% during high-impact events. With real-time management flexibility, the R B S should support critical-situation 00Z synoptic analyses. Operations should concentrate on the late-Fall, Winter, and early-Spring period or approximately 200 days per year and become dormant during the summer season. With real-time management, the R B S could awaken for high-impact summertime events. For the model resolution and cases studied it was found the correct N W P forecasts depend most strongly on getting the middle tropospheric structure correct. R B S profiles to 6 km altitude offer the optimum results; often 4 km profiles may be adequate. A multi-rocketsonde R B S with both 4 and 6 km rocketsondes should be developed. Forecast results from wind observations match the results from pressure-temperature observations. Including wind observations sometimes doubles the benefit, suggesting that the R B S should provide the wind observations. A 5% launch tilt was found to be acceptable. RBS Deployment A n initial "near coast R B S deployment" could be a three-RBS array located 200 km off of the B C and Washington outer coasts. A "medium range R B S deployment" could be a s ix -RBS array centered near 50N 145W and could have a cross configuration. A 12-15 R B S array over the eastern Pacific with the same spacing as the continental radiosonde network would be optimum. Coverage to the southwest seems necessary to provide lead-time for the tropical-origin heavy-rain events. Larger R B S arrays and real-time management would allow targeting activities and increase the R B S cost-effectiveness. If R B S technology can be adapted for the Arctic, it could be an affordable alternative to staffed radiosonde stations. 174 Forecast impact On average, the Pacific data void deteriorated the forecasts by 20%. Most forecast gain-per-buoy was with the smaller arrays; most total gain was with the larger arrays. The forecast improvements were case dependent, a function of the R B S locations and the synoptic situation. When a three-RBS array was strategically located, the data availability typically recovered 0.70 of the deterioration; sometimes nearly all o f the deterioration was recovered. The average recovery result was about 0.30 for a s ix-RBS array rising to about 0.60 for a twelve-RBS array. Targeting raised the recovery rates. The results seem to confirm the Pailleux et al. (1998) A S A P (real data) findings (Section 9.8.2). For the Arctic operational simulations, a s ix-RBS operation recovers up to 0.25 of the Arctic data void deterioration. Over non-arctic North America, radiosonde and aircraft-sounding data may be redundant. Enlarging the sounding network westward over the northeast Pacific Ocean at the expense of thinning the current radiosonde network over North America would lead to much better western North America forecasts. This spreading of the network wi l l result in a small short-term penalty paid by the eastern North America forecasts, but wi l l result in a compensating mid-term gain. With a 50% thinning of the radiosonde network, the forecast-skill deterioration was 0.05-0.10 during high impact events becoming very small for routine synoptic events. The savings accrued by thinning the radiosonde network could be reallocated towards R B S deployment and targeting operations. The forecasts over western North America would improve; the resolution of storms affecting eastern North America could be 175 enhanced with targeting operations, eliminating the "network-tmnning" penalty. A l l areas of North America would benefit. Comparison with Others Vertical profiles have more impact than flight-level observations, so R B S and dropsonde program development should be encouraged. R B S operations wi l l have an advantage over dropsonde operations; the decision to launch wi l l need a very short lead-time. Targeted-RBS locations could be determined manually based on the interpretation of satellite imagery alone. A dropsonde mission requires a much longer lead-time (24 hours or more) increasing the risk of misdirection. As well, the R B S sounding data would be synoptic, probably increasing the data effectiveness, unlike the dropsonde soundings that are spread over several hours or more. General Remarks These O S S E results, being the product of virtual-data simulations, should be viewed as estimates of the results that may occur i f analogous real data were used. The conclusions are valid only for the cases studied; there is no suggestion that the results should be generalized. More experimentation and robust statistics are necessary. L ike any other simulation procedure (e.g. N W P ) , the O S S E procedure should be verified; in this case analogous real-data O S E results are required. In the absence of real data, OSSEs are the only vehicle available for the study of proposed observing systems. 176 Chapter 10 Conclusions 10.1 Summary of Findings To minimize Numerical Weather Prediction (NWP) initialization and forecast errors, both data availability and data assimilation are equally important. In data-void areas like the Pacific, current data assimilation (DA) procedures do not always improve the analysis. Currently, the Eta Data Analysis System (EDAS) forecast failure rate is about 15%. Even when large-area verification shows an improved forecast, the analysis may deteriorate the forecast locally. Based on E T A 104 data-product verification results, E D A S seems to deliver a 20% averaged improvement every 12 hours. The Observing System Simulation Experiment (OSSE) design implemented in this study used the Successive Correction Method (SCM) and the three-dimensional variational ( 3 D V A R ) method for D A , and used the Mesoscale Model Generation-5 (MM5) model for N W P . S C M was used for the initial half of the study; 3D V A R was used afterward. Because o f its simplicity S C M was easier to work with. Overall, S C M and 3 D V A R delivered comparable results. For both, it was found that the data assimilated must be compatible with the background field; otherwise the D A wi l l fail to provide N W P value. The D A schemes must be well calibrated to deliver optimum results. Since S C M - D A is empirical and 3 D V A R - D A has semi-empirical settings, the sensitivity exercises required appreciable experimentation. M M 5 forecast error growth is largely determined by the ability of the model to simulate the weather regime, which can vary with the study period. During high-impact 