@prefix vivo: . @prefix edm: . @prefix ns0: . @prefix dcterms: . @prefix skos: . vivo:departmentOrSchool "Arts, Faculty of"@en, "Geography, Department of"@en ; edm:dataProvider "DSpace"@en ; ns0:degreeCampus "UBCV"@en ; dcterms:creator "Hunter, Cameron"@en ; dcterms:issued "2016-08-19T02:02:40"@en, "2016"@en ; vivo:relatedDegree "Master of Science - MSc"@en ; ns0:degreeGrantor "University of British Columbia"@en ; dcterms:description "The accuracy of gridded precipitation products in mountainous environments has been shown to be unreliable compared to other geographic areas due to the complex terrain and sparse network of stations typically found in these regions. This study is an analysis of the accuracy of the North American Regional Reanalysis (NARR) precipitation dataset for the province of British Columbia. Unlike similar gridded precipitation products, the NARR has not yet been evaluated to determine how reliably it reproduces observed patterns of precipitation in this region. The objective of this study is to assess the temporal and spatial patterns of precipitation in the NARR record in order to determine how closely it reproduces observed precipitation. A comparison of the NARR precipitation records with station precipitation records was conducted to evaluate the NARR’s ability to reproduce the interannual, monthly, and daily patterns of precipitation experienced. Streamflow records and NARR precipitation records for a number of basins throughout British Columbia were examined using a water balance approach to better understand the spatial variability of errors. A structural break in the NARR data was observed in 2003, which led to larger inaccuracies in the NARR record in subsequent years. This break was caused by a decision to exclude Canadian rain gauge data from the NARR’s data assimilation process from 2003 onwards. Several clear spatial patterns were observed in the NARR precipitation data. The NARR underpredicted precipitation in mountainous regions due to inaccuracies in its digital elevation model (DEM). The NARR overpredicted precipitation in the northern region of BC’s Interior Plateau due to a lack of available rain gauge data in this area. Finally, the NARR was found to be more accurate at modelling precipitation in areas with flat terrain and adequate station coverage, such as the southern region of the Interior and the Northeast Plateau. This analysis has shown the spatial and temporal variability of errors in the NARR dataset, allowing users to recognize both the strengths and potential shortcomings of this tool."@en ; edm:aggregatedCHO "https://circle.library.ubc.ca/rest/handle/2429/58887?expand=metadata"@en ; skos:note "An evaluation of the North American Regional Reanalysisprecipitation fields in a topographically complex domain,British Columbia, CanadabyCameron HunterB.Sc., University of Winnipeg, 2014a thesis submitted in partial fulfillmentof the requirements for the degree ofMaster of Scienceinthe faculty of graduate and postdoctoral studies(Geography)The University of British Columbia(Vancouver)August 2016c© Cameron Hunter, 2016AbstractThe accuracy of gridded precipitation products in mountainous environments has been shownto be unreliable compared to other geographic areas due to the complex terrain and sparsenetwork of stations typically found in these regions. This study is an analysis of the accuracyof the North American Regional Reanalysis (NARR) precipitation dataset for the province ofBritish Columbia. Unlike similar gridded precipitation products, the NARR has not yet beenevaluated to determine how reliably it reproduces observed patterns of precipitation in thisregion.The objective of this study is to assess the temporal and spatial patterns of precipitationin the NARR record in order to determine how closely it reproduces observed precipitation. Acomparison of the NARR precipitation records with station precipitation records was conductedto evaluate the NARR’s ability to reproduce the interannual, monthly, and daily patterns ofprecipitation experienced. Streamflow records and NARR precipitation records for a number ofbasins throughout British Columbia were examined using a water balance approach to betterunderstand the spatial variability of errors.A structural break in the NARR data was observed in 2003, which led to larger inaccuraciesin the NARR record in subsequent years. This break was caused by a decision to excludeCanadian rain gauge data from the NARR’s data assimilation process from 2003 onwards.Several clear spatial patterns were observed in the NARR precipitation data. The NARRunderpredicted precipitation in mountainous regions due to inaccuracies in its digital elevationmodel (DEM). The NARR overpredicted precipitation in the northern region of BC’s InteriorPlateau due to a lack of available rain gauge data in this area. Finally, the NARR was foundto be more accurate at modelling precipitation in areas with flat terrain and adequate stationcoverage, such as the southern region of the Interior and the Northeast Plateau.This analysis has shown the spatial and temporal variability of errors in the NARR dataset,allowing users to recognize both the strengths and potential shortcomings of this tool.iiPrefaceThis thesis is original work completed by the author. Guidance for this thesis was given by thesupervisory committee (Dan Moore, Ian McKendry, and Simon Donner).A version of this work has been published as a poster (Hunter, C., Moore, R.D., and McK-endry, I.G. Evaluating North American Regional Reanalysis using a spatially distributed waterbalance approach) on which the author acted as lead investigator, composing and present-ing the poster at the 2016 Canadian Geophysical Union (CGU)/Canadian Meteorological andOceanographic Society (CMOS) Joint Assembly in Fredericton, New Brunswick.iiiTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Motivation for the Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.2.1 Drivers of Spatial Variability in Precipitation . . . . . . . . . . . . . . . . 21.2.2 Overview of Canadian Precipitation Datasets . . . . . . . . . . . . . . . . 21.3 Research Objectives and Thesis Structure . . . . . . . . . . . . . . . . . . . . . . 52 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72.1 Study Area Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72.2 Data Sources and Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82.2.1 NARR Precipitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82.2.2 Station Precipitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92.2.3 Streamflow Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122.3 Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152.3.1 Statistical Comparison of NARR Precipitation and Station Precipitation . 152.3.2 Comparison of NARR Precipitation and Catchment Water Yield . . . . . 173 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193.1 Comparison of NARR Precipitation and Station Precipitation . . . . . . . . . . . 193.1.1 Analysis of Interannual Variability . . . . . . . . . . . . . . . . . . . . . . 193.1.2 Analysis of Precipitation Climatologies . . . . . . . . . . . . . . . . . . . . 27iv3.1.3 Analysis of Monthly Variability . . . . . . . . . . . . . . . . . . . . . . . . 303.1.4 Analysis of Daily Precipitation . . . . . . . . . . . . . . . . . . . . . . . . 363.2 Comparison of NARR Precipitation and Streamflow Measurements . . . . . . . . 394 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454.1 Temporal Variability of Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454.2 Spatial Variability of Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464.2.1 Coast Mountains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464.2.2 Interior Plateau . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474.2.3 Rocky Mountains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 484.2.4 Northeast Plateau . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 494.3 The Utility of the NARR Precipitation Estimates . . . . . . . . . . . . . . . . . . 495 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515.1 Summary of Key Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515.2 Future Research Direction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53vList of TablesTable 2.1 Station characteristics for the 18 Environment Canada Adjusted Precipitationstations used in this study, including the beginning and ending years of thestation record, the location, whether the station record is the result of joiningmore than one station, the elevation of the station calculated from a 1 kmDEM (m.a.s.l), and the elevation of the station calculated from the NARRDEM (m.a.s.l). All data except elevation data were obtained from Mekis andVincent (2011). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11Table 2.2 Characteristics of the Water Survey of Canada basins used in the study. . . . 13Table 3.1 Chow Test F statistics and p values. . . . . . . . . . . . . . . . . . . . . . . . . 21Table 3.2 NRMSE and NMBE values calculated from the total annual precipitation forthe 18 stations during the 1979-2002 and 2003-2015 periods. . . . . . . . . . . 22Table 3.3 Summary of regressions between station total annual precipitation and NARRtotal annual precipitation for the 1979-2002 and 2003-2015 periods. Notenough data was available to fit a regression for Prince Rupert following 2002. 26Table 3.4 NRMSE and NMBE values for the precipitation climatologies produced fromNARR monthly precipitation and station monthly precipitation for the 1979-2002 and 2003-2015 periods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27Table 3.5 NRMSE and NMBE values for the NARR monthly precipitation and stationmonthly precipitation for the 1979-2002 and 2003-2015 periods. . . . . . . . . 30Table 3.6 Adjusted R2 values from the linear model predicting station monthly precip-itation from NARR monthly precipitation and month and p-values from theANOVA comparing linear models 2.6 and 2.5 for the 1979-2002 and 2003-2015periods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32Table 3.7 Contingency analysis of NARR’s ability to predict wet vs dry days for the1979-2002 and 2003-2015 periods, showing when the NARR record correctlypredicted precipitation, incorrectly predicted precipitation (error of commis-sion) and failed to predict precipitation (error of omission). . . . . . . . . . . . 36viList of FiguresFigure 2.1 Elevation differences between the NARR DEM and a 1 km resolution DEM. 8Figure 2.2 Locations of the the 18 Environment Canada Adjusted Precipitation stationsand 46 Water Survey of Canada basins used in this study. Stations are shownas white triangles and basins as black polygons. A 1 km resolution DEM wasused for the background elevation. . . . . . . . . . . . . . . . . . . . . . . . . 10Figure 3.1 Total annual precipitation time series for the 18 climate stations (black lines)and the NARR grid cell in which the stations are located (grey lines) for 1979-2015. The red dashed line at 2002 shows the change in the data assimilatedinto the model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20Figure 3.2 Map of the NRMSE and NMBE values calculated from the total annualprecipitation for the 18 stations during the 1979-2002 (left panel) and 2003-2015 (right panel) periods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23Figure 3.3 Plot of station total annual precipitation vs NARR total annual precipitationfor the 1979-2002 period. Blue lines are the 1:1 lines and black lines are thelines of best fit through the scatter. . . . . . . . . . . . . . . . . . . . . . . . 24Figure 3.4 Plot of station total annual precipitation vs NARR total annual precipitationfor the 2003-2015 period. Blue lines are the 1:1 lines and black lines are thelines of best fit through the scatter. . . . . . . . . . . . . . . . . . . . . . . . 25Figure 3.5 Precipitation climatologies produced from the NARR monthly precipitation(grey line) and station monthly precipitation (black line) for the 1979-2002period. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28Figure 3.6 Precipitation climatologies produced from the NARR monthly precipitation(grey line) and station monthly precipitation (black line) for the 2003-2015period. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29Figure 3.7 Map of the NRMSE and NMBE values for the NARR monthly precipitationand station monthly precipitation for the 1979-2002 (left panel) and 2003-2015 (right panel) periods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31viiFigure 3.8 Plot of station total monthly precipitation vs NARR total monthly precipita-tion for the 1979-2002 period. Blue lines are the 1:1 lines and black lines arethe lines of best fit through the scatter calculated using the reduced model(Eq. 2.6). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33Figure 3.9 Plot of station total monthly precipitation vs NARR total monthly precipita-tion for the 2003-2015 period. Blue lines are the 1:1 lines and black lines arethe lines of best fit through the scatter calculated using the reduced model(Eq. 2.6). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34Figure 3.10 Plot of station total monthly precipitation vs NARR total monthly precip-itation for each month of the year for the Port Hardy station. Blue linesare the 1:1 lines and black lines are the lines of best fit through the scattercalculated using the full model (Eq. 2.5). . . . . . . . . . . . . . . . . . . . . 35Figure 3.11 Frequency plots of maximum daily rainfall for the NARR record (grey line)and the station record (black line). Lines fitted using method of momentsassuming the data follow a Gumbel distribution. The data were not splitinto two periods to provide a larger number of years to calculate the returnperiod of extreme events. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38Figure 3.12 Ratio of NARR to observed daily maximum rainfall by return period for the18 stations. Each dot represents one of the 18 stations. A value of < 1 impliesNARR maximum daily precipitation is lower than the station maximum dailyprecipitation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39Figure 3.13 Relation between mean annual catchment water yield and NARR mean an-nual catchment-average precipitation for the study basins during the 1979-2002 and 2003-2015 periods. The 1:1 line is shown in black with the greyshading showing ± 5%. Basins with points above the 1:1 line represent basinsin which the amount of water leaving the basin as streamflow is greater thanthe amount of NARR precipitation entering the basin, while basins below the1:1 line show a more realistic relationship where there is more water enteringthe basin than leaving as streamflow. . . . . . . . . . . . . . . . . . . . . . . . 40Figure 3.14 Map of the difference in the average NARR precipitation and average basinwater yield for the study basins during the 1979-2002 (left panel) and 2003-2015 (right panel) periods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41Figure 3.15 Time series of annual catchment water yield (black line) and total annualcatchment-averaged precipitation (grey line). The panel labels indicate theWSC station numbers (see Table 2.2 for catchment details). Part 1 of 3. . . . 42Figure 3.16 Time series of annual catchment water yield (black line) and total annualcatchment-averaged precipitation (grey line). The panel labels indicate theWSC station numbers (see Table 2.2 for catchment details). Part 2 of 3. . . . 43viiiFigure 3.17 Time series of annual catchment water yield (black line) and total annualcatchment-averaged precipitation (grey line). The panel labels indicate theWSC station numbers (see Table 2.2 for catchment details). Part 3 of 3. . . . 44ixAcknowledgmentsThere are many people I would like to thank for their help and support over the last two years.My time as a student at UBC was incredibly rewarding, and it was the support that I receivedfrom my advisors, my peers, and my family that allowed me to be successful in this experience.First, I would like to thank my supervisors Dan Moore and Ian McKendry for their constantfeedback and support throughout this process. Dan provided a wealth of knowledge, ideas, andattention to detail through the research process that helped the project towards its final product.Ian was invaluable in encouraging me to consider the project’s role in a broader research context.Both Ian and Dan have improved my research capabilities and I feel fortunate to have had thisopportunity to work with them. I would also like to thank Simon Donner for acting as myexternal reader, his thoughtful contributions to this thesis were much appreciated.Funding for this project was provided by the NSERC Canada Graduate Scholarships-Master’s (CGSM) Program.I would also like to acknowledge Joel Trubilowicz for providing me with several R scriptsthat helped me begin my analysis and Joey Lee for his support in formatting this manuscript.I would also like to thank Deirdre Loughnan for her support and friendship.Most importantly, I would like to thank Andrew and my family for their support andencouragement to move to Vancouver and work hard to be successful. Andrews countless hoursspent listening to my ideas and editing my thesis were invaluable, and I am so thankful for hissupport.xChapter 1Introduction1.1 Motivation for the StudyWeather and climate data are an integral component of many hydrologic analyses and modelling.However, collecting reliable weather and climate data continues to be a challenge. Observedweather data have an uneven distribution, and are particularly scarce at higher elevations inmountainous regions (Moore et al., 2013). In recent years, a growing number of gridded climateproducts based on statistical interpolation have become available with a range of temporaland spatial resolutions, including the Australian National University Spline (ANUSPLIN), TheCanadian Precipitation Analysis (CaPA), and the Precipitation-elevation Regression on Inde-pendent Slopes Model (PRISM) (Hutchinson et al., 2009; Mahfouf et al., 2007; Daly et al.,2008).The use of these gridded products for analysis and modelling is based on the assumptionthat the higher resolution of a gridded product represents an increase in information qualityrelative to observational data. In reality, the quality and quantity of input data and the inter-polation method used to calculate the gridded estimates can significantly impact the accuracyof the resultant data (Daly, 2006). Re-analysis products offer another means for producing datain areas with inadequate station coverage by providing atmospheric variables at multiple eleva-tions and sub-daily resolution. These products use observational data to constrain numericalweather models, and thus use physical laws and their numerical approximations to constrainthe interpolation.It is important for users of any dataset to understand how well their data represent thecomplex spatial and temporal patterns of climate by understanding the distribution of errorswithin their area of interest. Independent assessment of datasets can better inform the scientificcommunity on the strengths and appropriate uses of each dataset (Mahfouf et al., 2007; Daly,2006; Ruane, 2010; Berne and Krajewski, 2013).The North American Regional Reanalysis seeks to provide users with long-term, consis-tent, high-resolution climate data. The integration of a precipitation assimilation system intothe model is believed to allow for more accurate precipitation estimates for North America1(Mesinger et al., 2006).The accuracy of NARR precipitation data for British Columbia has not yet been analyzedand, therefore, the research below seeks to provide insight on the spatial and temporal distri-bution of errors for the NARR precipitation dataset in order to better inform future users onthe strengths and weaknesses of this tool.1.2 BackgroundThis section provides an overview of the major factors that influence precipitation patterns,as well as a discussion of several comparable precipitation datasets in order to provide contextwith regard to how this project fits within the broader research area.1.2.1 Drivers of Spatial Variability in PrecipitationSpatial patterns of observed precipitation are the result of many factors acting at the synopticand local scales. At the synoptic scale, the prevailing wind direction and position of the stormtrack direct movement of cyclonic precipitation over the region (Harman, 1991). More locally,topography can act as a barrier, forcing uplift and adiabatic cooling, which produces a generalincrease in precipitation with increasing elevation (Oke, 1987). This relationship between eleva-tion and precipitation shows local variation because terrain induced climate transitions, like therain shadow, produce sharp changes in precipitation between the windward side of the slope,where the largest amount of precipitation is commonly observed, and the leeward slope, whereprecipitation is at a minimum (Daly, 2006). Finally, large water bodies, which act as sources ofatmospheric moisture, cause an increase in the amount of precipitation to be recorded 50-100km from their coastlines (Daly, 2006).These climate-forcing factors interact to produce the complex spatial patterns of precipita-tion observed in mountainous regions such as western Canada. The high degree of variation inthese patterns makes accurately recording precipitation a challenge, in part because the volumeand density of climate stations in this area are insufficient to properly capture the complexityof the patterns (Daly, 2006). The poor representation of spatial patterns of precipitation is dueto the fact that stations are unevenly distributed across the region, and are generally located indensely populated, low elevation areas such as major urban centers (Ensor and Robeson, 2008).This bias makes areas of high elevation and sparse population, such as the northern portion ofBritish Columbia, largely unrepresented in the available data (Daly et al., 1994).1.2.2 Overview of Canadian Precipitation DatasetsGridded precipitation products provide a means of representing spatial patterns of precipitation.However, the methods used to calculate these estimates can add new errors and uncertaintiesto the data, which can ultimately influence their reliability. The success of these datasets isdependent on many assumptions, two of the most important being: 1) that station data must be2representative of the spatial patterns of precipitation, and 2) that the method used to calculatethe gridded estimates accounts for a complex relationship between terrain and precipitation.In the mountainous region of western Canada, these assumptions may not be met. Theterrain is extremely heterogeneous and station spacing is typically greater than 100 km, whilethe physiographic features that govern precipitation formation – elevation, coastal proximity,and terrain induced climate transitions – operate at much smaller spatial scales (10-50 km).Therefore, gridded precipitation products need to have their ability to reproduce spatial andtemporal patterns of precipitation assessed in order to determine their accuracy. A review ofseveral datasets that provide precipitation data for British Columbia is provided below.The Australian National University Spline (ANUSPLIN) dataset is produced using trivariatethin-plate smoothing splines to interpolate climate in three dimensions. This method uses adigital elevation model (DEM) to interpolate the climate grid from station data based onlongitude, latitude, and elevation (Hopkinson et al., 2011; Hutchinson, 1995; Hutchinson et al.,2009). This method has been shown to account for the changing relationship in elevation andprecipitation in space, which makes it a useful method for interpolating climate patterns overlarge regions. However, this dataset has difficulties reproducing precipitation patterns in areaswhere the station network does not provide sufficient coverage (Hutchinson et al., 2009). Asa result, ANUSPLIN precipitation data should be used with care in the mountainous regionsof British Columbia, as we know that station data do not properly represent the patterns ofprecipitation in this region.The Canadian Precipitation Analysis (CaPA) applies a statistical interpolation techniquethat combines and evaluates precipitation gauge data, short-range weather forecasts, and radarestimates to produce a more accurate spatial estimation of precipitation (Mahfouf et al., 2007).The CaPA uses short-range weather forecasts from the Global Environmental Multiscale modelas a background field, and compares that field with the observed precipitation gauge and radardata to produce an error analysis. The error analysis is then used to weight the observationaldata and interpolate it to a grid based on weighting factors (Mahfouf et al., 2007). The devel-opers of this dataset have warned that caution should be exercised when using this method tointerpret both winter time precipitation and precipitation in mountainous environments, as thereliability of these estimates should be considered questionable (Mahfouf et al., 2007).The Precipitation-elevation Regression on Independent Slopes Model (PRISM) data areproduced by interpolating climate data using station data, a DEM, and a knowledge-basedstatistical interpolation method (Daly et al., 2008). PRISM interpolates climate grids using alinear relationship between climate and