177 situations, wintertime-storm forecast-error is more than double that from summertime activity. On average wintertime error-growth rates are higher than summer and result in greater forecast-error (18-25%) than summertime activity. Much of the error-growth occurs in the first 12 hrs; afterward it is approximately linear. The Pacific data-void penalty paid by western North America (WNA) relative to eastern North America (ENA) is over 20% and can reach 35% for high impact events. The forecast ski l l over W N A and E N A equalizes after two to three days as the effects of the Pacific and polar data-voids spread over both regions. When E N A has a more active weather regime than W N A , the larger E N A forecast error-growth reverses the usual forecast skil l comparison. Over W N A , averaged OOZ analyses deliver better-forecast results than the 12Z analyses. The Rocketsonde Buoy System (RBS) engineering constraint is currently 200 of the 6 km altitude rocketsondes per buoy. If the R B S operational period is kept to seven months centered on the winter period delivering one 6 km profile at 12Z, the engineering constraint would be met. A study of 12-hour forecast errors shows a maximum just off the west coast during high impact events. A 254 winter-case average suggests a broad maximum near 50N 145W. The mountain and Arctic areas show local areas of high forecast error. The effect of the polar data-void is difficult to quantify. The ultimate R B S requirements may be 12 to 15 units deployed over the eastern Pacific about 300 km apart in a regular grid, equivalent to the non-polar North American radiosonde network. A phased-deployment strategy would suggest an initial deployment of three RBSs 200 km off the west coast of North America; these would protect the coastal 178 radiosondes from rejection by the D A system during high-impact events, and provide some lead-time when analyzing heavy-rain events. A subsequent deployment of six R B S s in a cross arrangement centered on 50N 145W would decrease the large 12 hour forecast error routinely found near there. The full complement of 12-15 RBSs could then follow. 12 RBSs deployed in the central Pacific w i l l move the data void further west. Based on the analysis of the E T A forecasts, the 20% averaged improvement is the maximum expected from the averaged O S S E results, and represents a realistic goal. During high-impact situations, both the three and six eastern Pacific R B S deployments can reduce the forecast error by about 14%, namely 0.70 of this goal. In certain cases, most o f the possible forecast error reduction was produced. The same impact was seen with a few targeted RBSs . On average, a s ix-RBS deployment delivers 0.30 of the possible improvement; the full deployment of 12-15 RBSs provides 0.60 of the possible improvement. These findings are in line with Pailleux et al. 1998; they found that at times one or two Atlantic (real data) A S A P radiosondes improved the analysis, but 10 were needed for routine improvement. The optimum profile height seems to be 6 km; higher apogees yield gains with diminishing returns. The R B S sondes should be designed to deliver both mass observations (temperature, humidity) and wind observations. For the cases examined, the mass and wind observations contribute to forecast improvement fairly equally, sometimes their contributions are cumulative. A 5% launch tilt was found to be acceptable. A targeting strategy for RBSs , dropsondes and aerosondes can deliver close to the averaged improvement goal of 20%. Simulations including some or all of these observing systems suggest that vertical profiles deliver more value than flight-level profiles. 179 During high-impact events, it was found that upstream eastern Pacific R B S observations are just as important as W N A radiosonde observations. Reallocation of observation resources from the continent to the eastern Pacific would help the W N A forecasts. During routine synoptic situations, a 50% thinning of the North American radiosonde network may deteriorate forecast skil l over E N A a very small amount. During high-impact events, forecast skil l may decrease 0.05 to 0.10 due to the less exacting initialization. Reallocation of resources from the radiosonde network to an R B S would improve the W N A forecasts by 0.60 or more (mentioned above), with only a short-term penalty by E N A during high-impact events. Dropsonde, aerosonde and R B S targeting over the Pacific/arctic areas and radiosonde targeting over the continent would eliminate any short-term E N A penalty while improving forecast skil l over all North America. 10.2 Critique of the OSSE Method A weakness of this O S S E method is the source of the virtual data. Data extracted from a model may not resemble the real world. The same errors in synoptic system strength and location may be common in both the virtual soundings and model first-guess. The result may be an underestimation of the improvements possible with a real R B S system. B y necessity, analyses over data-void areas cannot capture extremes; the virtual data is never far from the first-guess field. This is not necessarily so in the real world. This commonality would tend to lower the impact of the virtual data, leading to conservative O S S E results. The conservative nature of these results suggests that they could form a lower bound to similar results using Pacific real-data OSEs, when real R B S and other in-situ data become available. 180 Also, the O S S E procedure does not consider possible R B S data interactions with the other routinely available observations, except with simulated data having the same limitations. Since these other observations already provide value, their presence may diminish the full impact of regular R B S profiles. The full impact of R B S observations may be felt only when the other observations do not provide value during a routine data assimilation, a situation that often happens. The 3 D V A R and four-dimensional variational ( 4 D V A R ) methods have gained enormous popularity in both the research and the operational environments. They have a theoretical justification and account for observational error. After their development, the older empirical schemes (SCM) tend to be viewed with uncertainty; there is no theoretical justification, and observations are incorrectly assumed to be error-free. However, S C M -D A has certain advantages. The range of data impact is easy to envision. Quality-control criteria are easy to implement. The calibration exercise is straightforward. Computer resources are saved. S C M - D A can be a good tool for starting a goal-oriented O S S E design. In spite of their weaknesses, OSSEs are the only a-priori means available to comprehensively anticipate the impact of the R B S and other proposed observing systems. It is impossible to fully represent a real-world situation using a computer model. Therefore O S S E results are only estimates of what may happen in the real-world. Observing System Experiments (OSEs) should be performed with data extracted from the upcoming THORpex field tests. Such comparisons would test the validity of the OSSEs and help refine the procedure. 181 10.3 Tips for Future OSSE Work Here are guidelines for any successor project: • Make the regional domain as large as possible. • • For the initial and boundary conditions, provide the best data resolution. • Before beginning the mainstream study, experiment heavily with the D A scheme(s) using a variety of synoptic situations; the objective is the optimum calibration of the D A scheme, where every model implementation may require different D A settings. • Make up a detailed list of the experiments and the goals of each. • Use the same synoptic case set for as many of the experiments as possible. • Keep the areas of interest well away from the boundaries. • Archive the data products, as they w i l l be needed repeatedly for reruns. • Archive other supporting information, such as synoptic maps and G O E S satellite images available from other sources. • Lastly, use a historical collection of cases to broaden the representativeness. 10.4 Weather Observations in an Ideal World In the future, the data-voids must be filled. The observing system must be better managed. Better D A procedures must be developed. A n important advance has been the optimization of the data-collection process through targeting of observations; a strategy likely to be further embraced. The future observation suite could be placed into three categories; (1) the opportunistic in-situ and remotely sensed observations (the cheap observations), (2) the coarse set of synoptic vertical profiles taken by widely spaced over-land radiosondes, over-arctic and ocean 182 rocketsondes (taken routinely except where it has been predetermined that the data wi l l be of little value for the expected synoptic situation), (3) the targeted observations taken by mobile systems able to deploy quickly. A future observing system might have the following components. Commercial aircraft continue to provide observations, as do ships and buoys. Satellite sensors continue to provide active and passive radiance information. A coarse network of land radiosondes and ocean/arctic rocketsonde provides profiles at the main synoptic hours except when the data is determined to be of little value. The smart D A system provides regular analyses. The future operational scientists identify all features of interest and societal impact within their sector. The tools are their experience and the support of artificial intelligence systems driven by the latest generation of super-computers. Extraordinary-resolution N W P models make 10-day projections. The features are identified, the deployment plan is formulated and the mobile systems are activated. The following fictional scenario illustrates such a future observation system for an approaching cyclone. Routine observations from coarsely-spaced radiosonde and rocketsonde stations yield a synoptic analysis with no significant data voids. The analysis is used to drive a 10-day coarse resolution N W P forecast. The forecast suggest the genesis of an extratropical cyclone upstream from North America. Aerosondes are sent into the deployment area timed to expect the feature's further development. The smart DA-scheme provides a reanalysis as the data become available; the extraordinary-resolution N W P model makes more projections. The feature's further development is identified. A dropsonde mission is deployed saturating the area with observations. The feature continues to develop. Another reanalysis is generated. The extraordinary-resolution N W P model 183 provides an ominous projection. The 10-day N W P forecast shows a storm on day-9, with heavy wind, sleet, rain and snow over British Columbia. The emergency-response system is triggered, civic authorities are informed; day after day the public is warned of the day-9 storm approach; preparations are made. On day-9, the storm arrives right on schedule. British Columbia is braced for the storm; much damage is averted; lives are saved. 184 BIBLIOGRAPHY Arnold, C P . , and C. H. Dey, 1986: Observing Systems Simulation Experiments: Past, Present, and Future., Bull. Amer. Meteor. Soc, 67, 687-695. Bergot, T. and A . Doerenbecher, 2002: A study of the optimization of the deployment of targeted observations using adjoint-based methods. Q. J. R. Meteorol. Soc., 128, 1689-1712. Barker, D., and W. Huang, Y . Guo and A . Bourgeois, 2003: A Three-Diniensional Variational ( 3 D V A R ) Data Assimilation System for Use With M M 5 , M M M Division, N C A R , available at P.O. Box 3000, Boulder, C O , 80307-3000, U S A , 73 pp., or available on line at [http://www.mmm.ucar.edu/rmn53dvar/docs/3DVARTechDoc.pdf.] Bratseth, A . M . , 1986: Statistical interpolation by means of successive corrections. Tellus, 38A, 439-447 Chan, K., 2002: Forecasting Surface Winds Using the Pressure Slope Technique, Meterological Service of Canada, 26 pp. Available from: Chris Doyle, Meteorological Service o f Canada, Pacific Yukon Region, 200-401 Burrard Street, Vancouver, B C , V 6 C 3S5, Canada. Clark, D., 1965: A Meteorological Rocket Sonde, Journal of Scientific Instruments, Volume 42, Issue 10, pp. 733-736 (1965). Dabberdt, W. , and Shellhorn, R., 2003: Radiosondes, available from Vailsala Inc., Boulder, C O , U S A , 14 pp. Available from: http://rain.atmos.colostate.edu/dabberdt.pdf. Deng, X . and R. Stull, 2005: A Mesoscale Analysis Method for Surface Potential Temperature in Mountainous and Coastal Terrain. Mon. Wea. Rev., 133, 389-408. Doerenbecher, A . , and T. Bergot, 2001: Sensitivity to observations applied to F A S T E X cases, European Geophysical Society, 8, 467-481. 