elevation. This relationship is used to calculate a localregression function for each grid cell. The station data used to develop this local relationshipare weighted based on their elevation, topographic facet (orientation), coastal proximity, andtheir location in a two-layered atmosphere. Stations that are the closest and most similar tothe grid cell will be given the highest weight (Daly et al., 2008). After the local regressionfunction has calculated estimates for each grid cell, a quality control assessment is conducted3to ensure that the value of each cell falls within set boundaries, and that the estimated valuesare spatially consistent. This dataset has been shown to accurately reproduce precipitationpatterns in mountainous environments.It has been demonstrated that all three of these statistical interpolation methods have theirown strengths with respect to reproducing the variation of precipitation observed. However, allthese methods rely on station data and thus are not able to accurately reproduce the spatialvariation of precipitation in areas where there are few climate stations (Hutchinson et al., 2009;Mahfouf et al., 2007; Wang et al., 2012).One alternative to using gridded data that have been interpolated from precipitation gaugesis to estimate precipitation using weather radar. Weather radar has the ability to producedatasets with much higher spatial and temporal resolution because they can sample/sensethe area at a much higher frequency than traditional precipitation gauges. While radar doesnot directly measure precipitation rates, it can indirectly assess them by measuring the radarbackscatter from a precipitation event (Berne and Krajewski, 2013). While radar estimatesare currently used for the detection and monitoring of extreme weather events, there are stillsome uncertainties in estimating radar rainfall rates that must be considered. Radar signalscommonly interact with objects other than hydrometeors, causing inaccurate rainfall rates tobe estimated. These topography-beam interactions often cause ground echoes and beam atten-uation, which results in inaccurate precipitation amounts being recorded (Berne and Krajewski,2013).Radar remote sensing of precipitation in mountainous environments is uniquely challenging.Complex topography acts as a barrier, blocking the signal from sensing precipitation eventsoccurring behind it, which severely limits the range (Mo, 2008). Currently, the CanadianRadar Network is limited to the southern portion of the country. This limitation, combinedwith the difficulties associated with radar in mountainous environments, makes radar remotesensing of precipitation in the mountainous region of western Canada not viable.Another alternative to interpolated precipitation gauge data is to use reanalysis datasets.Reanalysis systems assimilate observational data to model the state of the atmosphere andproduce climate grids for a number of climatic fields, including precipitation, based on themodeled atmosphere. These datasets can offer continuous, long-term data as the reanalysissystems are kept unchanged for the entire period.The NCEP/NCAR Reanalysis system (R-1) and the NCEP-DOE AMIP-II Reanalysis (R-2) are examples of global reanalysis systems (Kalnay et al., 1996; Kanamitsu et al., 2002; Kistleret al., 2001). The precipitation fields from both R-1 and R-2 reanalysis systems have severalweaknesses. The first is that precipitation is one of the variables which is completely determinedby the model, meaning that the precipitation calculated is not constrained by observationaldata, but instead determined by the assimilation of other variables. Several authors warn thatthese precipitation datasets do not always match with observational data, and should be viewedas best estimates of the model (Bukovsky and Karoly, 2007; Kistler et al., 2001). Additionally,4the coarse resolution of the global reanalysis is not useful for regional studies, as it does notprovide adequate spatial variation in areas of complex terrain (Solaiman and Simonovic, 2010).The North American Regional Reanalysis (NARR) is one of many reanalysis products usedto provide data at a finer scale than global reanalyses. This particular product provides a long-term reanalysis for hydrology that addresses some of the shortcomings of the global reanalysisR-1 and R-2 (Mesinger et al., 2006). The NARR system provides data with a smaller grid cellsize (∼ 32 km compared to ∼ 210 km in the R-1 and R-2 systems) for the North Americandomain, providing a better estimation of spatial patterns in complex terrain. The NARR systemuses much of the same data as the R-2 reanalysis (temperature, wind, moisture, geopotentialheight, etc.), but has added precipitation data into the atmospheric analysis. The precipitationassimilation process uses observational precipitation data to constrain the land-atmospheremodel, forcing the modeled hydrologic cycle to be closer to the observed than if the modelwere left to forecast precipitation (Mesinger et al., 2006). This development has improvedprecipitation estimates from the R-1 and R-2 estimates, making the NARR precipitation valuescloser to those observed (Mesinger et al., 2006).The accuracy of NARR precipitation data has been extensively tested for the continentalUnited States (Mesinger et al., 2006; Becker et al., 2009; Bukovsky and Karoly, 2007; Ruane,2010). It has been shown that with the dense network of precipitation stations located in theUnited States, NARR is able to reproduce the patterns of precipitation experienced with severalexceptions, one of which is the reduction of extreme events and a bias towards more frequentlight precipitation events (Becker et al., 2009). In Canada, accuracy of NARR precipitationfields has not been as thoroughly tested (Eum et al., 2014; Solaiman and Simonovic, 2010).In particular, the mountainous region of British Columbia has gone largely unevaluated. Eumet al. (2014) have shown that NARR has difficulty modelling the orographic effect in highelevation areas of the Athabasca River Basin. Trubilowicz et al. (2016) provided an evaluationof several other NARR gridded fields (air temperature, vapour pressure, wind speed, and shortand long-wave radiations) showing that NARR can model air temperature and vapour pressurewith a high degree of accuracy in high elevation locations of BC. However, there are someuncertainties associated with wind speed, short-wave and long-wave radiation fields.Other datasets that rely on station data have been proven to be unable to reproduce spatialand temporal patterns of precipitation due to the complex circumstances (topographic andstation distribution) of this region (Hutchinson et al., 2009; Mahfouf et al., 2007). NARR shouldalso be evaluated to determine if its precipitation estimates experience the same inaccuraciesfacing other methods of measurement.1.3 Research Objectives and Thesis StructureSection 1.2 has shown that an independent assessment of the North American Regional Reanal-ysis precipitation data could be beneficial to uncovering the strengths and weaknesses of thisdataset for the province of British Columbia. The objective of this study is to complete such5an assessment in the hopes of better understanding:1. The temporal variability of errors in the NARR precipitation data through a comparisonof the seasonal and interannual variation of observed precipitation at selected weatherstations and NARR precipitation estimates at the nearest grid point.2. The spatial variability of errors in NARR precipitation data through a basin scale analysisof NARR precipitation estimates and streamflow observations to determine if the NARRcan reproduce the spatial variability of precipitation experienced in the province of BritishColumbia.Chapter 2 provides a description of the study area, data, and methodology used in thisproject. Chapter 3 highlights the results, while Chapter 4 provides a discussion of the temporaland spatial variability of error in the NARR dataset for British Columbia. Finally, Chapter 5provides conclusions and recommendations for future research.6Chapter 2Methods2.1 Study Area DescriptionThe study area covers the province of British Columbia, where major geographic features includethe Coast Mountains, the Interior Plateau, and the Rocky Mountains. The Coast Mountainsexperience the largest volume of precipitation, as the prevailing winds move storm systems overthe west coast of BC for most of the winter months. As a result, the terrain in this regioncreates an orographic effect that causes much of the Coast Mountain’s precipitation (Mooreet al., 2010). Summer weather is typically warmer and drier for this region. Generally speaking,the Interior of BC experiences less precipitation overall, as it exists in the rain shadow of theCoast Mountains. Greater seasonality is experienced here with cold dry winters and the summermonths bringing an increase in precipitation and temperature with most of the precipitationduring the summer months being a product of convective storms (Moore et al., 2010). Finally,a secondary maximum of precipitation is experienced along the eastern portion of the province,due to orographic precipitation forced by the Columbia and Rocky Mountains (Moore et al.,2010).This study area was chosen because it is a geographically diverse region with both largechanges in topography and variations in the spatial patterns of precipitation. British Columbiais also an interesting area of study because a large portion of its northern region has verysparse station data coverage, creating a degree of uncertainty with regard to the accuracy ofprecipitation estimates. NARR precipitation estimates for this area have not been evaluated.It is also important to note that because of the prevalence of stations at lower elevations, abias exists in that higher elevations may not be properly represented in this region (Daly et al.,2002).72.2 Data Sources and Processing2.2.1 NARR PrecipitationThe NARR project seeks to provide a long-term, consistent, high-resolution climate dataset forthe North American domain, (Mesinger et al., 2006). It does so through a reanalysis systemwhich provides climate grids at ∼32 km (0.3 degrees) resolution at 3 hour, daily, and monthlytime periods for North America.One result of the ∼32 km resolution of the NARR model is that its topography showsnoteworthy differences from reality. Figure 2.1 shows the differences between the NARR DigitalElevation Model (DEM) and a 1 km resolution DEM.Figure 2.1: Elevation differences between the NARR DEM and a 1 km resolution DEM.One improvement of the NARR over global reanalysis is its ability to assimilate observationalprecipitation data into the reanalysis system, better constraining the water cycle. The successof this system has led the developers of the NARR to suggest that it can now be used to model8shorter, more extreme events (such as flash floods) more reliably (Mesinger et al., 2006).In December of 2002, the NARR began processing data in real time, and as a result changedthe data used in its precipitation assimilation process. Prior to 2003, daily rain gauge datafor the United States, Mexico, and Canada were disaggregated to hourly precipitation esti-mates. Gridded estimates from the Climate Prediction Center Merged Analysis of Precipitation(CMAP) precipitation analysis were used over the oceans where no rain gauge data were avail-able. Following this change, fewer rain gauges were available in real time for the continentalU.S. and, as a result, WSR-88D radar precipitation estimates supplemented the sparse net-work of available rain gauge data. Similarly, Climate Prediction Center MORPHing technique(CMORPH) precipitation estimates were used in Mexico and over the oceans where CMAPdata were no longer available. Finally, rain gauge data for Canada were no longer availablein real time and thus all Canadian precipitation inputs have been excluded (Mesinger et al.,2006).2.2.2 Station PrecipitationPrecipitation data for a number of stations across BC were obtained from the EnvironmentCanada Adjusted Precipitation Dataset, which aims to provide long-term, statistically stable,observational data. Stations have been adjusted to account for rain gauge errors such as windovercatch and evaporation loss, and station observations from nearby stations have been joinedto produce more statistically stable, long-term observations (Mekis and Vincent, 2011).A total of 18 climate stations were chosen based on spatial coverage, length of time series,and completeness of time series. The chosen stations provided coverage for as much of BC, andfor as many years (minimum 10 years of data) of the NARR period (1979-2015), as possible.Figure 2.2 and Table 2.1 provide