185 Doyle, C , 2001: G E M Initialization and Forecast Errors in the Eastern Pac i f i c , Examination and Coping Strategies, Meteorological Service of Canada, 24 pp. Available from: Chris Doyle at the Meteorological Service of Canada/ E C , Pacific Yukon Region, 200-401 Burrard Street, Vancouver, B C , V 6 C 3S5, Canada. Dudhia, J . , D. G i l l , K. Manning, A . Bourgeois, W. Wang, and C. Bruyere, 2003: P S U / N C A R Mesoscale Modeling System Tutorial Class Notes and Users' Guide ( M M 5 Modeling System Version 3). National Center for Atmospheric Research, Mesoscale & Microscale Meteorology Division, 340 pp. Available from N C A R , P O Box 3000, Boulder, C O 80307-3000; or from [http:/^ox.rnrrim.ucar.edu/mm5/doc.html]. E C M W F , 2005: Available at [http://www.ecmwf.int]. Gilmour I., L. Smith, R. Buizza, 2001: Linear Regime Duration: Is 24 Hours a Long Time in Synoptic Weather Forecasting?, Journal of Atmospheric Sciences, 58, 3525-3539. Graham, R.J. , Anderson, S.R., Bader, M.J. , 2001: The Relative Uti l i ty of Current Observation Systems to Global-Scale N W P Forecasts, Q.J.R. Meteorol. Soc, 126, 2435-2460. Hacker, J.P, S.Krayenhoff, and R.B.Stull, 2003: Ensemble Experiments on Numerical Weather Prediction Error and Uncertainty for a Major North Pacific Forecast Failure. J. Weather and Forecasting, Vol. 18, No. 1, pp. 12—31. Kalnay, E., 2003: Data Assimilation. Atmospheric Modeling, Data Assimilation and Predictability, Cambridge Univ. Press. Cambridge, U.K. 130-204. Langland, R.H. , M .A . Shapiro, A . J . Thorpe, 2002: Proposal to the W W R P / W G N E : THe Observing System Research and predictability experiment THORpex, 40 pp. Available online at [http://www.mmm.ucar.edu/uswrp/thorpex/THORpex_wmo.pdf]. Lorenz, A . C . , 1975: Results of Observing System Simulation Experiments for the First G A R P Global Experiment. G A R P Working Group on Numerical Experimentation, Report #10, pp. 37-68. 186 Lorenz, R , 1993: The Essence of Chaos. University of Washington Press, Seattle. I S B N 0-295-97514-8, 240 pp. McCol lo r D., 2004, Heavy Rain Event, B C South Coast and Southern Interior, 17 to 22 January 2005, Western Canada Weather Workshop, 2004, 25 pp. Available online at [http://KStream.pyr.ec.gc.ca/flashcom/applications/WCWW/default.htm McNal ly , A .P . , 2000: The Occurrence of Cloud in Meteorologically Sensitive Areas and the Implication of Advance Infrared Sounders. Technical proceeding of the 11th International TOVS Study Conference, Budapest. Hungary, pp. 20-26. McTaggart-Cowan, R., 1998: Feasibility Analysis of a Buoy-Deployed mid-Tropospheric Measurement System in the North Pacific. Available from Professor Roland Stull. University of British Columbia. Mi l ler, P. A . and S. G. Benjamin, 1992: A system for the hourly assimilation of surface observations in mountainous and flat terrain. Mon. Wea. Rev., 120, 2342-2359. Morss R.E. , K. A . Emanuel, and C. Snyder, 2001: Idealized adaptive observation strategies for improving numerical weather prediction. J. Atmos. Sci:, 58, 210-234. N R L , 2002: N R L Predictability and Adaptive Observation Program. Available at http://www.nrlrnry.navy.mil/shared-bin/adap/adap _prod2.cgi?BASIN= hemispheric;DIR=sens_hemispheric_p72m72;PROD=ic Pailleux, J , M.Bader, W.Wergen, X.Huang.H.Bottter, March 1998: Improving E U C O S in Data-Sparse Areas by Conventional Means. Available from: htxp://www.eumetnet.eu.irg/ contexxos. html. Pielke, R. A . , and J. Kimpel , 1997: Societal aspects of weather. Report on the sixth prospectus development team of the U.S. weather research program to N O A A and N S F . Bull. Amer. Meteor. Soc, 78, 867-876. 187 Readyhough, C , 2003: Design of a Rocksonde Buoy System for Collecting Weather Sounding Data, M A S c thesis, Department of Mechanical Engineering. 120 pp. University of Brit ish Columbia. Available from [http://weather.eos.ubc.ca/ RocketsondeBuoy /rsbuoy.htm]. Rogers, E, T. Black, B. Ferrier, Y . L in , D. Parrish, and G. DiMego. Changes to the N C E P Meso Eta Analysis and Forecast System: Mesoscale Modeling Branch and * Science Applications International Corporation, Environmental Modeling Center, National Centers for Environmental Prediction, 34 pp. Available online at [http://www.emc.ncep.noaa.gov/ mmb/mmbpll/eta 12tpb/] Ruggiero, F., D. Norquist, T. Nehrkorn, G.D. Modica, M . Cerniglia, J .G. Michalakes, and Z. Zou, 2002: MM5V3-Based 4 D V A R : current status and future plans, 12th P S U / N C A R mesoscale model users'workshop, N C A R , June 24 - 25, 2002., 4 pp. Available online at [http://www.mmm.ucar.edu/mm5/workshop/ws02/Ruggiero.pdf]. Shapiro, M . A . and A . J . Thorpe, 2002: Program Overview: THe Observing-system Research and predictabUity experiment THORpex., 16 pp. Available from http://www.mmm.ucar.edu/uswrp/thorpex/thorpex_planl 3 .pdf. Shapiro, M .A . and A . J . Thorpe, 2004: T H O R P E X : THORPEX International Science Plan. 45 pp. Available from http://www.nimm.ucar.edu/uswrp/thorpex/ T H O R P E X _ L A T E S T _ 2 5 . p d f . Spagnol, J . , M . Horita, and P. Haering, 1980: Analyses of the eastern Pacific without Ship Papa data. Preprints, 8th Conf. on Weather Forecasting and Analysis, Denver, C O , Amer. Meteor. S o c , 150-155. ' Spagnol, J . , 2005: Rocketsonde Buoy web site. Available online at [http://weather.eos.ubc.ca/RocketsondeBuoy/rsbuoy.htm]. Spagnol, J and R.Stull, 2004: The Rocketsonde Buoy Project, Bulletin of American Meteorological Society, V o l 85, pp 9-11. 188 Stull, R., S. Green, M . Yedl in and R. Pawlowicz, 2003. Rocketsonde Buoy System, Research Grant Proposal to N O A A / T H O R p e x . Available from R. Stull, Earth & Ocean Sciences, U.B.C. , 6339 Stores Rd. , Vancouver, B.C. , V6T1Z4. Staudenmaier, M . , 1996: A Description of the Meso Eta Model , Western Region Technical Attachment, No. 96-06. N O A A / N W S , 5 pp. Available from National Weather Service, 125 South State St., Salt Lake City, U T 84103, or from http:// www.wrh.noaa.gov/index7.shtml , or more specifically http://www.wrh.noaa.gov/ wrhq/96TAs/TA9606/ta96-06.html. Szunyogh, I., Z . Toth, R.E. Morss, S.J. Majumdar, B.J. Etherton, and C . H . Bishop, 2000: The effect of targeted dropsonde observations during the 1999 Winter Storm Reconnaissance Program. Mon. Wea. Rev., 128, 3520-3537. Szunyogh, I., Z. Toth, A . V . Zimin, S.J. Majumdar, A . Persson, 2002: Propagation of the effect of targeted observations: The 2002 Winter Storm Reconnaissance Program. Mon. Wea. Rev., 130, 1144-1165. Toth, Z. , 2004: Winter Storm Reconnaissance Program. Available online at [http://vv ,ww.aoc.noaa.gov/article_winterstorm.htm]. Vargas, S., 2004: World Weather, Watch. Available online at: [http://www.wmo.ch/web/www/www.html]. Wong, P., 2004: Wi ld Weather in the West, Western Canada Weather Workshop, 2004, 47 pp. Available online at [http://KStream.pyr.ec.gc.ca/flashcorn/applications/WCWW/ default.html]. 189 APPENDIX A - Dropsondes Overview The dropsonde is a meteorological device that is launched from an aircraft or any other airborne platform. Descending through the atmosphere by parachute, the dropsonde measures atmospheric pressure, temperature, relative humidity and wind from the point of launch to the ground. The dropsonde is used with a receiving system that is located onboard an aircraft or a communications satellite. A GPS receiver in the sonde tracks the • dropsondes horizontal movement as it is borne by the wind, to infer wind speed and direction. The dropsonde electronics board has a microprocessor for measuring and controlling the sensor module and data transmission. The narrowband transmitter can be set anywhere in the 400Mhz meteorological band. A parachute deploys immediately upon launch. It slows and stabilizes the dropsonde's descent and ensures that it does not descend with a pendulum motion. The rate of descent is directly proportional to the size of the parachute and can be anywhere between 20 m/s and 5 m/s. The sensor vertical sampling rate is directly proportional to the rate o f descent. 190 Table A.1: Pressure, temperature and humidity sensors and their operating principles Pressure Sensor Operating Principle B A R O S W I T C H Pressure is measured by means of a baroswitch wh ich employs an expanding aneroid pressure cell to move a contact arm across a commutator bar as the pressure decreases (Ref. 1) ANEROID Capsule made of metal with elastic properties deflects with changes in atmospheric pressure (Ref. 1). BOURDON TUBE A tube of elliptical cross section changes in cross section and length as a function of atmospheric pressure (Ref. 1) BELLOWS T Y P E Flexible bellows respond to changes in atmospheric pressure (Ref. 1 ) . H Y P S O M E T E R A n electrical thermometer measures the boiling point of a liquid, which is a function of atmospheric pressure (Ref. 1) . B A R O C A P ANEROID SENSOR Small aneroid capsule responds to changes in pressure measured by capacitive transducer plates inside; utilizes a capacit ive transducer wi th a vacuum inside the capsule; entire unit is precision welded requiring no mechanical adjustment; unit is friction free and cont inuously variable. (Ref: ht tp: / /www.atd.ucar.edu/r t f / faci l i t ies/c lass/c lass.html B A R O C A P SILICON SENSOR Sil icon diaphragm bends and changes the height of the vacuum gap in the sensor; changes the capacitance of the sensor, which is measured and converted into a pressure reading (Ref: http: / /www.vaisala.com/page.asp7Sect ion = 5661) Temperature Sensor Operating Principle WIRE RESISTOR The resistance of a metal wire changes with temperature (Ref. 1). BIMETAL-COIL,STRIP Two metals of different expansivity (usually invar and steel or invar and brass) are riveted or welded together so that the element bends when heated (Ref. 1 ) . THERMISTOR-IROD, B E A D , CHIP) The electrical resistance of a an element changes with temperature (Ref. 1) T H E R M O C A P -CAPACIT IVE BEAD The capacitance of a dielectric ceramic bead changes with temperature (Ref: .http://www.atd.ucar.edu/rtf/facil i t ies/class/class.html) Humidity Sensor Operating Principle H U M A N HAIR The length of human hair increases with relative humidity; used to move capacitor plates (Ref. 1) G O L D B E A T E R S SKIN A membrane from the intestine of an ox (used also to separate gold leaf) changes length in response to humidity changes (Ref. 1 ) . Lithium Chloride HYGRISTOR The resistance of a strip coated with an electrolytic film of lithium chloride increases with increasing relative humidity (Refs:http://www1 .ncdc.noaa.gov/pub/data/stnhistory, ht tp: / /www.aos.wisc.edu/~hopkins/wx-inst/wxi-raob.htm). C A R B O N HYGRISTOR Finely divided carbon particles are suspended in a hygroscopic film Whose length changes wi th : humidity, and the resistance increases with (Refs:http:/ /www1 .ncdc.noaa.gov/pub/data/stnhistory, http://www.atd.ucar.edu/rtf/facil i t ies/dropsonde/dropsonde.html ) H-HUMICAP THIN FILM C A P I C A T O R , (heated twin sensor design) A thin-film sensor whose capacitance varies with relative humidity; polymer serves as a dielectric material on this sensor (Ref: .http://www.atd.ucar.edu/rtf/facil i t ies/class/class.html) 19.1 The dropsonde is the airborne counterpart to the conventional radiosonde (sometimes called an upsonde). Dropsondes were first developed in the 1960s for hurricane reconnaissance and were an adaptation of radiosonde technology. Sonde Sensors Thermodynamic sensor types vary widely among sondes currently in use throughout the world. Sensors and their operating principles are outlined in Table A . l (with internet references). Pressure measurements are typically made with an aneroid cell or piezoresistant sil icon sensor. Temperature sensors available include capacitance sensors, thermistors, resistance wires and bimetallic elements. The two most recently developed humidity sensors are carbon hygristors and planar thin-film capacitance. Descriptions of the sensors deployed on the Vaisala dropsonde are taken from the technical specification sheets and Dabberdt et al. 2003. The recent and most important sensors are discussed in greater detail later in the document. Pressure Sensors: Silicon sensors A n example is the Vaisala BAROCAP® capacitive absolute pressure sensor manufactured by sil icon micro-machining. When the pressure changes, the sil icon diaphragm bends and changes the height of the vacuum gap in the sensor. This changes the capacitance of the sensor, which is measured and converted into a pressure reading. Sensor frequency measurements are compared with the frequencies of reference capacitance transducers and these in turn are converted to physical measurements based on factory calibration measurements. These pressure sensors have a temperature dependence that is compensated by factory calibration. 