the locations and characteristics of these 18 stations.9Figure 2.2: Locations of the the 18 Environment Canada Adjusted Precipitation stationsand 46 Water Survey of Canada basins used in this study. Stations are shown aswhite triangles and basins as black polygons. A 1 km resolution DEM was used forthe background elevation.10Table 2.1: Station characteristics for the 18 Environment Canada Adjusted Precipitation stations used in this study, including thebeginning and ending years of the station record, the location, whether the station record is the result of joining more thanone station, the elevation of the station calculated from a 1 km DEM (m.a.s.l), and the elevation of the station calculatedfrom the NARR DEM (m.a.s.l). All data except elevation data were obtained from Mekis and Vincent (2011).Station Name Station ID BeginningYearEnd Year Latitude Longitude StationJoinedStationEleva-tion(DEM)StationEleva-tion(NARR)Comox 1021830 1936 2014 49.7 -124.9 Yes 26 315Dryad Point 1062544 1933 2014 52.2 -128.1 Yes 4 172Estevan Point 1032730 1924 2014 49.4 -126.6 No 7 36Fort Nelson 1192940 1938 2012 58.8 -122.6 No 382 379Fort St James 1092970 1895 2014 54.5 -124.3 No 686 821Fort St John 1183000 1931 2012 56.2 -120.7 Yes 695 708Germansen Landing 1183090 1952 2013 55.8 -124.7 No 766 1186Joe Rich Creek 1123750 1929 2008 49.9 -119.1 No 875 1340Langara 1054500 1937 2014 54.3 -133.1 No 41 0Port Alice 1036240 1924 2014 50.4 -127.5 No 21 245Port Hardy 1026270 1944 2013 50.7 -127.4 No 22 156Prince George 1096450 1913 2009 53.9 -122.7 Yes 691 821Prince Rupert 1066481 1909 2006 54.3 -130.4 Yes 35 43Saanichton 1016940 1914 2014 48.6 -123.4 No 61 106Sandspit 1057050 1949 2014 53.3 -131.8 No 6 105Shawnigan Lake 1017230 1911 2014 48.6 -123.6 No 138 107Vancouver 1108447 1896 2013 49.2 -123.2 Yes 4 58Williams Lake 1098940 1936 2012 52.2 -122.1 Yes 940 940112.2.3 Streamflow DataStreamflow data for the study area were obtained from the Water Survey of Canada (WSC) inorder to compare measured streamflow with basin precipitation estimates using a water balanceapproach. A number of basins throughout British Columbia were selected in order to compareNARR precipitation data with measured streamflow from the WSC. The basin selection processincluded three criteria:1. Basins must have a minimum 10 years of continuous WSC streamflow data within theNARR period (1979-2015) to provide a long enough period in which to compare measuredstreamflow and precipitation amount.2. Basins must have natural, unregulated flow patterns so that water is not being held backby a dam or diverted out of the catchment.3. Basins should have less than 3% of their surface area covered by glaciers.These criteria were chosen to provide a long enough period from which to compare precipita-tion and streamflow, and also to mitigate instances where precipitation is held in the basin andnot released within the same water year, which would ultimately create a discrepancy betweenmodelled precipitation and streamflow. As dams and diversions hold water in the basins, anybasins with regulated flow were excluded from the study. Similarly, as glacial melt providesa source of streamflow that would create these same discrepancies, basins with greater than3% of their area covered with glacier snow and ice were excluded, as 2-3% glacier coveragecan significantly influence streamflow measurements during summer months (Stahl and Moore,2006). A total of 46 basins across the province were selected based on these criteria. Figure2.2 shows the distribution of basins throughout the province, while Table 2.2 summarizes eachbasin’s key characteristics.12Table 2.2: Characteristics of the Water Survey of Canada basins used in the study.Basin ID Gauging Station Location Latitude Longitude BasinArea(km2)MeanBasinEleva-tion(DEM)MeanBasinEleva-tion(NARR)GlacierCover(%area)07EA004 Ingenika River above Swannell River 56.73 -125.10 4200 1628 1581 0.0107EB002 Ospika River above Aley Creek 56.53 -123.93 2220 1613 1463 0.0207EC002 Omineca River above Osilinka River 55.92 -124.57 5490 1421 1255 0.0507EC003 Mesilinka River above Gopherhole Creek 56.24 -124.64 2980 1537 1393 0.1007EC004 Osilinka River near End Lake 56.13 -124.80 1960 1519 1282 0.1007EE007 Parsnip River above Misinchinka River 55.08 -122.91 4900 1266 1206 0.4207FA006 Halfway River near Farrell Creek 56.25 -121.63 9350 1265 1225 0.0207FB002 Murray River near the Mouth 55.56 -121.20 5620 1346 1373 0.4007FB006 Murray River above Wolverine River 55.07 -121.02 2410 1456 1500 0.9308EB004 Kispiox River near Hazelton 55.43 -127.71 1870 981 982 0.6808EC013 Babine River at outlet of Kilkitwa Lake 55.43 -126.70 6790 1125 929 0.0208EE008 Goathorn Creek near Telkwa 54.65 -127.12 132 1365 1167 0.4008FB006 Atnarko River near the Mouth 52.36 -126.00 2430 1557 1415 0.7008FF002 Hirsch Creek near the Mouth 54.06 -128.60 347 1067 821 1.4508FF003 Little Wedeene River below Bowbyes Creek 54.14 -128.69 182 881 831 0.9708GD008 Homathko River at inlet to Tatlayoko Lake 51.67 -124.41 500 1576 1190 0.4708HC002 Ucona River at the Mouth 49.71 -126.10 185 991 661 0.7508HD011 Oyster River below Woodhus Creek 49.89 -125.24 298 1027 632 0.5108HE006 Zeballos River near Zeballos 50.01 -126.84 181 836 359 0.1508HF005 Nimpkish River above Woss River 50.21 -126.61 787 829 812 0.101308JB002 Stellako River at Glenannan 54.01 -125.01 3600 1088 901 0.1908JB003 Nautley River near Fort Fraser 54.09 -124.60 6030 1070 890 0.1208JD006 Driftwood River above Kastberg Creek 55.97 -126.68 407 1295 1148 0.3108KB006 Muller Creek Near the Mouth 54.30 -120.98 134 1507 1520 0.9908KD007 Bowron Rover below Box Canyon 54.02 -122.13 3420 1331 1206 0.1708KH001 Quesnel River at Likely 52.62 -121.57 5930 1419 1211 1.8008KH010 Horsefly River above McKinley Creek 52.29 -121.06 785 1627 1257 0.0908LB038 Blue River near Blue River 52.12 -119.30 280 1633 1613 1.4208LD001 Adams Rover near Squilax 50.94 -119.65 3080 1377 1331 1.7908LE024 Eagle River near Malakwa 50.94 -118.80 904 1494 1403 1.3208LE027 Seymour River near Seymour Arm 51.26 -118.95 805 1507 1384 2.5208MB006 Big Creek above Groundhog Creek 51.52 -123.11 1020 1961 1867 1.7408ME025 Yalakom River above Ore Creek 50.91 -122.24 575 1922 1522 0.1508MF065 Nahatlatch River below Tachewana Creek 49.95 -121.86 715 1607 1552 2.5408MG001 Chehalis River near Harrison Mills 49.30 -121.94 383 952 840 0.2008MH141 Coquitlam River above Coquitlam Lake 49.49 -122.79 54 154 510 0.6208NE006 Kuskanax Creek Near Nakusp 50.28 -117.75 337 1761 1378 0.0808NF001 Kootenay River at Kootenay Crossing 50.89 -116.04 420 1862 1851 0.5808NH005 Kaslo River below Kemp Creek 49.91 -116.95 453 1872 1551 1.5608NH132 Keen Creek below Kyawats Creek 49.87 -117.12 92.2 2080 1677 2.1908NJ013 Slocan River near Cresent Valley 49.46 -117.56 3320 1672 1479 1.0208NK002 Elk River at Fernie 49.51 -115.07 3110 1998 1863 0.7508NK016 Elk River near Natal 49.87 -114.87 1870 2090 2003 0.6210BE004 Toad River above Nonda Creek 58.86 -125.38 2570 1715 1582 2.7610BE007 Trout River at Kilometre 783.7 Alaska Highway 59.34 -125.94 1190 1502 1454 0.2710CB001 Sikanni Chief River near Fort Nelson 57.24 -122.69 2160 1566 1526 0.1814Daily streamflow measurements from WSC gauging stations were aggregated for each wateryear (October-September). Basin water yield was calculated as:WY = cQmaA(2.1)where WY is the annual water yield (mm), Qma the mean annual streamflow (m3s−1), A is thecatchment area (km2), and c is a coefficient of unit conversions (3.154 x 104).2.3 Data Analysis2.3.1 Statistical Comparison of NARR Precipitation and StationPrecipitationAnalysis of Interannual VariabilityTotal annual precipitation is defined as the amount of precipitation accumulated at each gridcell over the course of each water year (October - September) during the years 1979-2015.Comparing the total annual precipitation from the NARR with the station record across anumber of years evaluates the ability of NARR to reproduce the pattern of wet and dry yearsrecorded.Time series of total annual precipitation were first tested to determine if any structuralbreaks in the dataset existed due to a change in the model’s precipitation assimilation processin December 2002, when Canadian precipitation gauge data were excluded from the process.The Chow Test was used to detect structural breaks in the total annual precipitation timeseries between 2002-2003 (Eum et al., 2014). This test fits linear models to the two periods todetermine if the coefficients of linear regression are the same for each (Chow, 1960).Normalized root mean squared error (NRMSE) and normalized mean bias error (NMBE)were used to quantify the degree of agreement between the NARR grid cell record and stationrecords total annual precipitation for each of the 18 station as follows:NRMSE =√Σ(Ps−PNARR)2nPs(2.2)andNMBE =Σ(Ps − PNARR)Ps(2.3)where Ps is the total annual precipitation calculated from the station record (mm/yr) andPNARR is the total annual precipitation calculated from the NARR precipitation field (mm/yr)A simple linear regression between the NARR total annual precipitation and the stationrecord total annual precipitation was developed to determine whether the NARR can be bias15corrected to predict total annual precipitation observed in the station record:Yt = b0 + b1Xt + et (2.4)where Xt is total annual precipitation calculated from the NARR precipitation field (mm/yr)for year t, Yt is the total annual precipitation calculated from the station record (mm/yr) foryear t, b0 and b1 are an intercept and slope, respectively, to be estimated, and et is an error term.The strength of the relationship between the NARR measurement and the station measurementwas determined by the coefficient of determination (R2).Analysis of Precipitation ClimatologiesPrecipitation climatologies are the average amount of precipitation recorded during each monthfor the 1979-2015 period. This measure allows for the evaluation of NARR’s ability to reproduceprecipitation climatologies for each of the individual stations. Normalized RMSE and MBE wereagain used to quantify the relationship between station precipitation climatologies and NARRprecipitation climatologies from the grid cell the station was located in.Analysis of Monthly VariabilityTotal monthly precipitation is defined as the accumulated precipitation for each month duringthe period 1979-2015. This measurement was used to determine whether NARR is able toreliably reproduce the monthly pattern of precipitation at each individual station. NormalizedRMSE and MBE were once more used to quantify the relationship between station total monthlyprecipitation and NARR total monthly precipitation.To assess the potential for bias-correcting the NARR values, a regression model was fitbetween station total monthly precipitation and NARR total monthly precipitation, with monthadded into the model as a factor variable. The regression model can be expressed as follows:Yt = b0 + b1Xt + b2,mdm,t + b3dm,tXt + et (2.5)where Yt is station total monthly precipitation (mm/month), Xt is NARR total monthly pre-cipitation (mm/month), dm,t represents a set of 11 dummy variables (m ranges from 1 to 11)and et is an error term. The dummy variables are coded such that, for example, dm,t for m =1 is set equal to 1 for the month of February and 0 otherwise; dm,t for m = 2 is set equal to 1for the month of March and 0 otherwise; and so on. This model allows the slope and interceptof the regression to vary by month. A second reduced model was also fit:Yt = b0 + b1Xt (2.6)A one-way Analysis of Variance (ANOVA) was used to compare the models represented by16Eq. 2.5 and 2.6. A significant difference indicated that the full model 2.5 was the better fit;i.e., that the intercepts and slopes do vary by month.Analysis of Daily PrecipitationThe magnitude and timing of daily precipitation events from the NARR record were comparedwith the station record to assess NARR’s ability to reproduce daily precipitation patterns. Inorder to compare the two time series, all days that were missing data were excluded from bothdatasets. Daily precipitation values below 0.1 mm/day were converted to zero in both of thetime series to avoid complications with inconsistent definitions of trace values.Contingency tables were used to determine whether the timing of precipitation from theNARR data matched with the timing of precipitation recorded in the station record. Thenumber of days on which precipitation was recorded in both datasets, and the number of dayswhen no precipitation was recorded at both datasets, were compared with both the number ofdays NARR predicted precipitation when none occurred (error of commission) and the numberof days NARR failed to predict precipitation when it occurred (error of omission).Maximum daily precipitation amount from each year in the study period was extracted fromboth station and NARR records. Frequency curves were created using a Gumbel distribution,where plotting position was calculated by:T =n+ 1r(2.7)where T is the return period, n is the number of years, and r is the rank. A frequency curvewas fitted to the data using the method of moments, assuming the data follow a Gumbel (orExtreme Value I) distribution (Bedient and Huber, 1992).2.3.2 Comparison of NARR Precipitation and Catchment Water YieldTo gain a better understanding of the spatial variability of errors in the NARR, NARR precip-itation estimates were compared with streamflow measurements using a spatially distributedwater balance approach. The difference between the NARR water year precipitation and basinstreamflow was studied for a number of basins in the province to determine whether the rela-tionship between modelled precipitation and measured streamflow was realistic.The NARR grid (∼32 km) was re-sampled to a 1 km grid using a bilinear interpolationmethod, as many mountainous basins were located in only a single NARR grid cell. Thisre-sampling was conducted to provide a larger number of grid cells from which to calculatespatially averaged precipitation for these smaller basins.NARR precipitation estimates were integrated across each basin for every water year (October-September). The catchment water balance can be expressed as:P = WY + E + ∆S (2.8)17where P is the precipitation entering the basin (mm/yr), WY is the streamflow leaving thebasin expressed as basin water yield (mm/yr), E is the evapotranspiration loss (mm/yr), and∆S is the change in water storage in the basin (mm/yr).For catchments without significant glacier coverage or regulation, ∆S should approach 0when averaged over multiple years. The evapotranspiration term can be assumed to be positive.Realistic NARR precipitation estimates will result in positive differences between precipitationand basin water yield, which allows for evapotranspiration loss; a value of NARR P less thanWY is physically unrealistic.The average difference between precipitation and basin water yield for the study periodwas calculated for each basin, and the spatial pattern was mapped. Then, the NARR totalannual precipitation for each basin and water year was compared with the water yield fromeach basin to determine if the pattern of wet and dry years present in the streamflow recordwas reproduced by the NARR precipitation record.18Chapter 3Results3.1 Comparison of NARR Precipitation and StationPrecipitation3.1.1 Analysis of Interannual VariabilityThe time series of total annual precipitation (1979-2015) from the station record and closestNARR grid cell for each of the 18 stations are shown in Figure 3.1. A summary of the ChowTest results can be found in Table 3.1. The Chow Test found a significant structural break (α= 0.5) in the NARR total annual precipitation for all of the stations that were studied. As aresult of this test, the following analysis has been split into two periods: 1) the period in whichCanadian precipitation gauge data were assimilated into the NARR model (1979-2002), and 2)the period in which precipitation gauge data were excluded from the model (2003-2015).The normalized RMSE and MBE were calculated for each of the 18 stations (Table 3.2).In comparing the two periods of this study, a large increase in the NRMSE can be observedat many of the stations during the latter period (2003-2015). This increase indicates a greaterdisparity between the observed station record and the modelled NARR precipitation during the2003-2015 period, while the 1979-2002 period showed a higher degree of agreement for many ofthese same stations. Several stations (Shawnigan Lake, Vancouver) did exhibit a decrease inNRMSE during the 2003-2015 period.The difference between the two periods can also be observed with the NMBE. The 1979-2002 period has relatively small NMBE values; however, several stations (Langara, Port Alice,Shawnigan Lake) did report large positive NMBE values. These values indicate that the NARRrecord consistently underestimated the amount of total annual precipitation observed in thestation record at these locations. Conversely, during the 2003-2015 period, many stationsrecorded large negative NMBE values, indicating that the NARR model was overestimatingthe total annual precipitation during this period.Figure 3.2 presents a map of normalized RMSE and MBE for the two study periods. During19Figure 3.1: Total annual precipitation time series for the 18 climate stations (black lines)and the NARR grid cell in which the stations are located (grey lines) for 1979-2015. The red dashed line at 2002 shows the change in the data assimilated intothe model.20Table 3.1: Chow Test F statistics and p values.Station Name Station ID F Value p ValueComox 1021830 28.37 < 0.01Dryad Point 1062544 8.00 < 0.01Estevan Point 1032730 31.47 < 0.01Fort Nelson 1192940 19.04 < 0.01Fort St James 1092970 6.54 < 0.01Fort St John 1183000 24.82 < 0.01Germansen Landing 1183090 12.21 < 0.01Joe Rich Creek 1123750 9.29 < 0.01Langara 1054500 41.98 < 0.01Port Alice 1036240 26.04 < 0.01Port Hardy 1026270 26.71 < 0.01Prince George 1096450 6.85 < 0.01Prince Rupert 1066481 4.64 0.02Saanichton 1016940 19.00 < 0.01Sandspit 1057050 28.01 < 0.01Shawnigan Lake 1017230 8.34 < 0.01Vancouver 1108447 30.79 < 0.01Williams Lake 1098940 17.26 < 0.01the 1979-2002 period, large NRMSE values and the large positive NMBE values were observedat stations along the coast. The large NMBE suggests that the NARR record consistentlyunderestimated total annual precipitation observed in the station record along the coast duringthis period. On the other hand, during the 2003-2015 period, large NRMSE and large negativeNMBE were observed at stations in the Interior, suggesting that the NARR record consistentlyoverestimated total annual precipitation observed at these stations.Notably, stations located in the south of the province (Saanichton, Shawnigan Lake, Van-couver) do not follow the pattern observed above, in which the 1979-2002 period had strongeragreement between datasets than the 2003-2015 period. Instead, the error statistics for thesestations showed either minor increases or significant decreases during the 2003-2015 period.The NARR total annual precipitation records were more consistent with these station recordswhen the Canadian rain gauge data were no longer being assimilated into the NARR model.Figures 3.3 and 3.4 illustrate the relationship between NARR and total annual precipitationand Table 3.3 summarizes the regression fits. Generally, the R2 values were higher during the1979-2002 period.21Table 3.2: NRMSE and NMBE values calculated from the total annual precipitation forthe 18 stations during the 1979-2002 and 2003-2015 periods.Station Name Station ID1979-2002 2003-2015NRMSE NMBE NRMSE NMBEComox 1021830 0.256 -0.227 0.659 -0.609Dryad Point 1062544 0.096 -0.014 0.600 -0.624Estevan Point 1032730 0.160 -0.139 0.100 -0.018Fort Nelson 1192940 0.109 0.083 0.170 -0.153Fort St James 1092970 0.453 -0.442 0.933 -1.000Fort St John 1183000 0.124 0.059 0.195 -0.119Germansen Landing 1183090 0.278 -0.268 0.749 -0.715Joe Rich Creek 1123750 0.097 -0.020 0.216 -0.194Langara 1054500 0.303 0.292 0.216 0.172Port Alice 1036240 0.431 0.423 0.497 0.462Port Hardy 1026270 0.083 -0.033 0.233 -0.177Prince George 1096450 0.148 -0.110 0.521 -0.485Prince Rupert 1066481 0.138 0.118 0.160 0.156Saanichton 1016940 0.176 0.149 0.208 -0.164Sandspit 1057050 0.195 0.131 0.108 0.0214Shawnigan Lake 1017230 0.728 0.715 0.208 -0.005Vancouver 1108447 0.111 0.074 0.068 -0.029Williams Lake 1098940 0.221 -0.181 0.544 -0.48722Figure 3.2: Map of the NRMSE and NMBE values calculated from the total annualprecipitation for the 18 stations during the 1979-2002 (left panel) and 2003-2015(right panel) periods.23Figure 3.3: Plot of station total annual precipitation vs NARR total annual precipitationfor the 1979-2002 period. Blue lines are the 1:1 lines and black lines are the linesof best fit through the scatter.24Figure 3.4: Plot of station total annual precipitation vs NARR total annual precipitationfor the 2003-2015 period. Blue lines are the 1:1 lines and black lines are the linesof best fit through the scatter.25Table 3.3: Summary of regressions between station total annual precipitation and NARRtotal annual precipitation for the 1979-2002 and 2003-2015 periods. Not enough datawas available to fit a regression for Prince Rupert following 2002.Station Name Station ID1979-2002 2003-2015Intercept Slope R2 Intercept Slope R2Comox 1021830 275.2 0.63 0.59 604.0 0.29 0.41Dryad Point 1062544 520.0 0.79 0.51 1304.3 0.35 0.42Estevan Point 1032730 318.4 0.79 0.75 1475.9 0.52 0.17Fort Nelson 1192940 11.0 1.07 0.83 45.5 0.80 0.89Fort St James 1092970 8.9 0.68 0.65 209.1 0.32 0.23Fort St John 1183000 -75.6 1.21 0.81 -9.5 0.91 0.43Germansen Landing 1183090 44.3 0.73 0.70 77.3 0.50 0.50Joe Rich Creek 1123750 109.0 0.80 0.59 -230.6 1.17 0.90Langara 1054500 785.3 0.88 0.38 1690.1 0.10 0.00Port Alice 1036240 1385.6 1.04 0.52 2749.0 0.34 0.06Port Hardy 1026270 600.8 0.67 0.72 1292.0 0.31 0.47Prince George 1096450 42.9 0.84 0.56 565.7 0.11 0.21Prince Rupert 1066481 1097.6 0.67 0.50 NA NA NASaanichton 1016940 229.4 0.89 0.73 206.3 0.67 0.41Sandspit 1057050 965.4 0.41 0.16 1538.1 -0.10 0.02Shawnigan Lake 1017230 454.3 2.27 0.57 1009.5 0.19 0.05Vancouver 1108447 347.6 0.78 0.70 265.5 0.76 0.73Williams Lake 1098940 -284.3 1.31 0.66 163.8 0.44 0.29263.1.2 Analysis of Precipitation ClimatologiesThe precipitation climatologies for both periods (1979-2002, 2003-2015) at each of the 18 sta-tions and corresponding NARR grid cells are shown in Figure 3.5 and Figure 3.6. Visualcomparisons between the two figures highlight a stronger agreement between station precipi-tation climatologies and NARR precipitation climatologies during the 1979-2002 period. Thisagreement is further confirmed by both NRMSE and NMBE evaluations (Table 3.4), as theNRMSE were generally larger and NMBE were more negative in the 2003-2015 period than inthe 1979-2002 period.Table 3.4: NRMSE and NMBE values for the precipitation climatologies produced fromNARR monthly precipitation and station monthly precipitation for the 1979-2002and 2003-2015 periods.Station Name Station ID1979-2002 2003-2015NRMSE NMBE NRMSE NMBEComox 1021830 0.243 -0.219 0.692 -0.662Dryad Point 1062544 0.227 -0.010 0.783 -0.649Estevan Point 1032730 0.217 -0.145 0.085 0.016Fort Nelson 1192940 0.123 0.078 0.253 -0.139Fort St James 1092970 0.589 -0.446 1.296 -1.205Fort St John 1183000 0.102 0.062 0.214 -0.153Germansen Landing 1183090 0.423 -0.281 0.768 -0.736Joe Rich Creek 1123750 0.439 -0.049 0.439 -0.296Langara 1054500 0.304 0.288 0.219 0.201Port Alice 1036240 0.528 0.426 0.619 0.483Port Hardy 1026270 0.074 -0.026 0.237 -0.208Prince George 1096450 0.241 -0.111 0.548 -0.495Prince Rupert 1066481 0.157 0.122 0.204 0.075Saanichton 1016940 0.230 0.142 0.272 -0.180Sandspit 1057050 0.195 0.132 0.160 0.011Shawnigan Lake 1017230 0.880 0.719 0.216 -0.052Vancouver 1108447 0.120 0.073 0.190 -0.052Williams Lake 1098940 0.250 -0.182 0.586 -0.56027Figure 3.5: Precipitation climatologies produced from the NARR monthly precipitation(grey line) and station monthly precipitation (black line) for the 1979-2002 period.28Figure 3.6: Precipitation climatologies produced from the NARR monthly precipitation(grey line) and station monthly precipitation (black line) for the 2003-2015 period.293.1.3 Analysis of Monthly VariabilitySummaries of error indices are shown in Table 3.7 and Figure 3.7. In the southern part of theprovince, both NRMSE and NMBE during the 1979-2002 period were relatively small, withNMBE being predominantly negative. In the northern part of the province, while NRMSE wasrelatively small, NMBE was predominantly positive.Table 3.5: NRMSE and NMBE values for the NARR monthly precipitation and stationmonthly precipitation for the 1979-2002 and 2003-2015 periods.Station Name Station ID1979-2002 2003-2015NRMSE NMBE NRMSE NMBEComox 1021830 0.398 -0.219 0.792 -0.662Dryad Point 1062544 0.350 -0.009 0.870 -0.642Estevan Point 1032730 0.334 -0.146 0.267 0.015Fort Nelson 1192940 0.319 0.078 0.467 -0.141Fort St James 1092970 0.705 -0.446 1.478 -1.232Fort St John 1183000 0.320 0.062 0.451 -0.153Germansen Landing 1183090 0.530 -0.283 0.860 -0.738Joe Rich Creek 1123750 0.550 -0.046 0.569 -0.323Langara 1054500 0.381 0.288 0.306 0.201Port Alice 1036240 0.614 0.426 0.722 0.483Port Hardy 1026270 0.209 -0.026 0.363 -0.212Prince George 1096450 0.378 -0.111 0.641 -0.492Prince Rupert 1066481 0.245 0.122 0.266 0.086Saanichton 1016940 0.402 0.141 0.435 -0.180Sandspit 1057050 0.352 0.131 0.324 