192 Sil icon sensors offer good elasticity, low hysteresis, excellent repeatability, small temperature dependence and superior long-term stability. A further advantage is its good manufacturability with microelectronic techniques. There is a wide dynamic range and a built-in overpressure blocking mechanism. The sensor has excellent tolerance to mechanical and thermal shocks. Aneroid sensor Older pressure sensors use an encapsulated steel aneroid sensor that responds mechanically to pressure changes. It utilizes a capacitive transducer with a vacuum inside the capsule. The entire unit is precision welded requiring no mechanical adjustment. The unit is friction free and continuously variable. Temperature Sensors Capacitive Bead A n example is the Vaisala T H E R M O C A P temperature sensor which is a capacitive bead in glass encapsulation. The temperature change o f capacitive sensors is measured by the change in the dielectric constant of the sensor. The new capacitive temperature sensors are extremely small and fast, owing to a special twin wire construction. A water repellant treatment and metallisation of the surface ensures minimum radiation sensitivity and excellent performance in rain. Experience has shown that i f the sonde sensor arm is not protected or properly ventilated prior to launch, it can be adversely affected by soiar heating. This results in higher temperature readings. Due to the small thermal mass of the temperature sensor and its supporting structure this effect is not long-lived. The thermal time constant of the 193 sensor arm is 13 seconds and thus the problem goes away soon after launch and the sensor is adequately ventilated. Bead Thermistor A n example is the Fenwal bead thermistor. The electrical resistance of the element changes with temperature. The thermistor is dipped in a lead carbonate paint to provide a coating which reduces solar radiation absorption. The sensors are calibrated at the factory. The manufacturer's specification for the time constant is 2.5 seconds. The time constant of the thermistor produces a slight lag in temperature measurement through the sounding. However, with typical atmospheric lapse rates the resultant smoothing of the temperature profile is less than the accuracy of the thermistor. The smoothing resulting from the lag time becomes more significant when the sonde crosses frontal boundaries or goes through strong inversions. Relative-Humidity Sensors: Thin film capacitor A n example is the Vaisala H U M I C A P . Humidity is measured based on changes in the dielectric constant. This humidity sensing technology is based on so-called thin fi lms. The dielectric material is a very thin layer of special proprietary polymer that has an optimum combination o f measurement properties, including stability, repeatability, hysteresis, response time and temperature dependence. Thin-f i lm humidity sensors are calibrated to provide output in terms of percent relative humidity with respect to water; the temperature dependence is compensated by use of temperature-dependent calibration 194 coefficients determined from factory calibration tests. The probes incorporate two sensor elements that include heating of the sensor elements to minimize affects of water condensation. The two sensors are alternately heated in sequence, and the measurement is taken from the passive sensor. The sensor has good long-term stability and reliable response even at low temperature and after exposure to condensation. This sensor has improved humidity measurements over previous sensors, particularly in the high end of the humidity range. However, the sensor may have a dry bias i f it has been stored for a long time due to contamination from outgassing of the sonde packaging material Hygristor Hygristors use the principle that the resistance to conductivity changes with temperature in metals. A n example is the V I Z / N C A R hygristor. (Internet Reference 2) The sensor has a heated alumina substrate in the sensing area which prevents condensation from occurring. The temperature of the substrate is controlled by using a humidity set point (typically 75%) not allowing condensation to occur. The temperature of the substrate is controlled using the set point and monitored to allow for calculation of the ambient humidity (and dew point temperature). MODERN SENSOR SYTEM Finland's Vaisala Inc produces about 70%) of the world's sondes. Sensors used with Vaisala sondes are all of the capacitance type. The state of the art dropsonde system seems to be the model Vaisala RD93. The RD93 incorporates the latest pressure, 195 temperature, humidity sensor module (RSS903, see image). The sensors are small and designed for fast response. Changes in pressure, temperature and humidity result in changes in capacitance information from each sensor, which in turn is changed to a frequency signal by using sensor transducer electronics. A solid state switch connects each sensor in turn to the transducer electronics. A l l parameters are measured at approximately at 1.5 second intervals. REFERENCES Internet Reference 1: http://wwwl.ncdc.noaa.gov/pub/data/history Internet Reference 2: http://www.atd.ucar.edu/rtf/facilities/dropsonde/dropsonde.html Dabberdt, et al, 2003, located at http://rain.atmos.colostate.edu/dabberdt.pdf Images were taken from http://www.aoc.noaa.gov/instrumentation.htm 196 Appendix B UBC Computer Systems The M M 5 model simulations were run on two computer systems operated by the Geophysical Disaster Computational Fluid Dynamics Centre (GDCFD) , Faculty of Science, University of British Columbia. They are: (1) A 24-processor Beowulf cluster, with Linux as its operating system, where each processor is a Pentium III mnning at 700 Mhz. (2) A 256-processor high performance computing linux super-cluster named Monster, where each processor is a Pentium III running at 1 GHz . Wide-bandwidth internode communications are via Myrinet switches. Monster was bought from the grant awarded to G D C F D by Canadian Foundation for Innovation (CFI), B C Knowledge Development Fund (KDF) , and The University of British Columbia (UBC) . More information may be found at http://www.gdcfd.ubc.ca/Monster/. For the M M 5 N W P integrations, 16 processors were routinely used on Monster and Beowulf. The REF12 runs took about TA hours to complete; the REFOO and O S S E runs took a little over six hours to complete. For the OSSEs with the greatest number of observations (BEN), S C M - D A took about one hour on 2 processors of Monster to complete; 3 D V A R - D A took about 1 1/2 hours on 6 processors to complete. 