0.013Shawnigan Lake 1017230 0.978 0.719 0.381 -0.052Vancouver 1108447 0.300 0.073 0.328 -0.058Williams Lake 1098940 0.419 -0.181 0.685 -0.563For 16 of the 18 stations, the model including month was significantly better than thereduced model, indicating that the relations between observed and NARR monthly precipitationvary by month (Table 3.6). As an example, Figure 3.10 presents the month by month regressionsfor the Port Hardy station.The regression models including month (Eq. 2.5) explained 66-92% of the variance for the1979-2002 period, but performed more poorly for the 2003-2015 period, explaining 20-84% ofthe the variation (Figure 3.8, Figure 3.9 and Table 3.6).30Figure 3.7: Map of the NRMSE and NMBE values for the NARR monthly precipitationand station monthly precipitation for the 1979-2002 (left panel) and 2003-2015(right panel) periods.31Table 3.6: Adjusted R2 values from the linear model predicting station monthly precipi-tation from NARR monthly precipitation and month and p-values from the ANOVAcomparing linear models 2.6 and 2.5 for the 1979-2002 and 2003-2015 periods.Station Name Station ID1979-2002 2003-2015R2 p-value R2 p-valueComox 1192940 0.87 < 0.01 0.84 < 0.01Dryad Point 1026270 0.79 < 0.01 0.71 < 0.01Estevan Point 1032730 0.90 < 0.01 0.89 0.02Fort Nelson 1108447 0.85 < 0.01 0.84 0.04Fort St James 1098940 0.66 < 0.01 0.40 < 0.01Fort St John 1054500 0.83 0.20 0.83 0.02Germansen Landing 1017230 0.75 < 0.01 0.43 < 0.01Joe Rich Creek 1123750 0.73 < 0.01 0.20 < 0.01Langara 1066481 0.76 < 0.01 0.72 0.19Port Alice 1062544 0.86 < 0.01 0.82 < 0.01Port Hardy 1021830 0.92 < 0.01 0.91 < 0.01Prince George 1016940 0.74 < 0.01 0.59 0.02Prince Rupert 1057050 0.88 < 0.01 0.84 0.51Saanichton 1183000 0.83 < 0.01 0.80 < 0.01Sandspit 1096450 0.78 < 0.01 0.71 < 0.01Shawnigan Lake 1092970 0.88 < 0.01 0.85 < 0.01Vancouver 1183090 0.85 0.12 0.84 < 0.01Williams Lake 1036240 0.73 < 0.01 0.64 0.4132Figure 3.8: Plot of station total monthly precipitation vs NARR total monthly precipi-tation for the 1979-2002 period. Blue lines are the 1:1 lines and black lines are thelines of best fit through the scatter calculated using the reduced model (Eq. 2.6).33Figure 3.9: Plot of station total monthly precipitation vs NARR total monthly precipi-tation for the 2003-2015 period. Blue lines are the 1:1 lines and black lines are thelines of best fit through the scatter calculated using the reduced model (Eq. 2.6).34Figure 3.10: Plot of station total monthly precipitation vs NARR total monthly precip-itation for each month of the year for the Port Hardy station. Blue lines are the1:1 lines and black lines are the lines of best fit through the scatter calculatedusing the full model (Eq. 2.5).353.1.4 Analysis of Daily PrecipitationThe NARR precipitation records matched the timing of daily precipitation events in the stationrecords between 67-84% of the days during the 1979-2002 period (Table 3.7). Similar resultswere observed during the 2003-2015 period, with the NARR properly modelling the timing ofdaily events between 63-81% of the time. In comparing the 1979-2002 period with the 2003-2015 period, the number of times the NARR predicted precipitation when no precipitation wasrecorded in the station record increased at 17 of the 18 stations, and in some cases by as muchas 10-15%. This increase is accompanied by slight decreases in both the occurrence of theNARR correctly predicting precipitation, and the occurrence of the NARR failing to predictprecipitation. These changes suggest that the NARR is modelling more frequent precipitationevents in the 2003-2015 period than has been recorded at the stations themselves.Table 3.7: Contingency analysis of NARR’s ability to predict wet vs dry days for the 1979-2002 and 2003-2015 periods, showing when the NARR record correctly predictedprecipitation, incorrectly predicted precipitation (error of commission) and failed topredict precipitation (error of omission).Station Name1979-2002 2003-2015Correct Commission Omission Correct Commission OmissionComox 81.58% 13.34% 5.08% 77.23% 20.17% 2.60%Dryad Point 84.15% 9.89% 5.97% 79.96% 15.85% 4.19%Estevan Point 80.72% 13.79% 5.50% 77.53% 16.79% 5.68%Fort Nelson 74.36% 17.82% 7.82% 68.79% 27.50% 3.72%Fort St James 67.33% 28.23% 4.43% 54.15% 43.16% 2.69%Fort St John 70.71% 23.21% 6.07% 63.61% 33.83% 2.56%Germansen Landing 71.64% 23.24% 5.13% 70.53% 24.67% 4.80%Joe Rich Creek 72.34% 21.75% 5.91% 69.65% 25.54% 4.81%Langara 83.35% 9.22% 7.43% 81.10% 15.56% 3.33%Port Alice 81.24% 11.20% 7.56% 79.25% 12.80% 7.95%Port Hardy 81.96% 12.86% 5.18% 77.96% 19.79% 2.25%Prince George 75.30% 17.03% 7.67% 73.98% 21.48% 4.54%Prince Rupert 83.85% 12.24% 3.90% 81.87% 11.97% 6.17%Saanichton 78.84% 16.08% 5.08% 71.75% 23.21% 5.04%Sandspit 82.67% 9.42% 7.91% 79.47% 15.51% 5.01%Shawnigan Lake 80.02% 8.94% 11.04% 70.90% 25.61% 3.49%Vancouver 83.76% 9.51% 6.73% 77.64% 19.04% 3.33%Williams Lake 71.59% 23.90% 4.51% 68.13% 29.22% 2.65%The frequency curves of maximum daily precipitation calculated from the station recordsand the NARR records (Figure 3.11) show that the NARR failed to represent the occurrenceof extreme events at many stations. The ratio of NARR maximum daily rainfall to stationmaximum daily rainfall consistently fell around 75% for the 2, 5, 10, 20, 50, and 100 yearevents (Figure 3.12). On average, the NARR underpredicted the magnitude of these events by36approximately 25%. Several stations (Joe Rich Creek, Port Alice) showed as much as a 50%reduction in the magnitude of these events.37Figure 3.11: Frequency plots of maximum daily rainfall for the NARR record (grey line)and the station record (black line). Lines fitted using method of moments as-suming the data follow a Gumbel distribution. The data were not split into twoperiods to provide a larger number of years to calculate the return period ofextreme events.38Figure 3.12: Ratio of NARR to observed daily maximum rainfall by return period forthe 18 stations. Each dot represents one of the 18 stations. A value of < 1 impliesNARR maximum daily precipitation is lower than the station maximum dailyprecipitation.3.2 Comparison of NARR Precipitation and StreamflowMeasurementsThe relationship between NARR average water year precipitation and basin water yield duringeach of the two periods is shown in Figure 3.13. During the 1979-2002 period, 18 stationshad a greater amount of water yield than NARR precipitation. During the 2003-2015 period,only 12 stations exhibited this same relationship. In both periods, four basins had nearly twicethe amount of water yield than precipitation. The spatial pattern of the difference betweenprecipitation and streamflow is shown in Figure 3.14. During the 1979-2002 period, basinswith negative values were clustered in the Coast Mountains and along the western edge of theRocky Mountains. During the 2003-2015 period, fewer basins along the western edge of theRocky Mountains showed these unrealistic negative values. During both periods, basins in theInterior of the province exhibited a more realistic scenario with positive differences. Interiorbasins show an interesting spatial pattern, in which basins farther north have larger differences39between precipitation and streamflow. In some cases, differences greater than 500 mm/yr areobserved in these basins.Figure 3.13: Relation between mean annual catchment water yield and NARR meanannual catchment-average precipitation for the study basins during the 1979-2002and 2003-2015 periods. The 1:1 line is shown in black with the grey shadingshowing ± 5%. Basins with points above the 1:1 line represent basins in whichthe amount of water leaving the basin as streamflow is greater than the amountof NARR precipitation entering the basin, while basins below the 1:1 line showa more realistic relationship where there is more water entering the basin thanleaving as streamflow.The plots of both the NARR total annual precipitation and basin water yield display thesame pattern as observed in the average water year difference between precipitation and stream-flow (Figures 3.15, 3.16, and 3.17). The mountainous basins are apparent, with higher amountsof recorded streamflow than modelled precipitation, showing the unrealistic situation in whichsome basins have a water deficit of greater than 1000 mm/yr. As well, the basins with largepositive differences between modelled precipitation and observed streamflow are also apparentin the figures.40Figure 3.14: Map of the difference in the average NARR precipitation and average basin water yield for the study basins duringthe 1979-2002 (left panel) and 2003-2015 (right panel) periods41Figure 3.15: Time series of annual catchment water yield (black line) and total annualcatchment-averaged precipitation (grey line). The panel labels indicate the WSCstation numbers (see Table 2.2 for catchment details). Part 1 of 3.42Figure 3.16: Time series of annual catchment water yield (black line) and total annualcatchment-averaged precipitation (grey line). The panel labels indicate the WSCstation numbers (see Table 2.2 for catchment details). Part 2 of 3.43Figure 3.17: Time series of annual catchment water yield (black line) and total annualcatchment-averaged precipitation (grey line). The panel labels indicate the WSCstation numbers (see Table 2.2 for catchment details). Part 3 of 3.44Chapter 4Discussion4.1 Temporal Variability of ErrorsA clear break in the NARR precipitation time series was observed at all 18 stations as theChow Test showed a significant step change in the data between the 1979-2002 and 2003-2015periods. Changes in the total annual precipitation, precipitation climatologies, monthly totalprecipitation, and timing of daily events were all recorded.Many of the NARR records for the 18 stations (Comox, Dryad Point, Fort St. James, FortSt. John, Germansen Landing, Prince George, Shawnigan Lake, and Williams Lake) exhibiteda change in the amount and variability of total annual precipitation during the 2003-2015period. In many of these cases, this change caused the NARR records and station records togreatly differ, as indicated by the NRMSE and NMBE values. During the 2003-2015 period,the increase in NRMSEs and NMBEs was again observed in the precipitation climatologies andtotal monthly precipitation results. A greater occurrence of the NARR modelling precipitationwhen no precipitation was recorded was also observed during this period.One interesting pattern suggests that stations located near the U.S.-Canada border seemedto be less affected by the change in the data assimilation process. A comparison of the 1979-2002 and 2003-2015 periods showed that NRMSE and NMBE for stations near the border didnot show the same increase observed at other stations. These border stations either experiencedsimilar levels of agreement between the NARR and station records (Saanichton, Joe Rich Creek)or, in one case (Shawnigan Lake), showed a greater level of agreement between the NARR andstation records. Since precipitation data from the U.S. were still being included in the dataassimilation process, it is possible that these southern stations continued to benefit from theirproximity to the U.S., and thus did not show the large discrepancies between the NARR recordsand station records that are observed at many of the stations further north.A similar split in the NARR’s precipitation data was observed by Eum et al. (2014) whorecorded a similar increase in the amount of annual precipitation in the Athabasca River basin.This has been confirmed by the Chow Test above, as well as by visual inspection of total45annual precipitation plots. Another discontinuity in precipitation rate exists at the Canada-U.S. border, due to the different spatial resolution of input data for both Canada and the U.S.(Jarosch et al., 2012).At the daily time scale the NARR records for all 18 stations underpredicted the occurrenceof extreme events. This inability to capture the magnitude of extreme events is not uniqueto the NARR, as many other reanalysis and large-scale weather and climate models reportsimilar findings (Isotta et al., 2015; Rienecker et al., 2011; Becker et al., 2009). Many ofthese datasets report a tendency to overpredict the occurrence of light precipitation eventswhile simultaneously underpredicting extreme events (Isotta et al., 2015; Rienecker et al., 2011;Becker et al., 2009).4.2 Spatial Variability of ErrorsThe spatial variability of errors in the province of British Columbia varies with the diversetopographic features of the region. The following section will discuss the patterns of errorexperienced in the province by region and attempt to identify the potential causes for theseerrors.4.2.1 Coast MountainsGenerally speaking, the NARR tends to underestimate precipitation in the Coast Mountains.At the annual time scale, NARR precipitation estimates were substantially lower than thestation records at many of the 18 stations. In addition, an examination of the water balance foreach of the basins in the Coast Mountains shows more water leaving the basin as streamflowthan entering as precipitation.At a monthly time scale, the NARR is able to reproduce the distribution of precipitationin the Coast Mountains throughout the year. While it can replicate the seasonal pattern ofwet winters and dry summers experienced in this region, it fails to replicate the magnitudeof precipitation during the winter months. The precipitation climatologies for Langara, PortAlice, and Shawnigan Lake all demonstrate that, on average, less precipitation is modelledby the NARR than is recorded during the winter months – in some cases a difference greaterthan 200 mm monthly. The monthly variability of precipitation showed that the monthlyprecipitation estimates at these stations are substantially lower than the observed record.Finally, the daily precipitation records showed that the NARR precipitation data fails torepresent the magnitude of extreme precipitation events in the Coast Mountains. These resultssuggest that the NARR is unable to properly model the orographic precipitation experiencedin this region, as stations located on the westward side of the Coast Mountains showed thegreatest discrepancies between observed and modeled precipitation.A study of the NARR’s digital elevation model (DEM) reveals that terrain smoothing hasoccurred as a result of the relatively coarse grid size. This terrain smoothing has caused largemisrepresentations in the topography of the Coast Mountains, as is shown in Figure 2.1. In46small catchments in the Coast Mountains, where the discrepancy between the NARR DEM andthe observed topography is greatest (as large as 500 m; see Table 2.2), the NARR seems to failto model the magnitude of the orographic effect.One strong example that highlights the NARR’s inability to model the magnitude of pre-cipitation experienced on the windward slopes of the Coast Mountains is the difference betweenthe Port Alice and Port Hardy stations. Though both of these stations are located at a similarlatitude, and show similar patterns of precipitation, the Port Alice station is located on thewindward side of Vancouver Island while the Port Hardy station is located on the leeward side.Presumably, as a result of their locations, the Port Alice station experiences a difference of 1000mm between the observed station record and the NARR record, while the Port Hardy stationexperiences virtually no difference.The generalization of the NARR DEM has likely caused the magnitude of precipitationmodelled at the Port Alice station to be substantially less than the actual recorded precipitation.It is important to note that local topographic features may also influence the NARR’s inabilityto properly model precipitation patterns. As a result of being located in a narrow fjord, PortAlice may exhibit a degree of enhanced precipitation which is not resolved by the NARR model.This topographic feature is too small to be resolved by the NARR’s digital elevation model,and thus the NARR does not fully reflect the precipitation patterns experienced in this area.Despite these local topographic considerations, the dominant pattern to recognize in theseresults is that the NARR consistently underestimates precipitation along the windward slopesof the Coast Mountains.4.2.2 Interior PlateauStations in the southern Interior of British Columbia show that the NARR is better able tomodel precipitation patterns in this region than it was along the Coast Mountains. Morespecifically, the Joe Rich Creek and Williams Lake stations show a strong agreement betweenNARR total annual precipitation and observed total annual precipitation. The comparison ofprecipitation and streamflow in the Interior shows a much more realistic relationship betweenmodelled precipitation and measured streamflow. In basins located in this region, more wateris entering (as NARR precipitation estimates) than is leaving (as streamflow). This differencebetween precipitation and streamflow allows for evapotranspiration loss, which differs from theCoast Mountains, where a deficit existed in the water balance.Similarly, NARR total monthly precipitation in this region showed strong agreement withthe observed total monthly precipitation, as evidenced by relatively small NRMSE and NMBEvalues. Regression models using NARR total monthly precipitation, and month as a factorvariable, were able to explain much of the variability experienced in the observed total monthlyprecipitation. However, analysis of mean monthly precipitation shows that the NARR is ableto reproduce the seasonal distribution of precipitation experienced at Williams Lake, but failsto follow the pattern of seasonal precipitation experienced at Joe Rich Creek. Additionally, the47NARR fails to reproduce the occurrence of extreme precipitation events at both stations.In the northern Interior region (Fort St. James, Germansen Landing, Prince George), NARRprecipitation estimates are unable to reliably model observed precipitation patterns. NARRestimates in this region tend to overestimate the amount of total precipitation received whencompared to actual precipitation experienced. There are also inconsistencies with respect toseasonal distribution of precipitation: while summer is modelled relatively accurately, precipi-tation in winter is regularly overestimated. Overestimation can also be observed in the monthlyvariability of this region, where the total monthly precipitation from the NARR record tends tobe greater than the total monthly precipitation observed in the station record. When exploringthe relationship between measured streamflow and NARR precipitation in the northern Interior,it is clear that there is more precipitation entering the basins then leaving as streamflow.In examining the differences between the northern and southern Interior, a north-southgradient in the accuracy of the NARR precipitation is apparent. This gradient is readilyobserved in the magnitude of difference between precipitation and streamflow. Figure 2.1shows that there is a smaller difference between the amount of precipitation and the amount ofstreamflow in the south, and a larger difference in basins in the north. This relationship is alsoshown in the comparison of the NARR and station records where, in the south, a strong matchexists between station and NARR data, while in the north, the station data do not match theNARR due to an overestimation of precipitation.The NARR model has been shown to be dependent on the assimilation of precipitation datainto the model, as evidenced by the clear break in data between 2002 and 2003, when Canadianprecipitation gauge data were excluded from the model. As there are fewer precipitation gaugedata available for the northern part of British Columbia, the NARR precipitation estimates areless accurate overall.The Interior region highlights that the accuracy of the NARR’s precipitation estimates isvery much dependent on the quantity of rain gauge data being assimilated into the model. Theprecipitation estimates for the southern Interior can more accurately reproduce the observedpatterns of precipitation due to a greater number of precipitation gauge stations in this region.Conversely, precipitation estimates for the northern Interior have been shown to be much lessaccurate, likely due to the lack of available precipitation gauge stations.4.2.3 Rocky MountainsThere are fewer data available in the Rocky Mountain region as a result of there being noavailable stations in this region. However, basins located along the western edge of the regionwere used to show that precipitation is again being underestimated by the NARR. These findingsare consistent with Eum et al. (2014) who provide analysis showing the NARR underestimatesprecipitation in the Athabasca River basin located on the eastern edge of the Rockies.Basins in the south of the region show a slight deficit in the water balance, as the measuredstreamflow is greater than the estimated NARR values. Basins in the north of this region show48similar patterns.This pattern is similar to that of the Coast Mountains, as terrain smoothing in the NARRDEM exist. This pattern, however, is less apparent than it is in the Coast Mountains, likelydue to the fact that less precipitation is recorded in this region overall. Again, differences inthe digital elevation data of the NARR model result in inaccurate precipitation estimates overcomplex terrain.4.2.4 Northeast PlateauIn British Columbia’s northeast region, Fort Nelson and Fort St. John stations showed NARR’sability to reproduce observed patterns of precipitation. NARR precipitation estimates wereable to accurately reproduce the interannual variability, seasonal pattern of precipitation, andmonthly variability, but the NARR failed to capture the occurrence of extreme precipitationevents, as it again underestimated annual maximum daily rainfall.The NARR is able to accurately reproduce precipitation patterns in this region because itsterrain is fairly homogenous and analysis of the NARR DEM shows little difference between theNARR and actual terrain. The station data for this region are sufficient in their ability to recordthe spatial pattern of precipitation. Because the NARR model has an accurate representationof the terrain and sufficient precipitation data to constrain the water cycle, NARR precipitationestimates are shown to match with observed precipitation patterns.4.3 The Utility of the NARR Precipitation EstimatesIn evaluating NARR’s accuracy in BC, it becomes clear that many issues such as terrain andbias in the station record impact the NARR’s effectiveness. Many topographic features in-fluencing precipitation formation become important at scales less than 50 km (Daly, 2006).Thus, NARR’s inability to accurately model the terrain of the region impact the accuracy of itsprecipitation estimates over heterogeneous areas. European Regional Climate Models (RCMs)have reported similar findings in the European Alps where, at scales similar to the NARRand with similar biased input data, the magnitude of precipitation in mountainous areas wasincorrectly modelled (Frei et al., 2003).Another important failing of the NARR is the loss of data continuity with the removal ofCanadian precipitation gauge data from the assimilation process. This change caused large in-accuracies in NARR data from 2003 onwards. At many stations, a large increase in the amountof total annual precipitation was observed. This increase in precipitation is similar to precipita-tion modelled by the NCEP/NCAR global reanalysis R-1 and R-2, which reported much higherprecipitation values than observed in North America (Kalnay et al., 1996; Kistler et al., 2001).Similar to these global reanalyses, the NARR estimates larger amounts of precipitation thanobserved in many areas of BC when it is no longer assimilating precipitation data.However, NARR precipitation data still provide useful information to areas of BC lackingother data sources. The analysis suggested that during the 1979-2002 period, much of the49variability in the observed precipitation data could be explained by variability in NARR data(Table 3.6), making it possible for precipitation data to be corrected and better reflect stationdata. Jarosch et al. (2012) developed a method for predicting spatially distributed precipita-tion at 1 km resolution by combining NARR-predicted atmospheric moisture transport withan orographic precipitation model. This scheme outperformed raw NARR precipitation overthe period 1990-2002. In recent years (2003-present), NARR precipitation estimates are lessaccurate. Correcting these values may not be possible, as NARR records can no longer explainmuch of the variability in the station record. These inaccuracies call the utility of this datasetinto question.50Chapter 5Conclusion5.1 Summary of Key ResultsThis study has identified several key failings of the NARR precipitation dataset for BritishColumbia. First, the exclusion of Canadian precipitation gauge data from the data assimilationprocess caused significant changes to NARR precipitation data from 2003 onwards. With theexclusion of Canadian data, the ability of NARR precipitation estimates to reproduce observedpatterns of precipitation deteriorated, as observational data were no longer constraining theNARR model.Second, while the NARR shows a strong ability to reproduce observed patterns of precip-itation in areas where both the terrain is properly modelled by the NARR DEM and