197 Appendix C Case Dates and Assorted Information This appendix lists the cases that were used for the experiments. A l l cases before 08 August 2002 were used for the Test Period results; all cases after 08 August 2002 were used for the Study Period results. The date format is Y Y M M D D where Y Y are the last two digits of the year; M M is the month number, D D is the day. Each case includes the REFOO on the OOZ hr and the R E F 12 the previous 12Z. For example the 011213 case includes R E F 12 starting at 12Z, 12 December 2001 and REFOO starting at OOZ, 13 December 2001. In the text, the same case was referred to as 13Dec01; i f a number format was used the leading " 0 " would have been dropped to save space in the table columns. When D A was performed, cases from December 2001 to the end of Apr i l 2003 were assimilated with M M 5 - S C M ; afterward the scheme used was M M 5 - 3 D V A R . Three cases were assimilated with both S C M and 3 D V A R (9.5.3). 9.2.1 EDAS Failures The cases listed below were analyzed to determine which subsets yielded E D A S forecast failures. Test Test 020811 020923 021005 Period Period Study 020814 020924 021006 winter summer Period 020815 020925 021008 (5 cases) (4 cases) (373 020820 020926 021009 011213 020702 cases; 020830 020927 021010 020218 020702 summer; 020907 020928 021013 020316 020802 68; 020912 020929 021014 020412 020806 winter; 020916 020930 021015 020506 305) 020919 021001 021016 020808 020920 021002 021017 020809 020921 021003 021018 020810 020922 021004 021019 198 021020 021021 021022 021023 021024 9.2.1 Study Period Cont 'd 021025 021026 021027 021028 021029 021030 021031 021101 021102 021103 021104 021105 021106 021107 021108 021109 021110 021111 021112 021113 021114 021116 021117 021118 021119 021120 021123 021124 021125 021126 021127 021129 021130 021201 021202 021203 021204 021205 021206 021207 021208 021209 021210 021211 021212 021213 021214 021215 021216 021217 021218 021219 021220 021221 021222 021223 021224 021225 021226 021227 021228 021229 021230 021231 030101 030102 030104 030105 030106 030107 030108 030109 030110 030111 030113 030114 030115 030116 030117 030118 030119 030120 030121 030122 030123 030124 030125 030126 030127 030128 030129 030130 030131 030201 030202 030203 030204 030206 030207 030208 030210 030211 030212 030213 030214 030215 030216 030217 030218 030220 030221 030222 030223 030224 030225 030226 030227 030228 030301 030302 030303 030304 030305 030306 030307 030308 030309 030311 030312 030313 030314 030315 030316 030317 030318 030319 030320 030321 030322 030323 030324 030325 030326 030327 030328 030329 030330 030331 030401 030402 030403 030404 030405 030406 030407 030408 030409 030410 030411 030412 030413 030414 030415 030416 030417 030418 030419 030420 030421 030422 030423 030424 030425 030426 030427 030428 030429 030430 030501 031004 031011 031014 031015 031016 031017 031026 031111 031112 031114 031115 031116 031117 031118 031120 031127 031128 031129 031130 031201 031202 031203 031204 031205 031212 040101 040103 040105 040106 040107 040108 040109 040110 040111 040112 040113 040114 040115 040116 040117 040118 040119 040120 040121 040122 040127 040128 040127 040128 040130 040131 040202 040203 040204 040205 040206 040207 040208 040209 040210 040211 040212 040213 040215 040216 040217 040218 040219 040220 040221 040222 040223 040224 040225 040226 040227 040228 040229 040230 040301 040302 040303 040304 040305 040306 040307 040308 040309 040316 040317 040318 040319 040425 040426 040427 040428 040706 040709 040712 040716 040717 040718 040719 040720 040721 040722 040723 040724 040725 040726 040727 040728 040729 040730 040731 040801 040802 040803 040804 040805 040806 040807 040808 040809 040810 040811 040812 040813 040814 040815 040816 040817 040818 040819 040820 040821 040822 040823 040824 199 9.2.2 EDAS Value These cases (143 cases) 020922 020923 .020924 020925 020927 020928 021004 021016 021026 021107 021116 021120 021124 021216 021228 030118 030225 030303 listed below 030328 030329 030330 030401 030402 030407 030408 030409 030410 030411 030412 030413 030414 030415 030416 030417 030418 030419 030420 030421 were used to measure 030422 030423 030424 030425 030426 030427 030428 030429 030220 030221 030222 030223 030224 030225 030226 030227 030228 030301 030302 030303 the E D A S wa\ 030304 030305 030306 030307 030308 030309 030311 030312 030313 030314 030315 ,030316 030317 030318 030319 030320 030321 030322 030323 030324 ue. 030325 030326 030327 030328 030329 030330. 030331 030401 030402 030403 030404 030405 030406 030407 030408 030409 030410 030411 030412 030413 9.3.1 Synoptic Hour Support Same case list as 9.2.1; E D A S failure (check above). These cases were used to compare REFOO vs R E F 12 results for summer and winter. 9.3.2 Summer Period Denial Same case list as 9.2.1; E D A S failure (check above). These cases were used to compare summer forecast results with winter forecast results. 9.3.3 WNA Quiet Period Periods W N A quiet periods were responsible for lowering the Study-Period winter forecast error. The Study-Period winters had many quiet periods. The quiet periods that were analyzed included the following cases. This is a subset of the case list for 9.2.1. Study Period (15 cases) 030202 030203 030204 030205 030206 030207 030208 030209 030210 030211 030212 030213 030214 030215 030216 9.4.1 Sources of Forecast Error Same Study-Period case list as 9.2.1; E D A S failure (check above). VERwest forecast were compared to VEReast forecasts. 200 9.4.2. Initialization & Model Attributes Same Study-Period case list as 9.2.2; E D A S value measurement. E T A forecasts were compared to M M 5 forecasts. 9.4.3 Effect of Boundary Conditions For the Study-Period cases listed below, the M M 5 lateral boundary-condition tendencies were updated at different frequencies. 031015 | 031118 | 04010 9.5.1 SCM-DA Sensitivity Tests/Calibration For the Study Period cases listed below, the maximum ROI was varied from 3 to 20 grid points. The S C M ROI sequences; Starting with 20: 20, 17, 14, 11, 8, 6, 4 , 3 , 2, 1 Starting with 15: 15, 13, 1 1 , 9 , 7 , 5 , 4 , 3 , 2 , 1 Starting with 10: 10, 9, 8, 7, 6, 5, 4, 3, 2, 1 Starting with 5: 5, 4 , 3 , 2, 1 Starting with 3: 3 ,2 , 1. 021016 | 021116 | 030123 | 030324 | 030408 | 030416 9.5.2 3DVAR-DA Sensitivity Tests/Calibration The cases listed below were used to discover the optimum calibration for 3 D V A R - D A . 031015 | 031016 9.5.3 SCM/3DVAR Comparison For the cases listed below, the results of S C M - D A and 3 D V A R - D A were compared. These cases are a subset of the case list for 9.8.1 etc. 021026 | 021116 | 021216 9.5.4 SCM Quality Control The effect of Q C - O N and Q C - O F F on forecast results was studied for each of the cases listed below. 