wheresufficient station data exist to constrain the NARR model, it has trouble estimating precip-itation over complex terrain or in areas lacking station data to constrain the NARR model.Therefore, the NARR properly represents spatial and temporal patterns of precipitation inboth the southern part of the Interior and the Northeast Plateau, as topography in the NARRis similar to reality and the regions station network provides sufficient coverage. However, theNARR data tend to underestimate precipitation in mountainous regions, due in part to thecoarse resolution of the NARR DEM, causing both terrain smoothing and failing to model ele-vation extremes. Also, NARR overestimates precipitation in the northern part of the Interior,where less rain gauge stations are available to be assimilated into the NARR model.This study has demonstrated the errors associated with the North American Regional Re-analysis precipitation dataset by highlighting areas in which it fails to accurately reproduceobserved patterns of precipitation. It has also highlighted that NARR struggles to properlyestimate precipitation over mountainous terrain and in situations where sufficient station datais not available. The accuracy of precipitation data from reanalysis systems has been shown tobe reliant on both a DEM that can accurately model terrain and sufficient observational datato constrain the water cycle of the model. Given the need for these two features in reanalysissystems, caution should be taken when using reanalysis precipitation data in areas similar to51the mountainous region of British Columbia (heterogeneous terrain, and sparse stations cover-age). Given the errors highlighted here, this study should highlight the need for independentassessments of precipitation datasets. This study could be used as an outline to evaluate spatialand temporal variability of errors for other datasets in other regions, as it provides a meansof observing if a dataset can accurately model observed spatial and temporal patterns of pre-cipitation in a given region. Given the heterogeneous nature of precipitation data, users ofany dataset in any region should understand the errors and limitations of their chosen datasetbefore use.5.2 Future Research DirectionFuture work on the NARR precipitation data could focus on developing a means of correctingprecipitation estimates for mountainous environments in British Columbia. The NARR wasable to model the pattern of wet and dry years, but failed to properly model the magnitude ofprecipitation received. A correction that scales up NARR precipitation estimates could improvethe accuracy of the NARR precipitation fields. Jarosch et al. (2012) provided a method ofdownscaling NARR data, but the accuracy of this method was only tested for 1990-2002. Thismethod of downscaling could be tested to determine whether it could increase the accuracy ofNARR precipitation estimates for data produced after 2003.Similar to this study, an assessment of the NARR’s snow cover data could be conductedto determine the accuracy of the NARR snow depth field, and whether similar errors areassociated with its snow depth estimates. While accurate spatial estimates of snow depth forBritish Columbia would be useful in better understanding the cold weather hydrology of thisregion, NARR estimates should first be assessed to determine their accuracy before being used,mainly due to the inaccuracies apparent in the data’s precipitation fields.52ReferencesBecker, E. J., Berbery, E. H., and Higgins, R. W. Understanding the characteristics of dailyprecipitation over the United States using the North American Regional Reanalysis.Journal of Climate, 22:6268–6286, 2009. doi:10.1175/2009JCLI2838.1. → pages 5, 46Bedient, P. B. and Huber, W. Hydrology and Floodplain Analysis. Addison-Wesley, Reading,Massachusetts, second edition edition, 1992. → pages 17Berne, A. and Krajewski, W. F. Radar for hydrology: Unfulfilled promise or unrecognizedpotential? Advances in Water Resources, 51:357–366, 2013.doi:10.1016/j.advwatres.2012.05.005. → pages 1, 4Bukovsky, M. S. and Karoly, D. J. A brief evaluation of precipitation from the NorthAmerican Regional Reanalysis. Journal of Hydrometeorology, 8:837–846, 2007.doi:10.1175/JHM595.1. → pages 4, 5Chow, G. C. Tests of equality between sets of coefficients in two linear regressions.Econometrica: Journal of the Econometric Society, 28:591–605, 1960. doi:10.2307/1910133.→ pages 15Daly, C. Guidelines for assessing the suitability of spatial climate data sets. InternationalJournal of Climatology, 26:707–721, 2006. doi:10.1002/joc.1322. → pages 1, 2, 49Daly, C., Neilson, R. P., and Phillips, D. L. A statistical-topographic model for mappingclimatological precipitation over mountainous terrain. Journal of Applied Meteorology, 33:140–158, 1994. doi:10.1175/1520-0450(1994)033〈0140:ASTMFM〉2.0.CO;2. → pages 2Daly, C., Gibson, W. P., Taylor, G. H., Johnson, G. L., and Pasteris, P. A knowledge-basedapproach to the statistical mapping of climate. Climate Research, 22:99–113, 2002.doi:10.3354/cr022099. → pages 7Daly, C., Halbleib, M., Smith, J. I., Gibson, W. P., Doggett, M. K., Taylor, G. H., Curtis, J.,and Pasteris, P. P. Physiographically sensitive mapping of climatological temperature andprecipitation across the conterminous United States. International Journal of Climatology,28:2031–2064, 2008. doi:10.1002/joc.1688. → pages 1, 3Ensor, L. A. and Robeson, S. M. Statistical characteristics of daily precipitation: comparisonsof gridded and point datasets. Journal of Applied Meteorology and Climatology, 47:2468–2476, 2008. doi:10.1175/2008JAMC1757.1. → pages 2Eum, H.-I., Dibike, Y., Prowse, T., and Bonsal, B. Inter-comparison of high-resolutiongridded climate data sets and their implication on hydrological model simulation over the53Athabasca Watershed, Canada. Hydrological Processes, 28:1099–1085, 2014.doi:10.1002/hyp.10236. → pages 5, 15, 45, 48Frei, C., Christensen, J. H., Deque, M., Jacob, D., Jones, R. G., and Vidale, P. L. Dailyprecipitation statistics in regional climate models: Evaluation and intercomparison for theEuropean Alps. Journal of Geophysical Research: Atmospheres, 108, 2003. ISSN 2156-2202.doi:10.1029/2002JD002287. → pages 49Harman, J. R. Synoptic climatology of the westerlies: process and patterns. ResourcePublications in Geography. Association of American Geographers, Washington D.C., 1991.→ pages 2Hopkinson, R. F., McKenney, D. W., Milewska, E. J., Hutchinson, M. F., Papadopol, P., andVincent, L. A. Impact of aligning climatological day on gridding daily maximum-minimumtemperature and precipitation over Canada. Journal of Applied Meteorology andClimatology, 50:1654–1665, 2011. doi:10.1175/2011JAMC2684.1. → pages 3Hutchinson, M. F. Interpolating mean rainfall using thin plate smoothing splines.International Journal of Geographical Information Systems, 9:385–403, 1995.doi:10.1080/02693799508902045. → pages 3Hutchinson, M. F., McKenney, D. W., Lawrence, K., Pedlar, J. H., Hopkinson, R. F.,Milewska, E., and Papadopol, P. Development and testing of canada-wide interpolatedspatial models of daily minimum-maximum temperature and precipitation for 1961-2003.Journal of Applied Meteorology and Climatology, 48:725–741, 2009.doi:10.1175/2008JAMC1979.1. → pages 1, 3, 4, 5Isotta, F. A., Vogel, R., and Frei, C. Evaluation of European regional reanalyses anddownscalings for precipitation in the Alpine region. Meteorologische Zeitschrift, 24:15–37,2015. doi:10.1127/metz/2014/0584. → pages 46Jarosch, A. H., Anslow, F. S., and Clarke, G. K. C. High-resolution precipitation andtemperature downscaling for glacier models. Climate Dynamics, 38:391–409, 2012.doi:10.1007/s00382-010-0949-1. → pages 46, 50, 52Kalnay, E., Kanamitsu, M., Kistler, R., Collins, W., Deaven, D., Gandin, L., Iredell, M.,Saha, S., White, G., and Woollen, J. The NCEP/NCAR 40-year reanalysis project.Bulletin of the American Meteorological Society, 77:437–471, 1996.doi:10.1175/1520-0477(1996)077〈0437:TNYRP〉2.0.CO;2. → pages 4, 49Kanamitsu, M., Ebisuzaki, W., Woollen, J., Yang, S.-K., Hnilo, J. J., Fiorino, M., and Potter,G. L. NCEP-DOE AMIP-II Reanalysis (R-2). Bulletin of the American MeteorologicalSociety, 83:1631–1643, 2002. doi:10.1175/bams-83-11-1631. → pages 4Kistler, R., Collins, W., Saha, S., White, G., Woollen, J., Kalnay, E., Chelliah, M., Ebisuzaki,W., Kanamitsu, M., Kousky, V., Dool, H., van den, Jenne, R., and Fiorino, M. TheNCEP-NCAR 50-year reanalysis: Monthly means cd-rom and documentation. Bulletin ofthe American Meteorological Society, 82:247–267, 2001.doi:10.1175/1520-0477(2001)082〈0247:tnnyrm〉2.3.co;2. → pages 4, 49Mahfouf, J.-F., Brasnett, B., and Gagnon, S. A Canadian Precipitation Analysis (CaPA)project: Description and preliminary results. Atmosphere-Ocean, 45:1–17, 2007.doi:10.3137/ao.v450101. → pages 1, 3, 4, 554Mekis, E. and Vincent, L. An overview of the second generation adjusted daily precipitationdataset for trend analysis in canada. Atmosphere-Ocean, 49:163–177, 2011.doi:10.1080/07055900.2011.583910. → pages vi, 9, 11Mesinger, F., DiMego, G., Kalnay, E., Mitchell, K., Shafran, P. C., Ebisuzaki, W., Jovic, D.,Woollen, J., Rogers, E., Berbery, E. H., Ek, M. B., Fan, Y., Grumbine, R., Higgins, W., Li,H., Lin, Y., Manikin, G., Parrish, D., and Shi, W. North American Regional Reanalysis.Bulletin of the American Meteorological Society, 87:343–360, 2006.doi:10.1175/bams-87-3-343. → pages 2, 5, 8, 9Mo, R. Radar-based Hovmller diagrams: A case study on the time-space coherence ofprecipitation in the Sea-to-Sky Corridor. Techincal report, National Laboratory for Coastaland Mountain Meteorology, Enviroment Canada, 2008. → pages 4Moore, R. D., Spittlehouse, D., Whitfield, P., and Stahl, K. Weather and climate. In Pike, R.,Redding, T., Moore, R., Winkler, R., and Bladon, K., editors, Compendium of foresthydrology and geomorphology in British Columbia, chapter 3, pages 47–84. B.C. Min. For.Range, For. Sci. Prog., Victoria, B.C. and FORREX Forum for Research and Extension inNatural Resources, Kamloops, B.C., 2010. → pages 7Moore, R. D., Woods, R. W., and Boyle, D. P. Putting PUB into practice in mountainousregions. Streamline Watershed Management Bulletin, 15:12–21, 2013. → pages 1Oke, T. R. Boundry Layer Climates. Routledge, 1987. ISBN 0415043190. → pages 2Rienecker, M. M., Suarez, M. J., Gelaro, R., Todling, R., Bacmeister, J., Liu, E., Bosilovich,M. G., Schubert, S. D., Takacs, L., Kim, G., Bloom, S., Chen, J., Collins, D., Conaty, A.,Silva, A., da, Gu, W., Joiner, J., Koster, R. D., Lucchesi, R., Molod, A., Owens, T.,Pawson, S., Pegion, P., Redder, C. R., Reichle, R., Robertson, F. R., Ruddick, A. G.,Sienkiewicz, M., and Woollen, J. MERRA: NASAs Modern-Era Retrospective Analysis forResearch and Applications. Journal of Climate, 24:3624–3648, 2011.doi:10.1175/JCLI-D-11-00015.1. → pages 46Ruane, A. C. NARR’s atmospheric water cycle components. Part I: 20-year mean and annualinteractions. Journal of Hydrometeorology, 11:1205–1219, 2010. doi:10.1175/2010JHM1193.1.→ pages 1, 5Solaiman, T. A. and Simonovic, S. P. National Centers for Environmental Prediction-NationalCenter for Atmospheric Research (NCEP-NCAR) reanalyses data for hydrologic modellingon a basin scale. Canadian Journal of Civil Engineering, 37:611–623, 2010.doi:10.1139/L10-005. → pages 5Stahl, K. and Moore, R. D. Influence of watershed glacier coverage on summer streamflow inBritish Columbia, Canada. Water Resources Research, 42, 2006. ISSN 1944-7973.doi:10.1029/2006WR005022. → pages 12Trubilowicz, J. W., Shea, J. M., Jost, G., and Moore, R. D. Suitability of North AmericanRegional Reanalysis (NARR) output for hydrologic modelling and analysis in mountainousterrain. Hydrological Processes, 30:2332–2347, 2016. doi:10.1002/hyp.10795. → pages 5Wang, T., Hamann, A., Spittlehouse, D. L., and Murdock, T. Q. ClimateWNAhigh-resolution spatial climate data for western North America. Journal of AppliedMeteorology and Climatology, 51:16–29, 2012. doi:10.1175/JAMC-D-11-043.1. → pages 455"@en ; edm:hasType "Thesis/Dissertation"@en ; vivo:dateIssued "2016-09"@en ; edm:isShownAt "10.14288/1.0308672"@en ; dcterms:language "eng"@en ; ns0:degreeDiscipline "Geography"@en ; edm:provider "Vancouver : University of British Columbia Library"@en ; dcterms:publisher "University of British Columbia"@en ; dcterms:rights "Attribution-NonCommercial-NoDerivatives 4.0 International"@* ; ns0:rightsURI "http://creativecommons.org/licenses/by-nc-nd/4.0/"@* ; ns0:scholarLevel "Graduate"@en ; dcterms:title "An evaluation of the North American Regional Reanalysis precipitation fields in a topographically complex domain, British Columbia, Canada"@en ; dcterms:type "Text"@en ; ns0:identifierURI "http://hdl.handle.net/2429/58887"@en .