021016 | 021107 | 021116 | 030308 9.6.3 NWP Error Linear Growth For each of the cases listed below, the B E N G A I N results are measured relative to R E F l 2 to REFOO R M S D improvement. 201 Study Period SCM-DA (42 cases) 020923 020927 020928 021004 021016 021026 021027 021029 021030 021103 021104 021105 021107 021114 021116 021120 021202 021203 021211 021216 021226 021228 030117 030118 030122 030123 030127 030128 030204 030211 030222 030223 030225 030302 030305 030308 030319 030324 030326 030401 030408 030416 Study Period 3DVAR-DA (29 cases) 031015 031016 031118 031128 031212 040101 040103 040105 040106 040107 040108 040109 040110 040111 040112 040113 040114 040115 040116 040117 040118 040119 040120 040121 040122 040127 040130 040202 040218 9.6.4 The Uneven Pacific Analysis This case was used to study the varying impact of the different buoy arrays on the forecast results. 020220 9.7.1 Pacific Data Void Fore the cases listed below, western North America and the Pacif ic areas were denied data. Test SCM-DA 021116 030127 Study 040103 Period (31 cases) 021120 030204 Period 040127 SCM-DA 020923 021202 030211 3DVAR- 040130 (5 cases) 020927 021203 030222 DA (11 040202 011213 021004 021211 030225 cases) 040218 020218 021016 021216 030305 031015 020316 021027 021226 030319 031016 020412 021104 021228 030324 031118 020506 021105 030117 030401 031128 Study 021107 030122 030408 031212 Period 021114 030123 030416 040101 9.7.2 Polar Data Void For the cases listed below, the arctic area was denied data. Study Period 3DVAR-DA (16 cases) 040106 040107 040108 040109 040110 040111 040112 040113 040114 040115 040116 040117 040118 040119 040120 040121 04012 202 9.8.1 Pacific RBS Locations Test Period S C M - D A (5 cases) 011213 020218 020316 020412 020506 Study Period 3 D V A R -D A (254 cases) 021116 021117 021118 021119 021120 021123 021124 021125 021126 021127 021129 021130 021201 021202 021203 021204 021205 021206 021207 021208 021209 021210 021211 021212 021213 021214 021215 021216 021217 021218 021219 021220 021221 021222 021223 021224 021225 021226 021227 021228 021229 021230 021231 030101 030102 030104 030105 030106 030107 030108 030109 030110 030111 030113 030114 030115 030116 030117 030118 030119 030120 030121 030122 030123 030124 030125 030126 030127 030128 030129 030130 030131 030201 030202 030203 030204 030206 030207 030208 030210 030211 030212 030213 030214 030215 030216 030217 030218 030220 030221 030222 030223 030224 030225 030226 030227 030228 030301 030302 030303 030304 030305 030306 030307 030308 030309 030311 030312 030313 030314 030315 030316 030317 030318 030319 030320 030321 030322 030323 030324 030325 030326 030327 030328 030329 030330 030331 030401 030402 030403 030404 030405 030406 030407 030408 030409 030410 030411 030412 030413 030414 030415 030416 030417 030418 030419 030420 030421 030422 030423 030424 030425 030426 030427 030428 030429 030430 030501 031004 031011 031014 031015 031016 031017 031026 031111 031112 031114 031115 031116 031117 031118 031120 031127 031128 031129 031130 031201 031202 031203 031204 031205 031212 040101 040103 040105 040106 040107 040108 040109 040110 040111 040112 040113 040114 040115 040116 040117 040118 040119 040120 040121 040122 040127 040128 040127 040128 040130 040131 040202 040203 040204 040205 040206 040207 040208 040209 040210 040211 040212 040213 040215 040216 040217 040218 040219 040220 040221 040222 040223 040224 040225 040226 040227 040228 040229 040230 040301 040302 040303 040304 040305 040306 040307 040308 040309 040316 040317 040318 040319 040425 040426 040427 040428 203 9.8.2 Pacific RBS Array Size These cases were used to study the effect of many combinations buoy locations in the Pacif ic, array size and profile heights. Test Study 021107 030211 Study 031118 Period Period 021114 • 030225 Period 031128 (2 cases) S C M - D A 021116 030319 3 D V A R - 040101 011213 (16 cases) 021120 030324 D A ( 9 040103 020412 021016 021216 030401 cases) 040127 021104 021228 030408 031015 040130 021105 030204 031016 040202 9.9.1 RBS Profile Height These cases were used to study the effect of many combinations Pacific locations, array size and profile heights. Test Study 021107 030211 Study 031015 Period Period 021114 030225 Period 040101 (1 cases) S C M - D A 021116 030319 3 D V A R - 040103 011213 (16 cases) 021120 030324 D A ( 3 021016 021216 030401 cases) 021104 ' 021228 030408 021105 030204 9.9.2. Parameters To determine the relative impact of the mass and wind observations, each case was run with the full suite of observations, and then alternatively denied the wind and mass observations. Study Period S C M - D A (2 cases 021120 030324 Study Period 3 D V A R -D A (6 cases) 031015 031016 040101 040103 040127 040130 9.9.3 Observation-error Sensitivity Study For the following case, observation error variances were halved and then doubled. 031015 9.10 RBS Targeting Using the cases listed below, the effect of targeting the most active meteorological areas was studied. 031015 I 031016 204 9.11 RBS Deployment Using the cases listed below, the effect of a strategic phased deployment of an expanding R B S on forecast results was studied. Test Period (5 cases) 011213 020218 020316 020412 020506 Study Period S C M - D A (3 cases) 021107 021120. 030324 Study Period 3DVAR-D A (5 cases 031015 031016 031017 040101 040103 9.12 Arctic Operations Using the cases listed below, the effect of an R B S equivalent on forecast results was studied. Study Period S C M - D A (7 cases) 020923 020927 021124 021126 030127 030324 030326 9.13.1 Extending the Network For the case listed below, the effect of adding radiosonde soundings over North America was studied. 011213 9.13.2 Balanced Operational Network For the case listed below, the effect of thinning the radiosonde network over North America studied. Test Study 021016 030408 Study 031128 Period Period 021120 030416 Period 040130 (1 case) S C M - D A 030123 3DVAR- 040202 020316 (8 cases) 030127 DA (3 030222 cases) 030324 9.14.2 Simulating the Mix of Observing Systems For the cases listed below, the effect of aircraft, surface, rocketsonde, aerosonde and dropsonde data on forecast results was studied 031015 031016 205 E N D OF D O C U M E N T Thank Y o u John Spagnol July, 2005 206 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0052550/manifest

Comment

Related Items