Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Sea ice monitoring using spaceborne multi-polarization and polarimetric SAR imagery Scheuchl, Bernd 2006

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

Item Metadata

Download

Media
831-ubc_2006-199060.pdf [ 26.61MB ]
Metadata
JSON: 831-1.0065869.json
JSON-LD: 831-1.0065869-ld.json
RDF/XML (Pretty): 831-1.0065869-rdf.xml
RDF/JSON: 831-1.0065869-rdf.json
Turtle: 831-1.0065869-turtle.txt
N-Triples: 831-1.0065869-rdf-ntriples.txt
Original Record: 831-1.0065869-source.json
Full Text
831-1.0065869-fulltext.txt
Citation
831-1.0065869.ris

Full Text

Sea Ice Monitoring Using Spaceborne Multi-Polarization and Polarimetric SAR Imagery by Bernd Scheuchl Dipl.-Ing., Technical University Graz, Austria, 1998 A THESIS S U B M I T T E D IN P A R T I A L F U L F I L L M E N T O F T H E R E Q U I R E M E N T S F O R T H E D E G R E E O F D o c t o r o f P h i l o s o p h y in T H E F A C U L T Y O F G R A D U A T E S T U D I E S (Electrical and Computer Engineering) The University of Brit ish.Columbia A p r i l 2006 © Bernd Scheuchl, 2006 Abstract The imminent launch of RADARSAT-2 , the most advanced of the second-generation space-borne SARs, has stimulated renewed interest in polarization diversity for sea ice monitor-ing. The primary objective of this work is to assess the potential value of RADARSAT-2 multi-polarization and polarimetric C-band SAR imagery for classification of sea ice and to develop improved classifiers that account for the characteristics of such imagery. Our review of the ice information requirements of the Canadian Ice Service reveals the importance of daily revisit for operations. The need to determine the ice edge location, the ice concentration, and stage of development of ice can be addressed by accurate classification of ice types in SAR imagery. Our application of target decomposition to AIRSAR airborne polarimetric imagery of sea ice reveals that surface scattering dominates the majority of the scene. Pixel by pixel application of target decomposition methods can be used to distinguish thin ice, first-year ice, and multi-year ice with some success. When classifying sea ice, we show that a recently proposed K-means clustering algo-rithm which uses a Wishart classifier can be substantially simplified by initializing it with a seed based solely on backscatter levels. Our analysis of A I R S A R airborne polarimetric imagery of sea ice suggests that classification accuracy obtained using dual-polarization im-agery is similar to that of polarimetric imagery and better than that of single-polarization imagery. Our analysis of simulated RADARSAT-2 polarimetric imagery derived from airborne CV-580 imagery indicates that speckle noise degrades our ability to distinguish between ice types more than the increase in NESZ but can easily be reduced through spatial filtering. In simulated dual-polarization ScanSAR imagery, open water and sea ice can be easily distinguished by using both co- and.cross-polarized image in spite of the high NESZ level. In our analysis of ENVISAT A S A R A P imagery that covers a full ice season, multi-year ice could be distinguished from other types in eight out of ten scenes available. When antenna pattern correction causes a variation of NESZ over the swath, we show that adaptive classification scheme can compensate for such variation. ii Contents Abstract ii Contents - iii List of Tables viii List of Figures xi Acknowledgements xiv 1 Introduction 1 1.1 Sea Ice and the Environment 1 1.2 Ice Information 2 1.3 Use of Multi-Polarization and Polarimetric S A R Imagery for Sea Ice Monitoring . . 6 . 1.4 Motivation 7 1.5 Objectives 8 1.6 Thesis Outline 9 i i i 2 Potential of RADARSAT-2 Imagery for Operational Sea Ice Moni-toring 11 2.1 Introduction 11 2.2 Ice Information Requirements 13 2.2.1 Primary Ice Information Requirements . . 16 2.2.2 Secondary Ice Information Requirements 17 2.2.3 Iceberg Detection 18 2.3 Expected Improvements Using R A D A R S A T - 2 Imagery 19 2.3.1 Dual-Polarization Modes 20 2.3.2 Polarimetric Modes 22 2.4 Discussion 26 3 Information Content and Classification Potential of Airborne Po-larimetric SAR Imagery of Sea Ice 27 3.1 Description of the Imagery and the Ice Situation 30 3.2 Target Decomposition for Sea Ice Monitoring . 33 3.2.1 Dominant scattering Based on Freeman-Durden 33 3.2.2 H / A / a Decomposition 36 3.3 Maximum Likelihood Classification Based on the Wishart Statistics of the Covariance Matr ix 40 3.4 Evaluation of Classification Results using Limited Ground Truth . . . 43 3.4.1 Dual Frequency Classification 43 3.4.2 Classification Confidence for the Wishart Classifier 47 3.4.3 Adaptive Post-Classification Filter 49 iv 3.5 Single Frequency Classification Results 51 3.5.1 Classification of Polarimetric Imagery of Sea Ice 51 3.5.2 Classification of Dual-Polarization Imagery of Sea Ice 56 3.5.3 Classification of Single-Polarization Imagery of Sea Ice . . . . 57 3.6 Discussion of Results • • • 59 4 Evaluation of Simulated RADARSAT-2 Imagery 62 4.1 Simulation of R A D A R S A T - 2 Polarimetric Imagery 64 4.1.1 CV-580 Imagery from the Northumberland Strait . . . . . . . 64 4.1.2 Classification Result - Airborne . 67 4.1.3 Simulation of R A D A R S A T - 2 Polarimetric Imagery . . . . . . 70 4.1.4 Effect of Noise and Speckle on Classification Capability . . . . 77 4.2 Simulation of R A D A R S A T - 2 ScanSAR Imagery , 81 4.2.1 E N V I S A T A S A R Alternating Polarization Burst 82 4.2.2 R A D A R S A T - 2 ScanSAR Burst 83 4.2.3 E N V I S A T A S A R Processor Settings 84 4.2.4 Example Simulation . . 87 4.3 Discussion of Results - - 91 5 Analysis of a Full-Season ENVISAT ASAR APX Imagery 93 5.1 Resolute Test Area Data Set 95 5.1.1 A S A R A P Imagery 95 5.1.2 Auxil iary Data 95 5.2 Data Analysis 99 v 5.3 Classification Results 106 5.3.1 Standard Classification 106 5.3.2 Consideration of N E S Z Variation Over the Swath . I l l 5.4 Discussion of Results 121 6 Conclusions and Recommendations 123 6.1 Conclusions 123 6.2 Implications of Our Results 127 6.3 Recommendations for Future Work 128 Appendix A Properties of Sea Ice 130 A . l Sea Ice Stages of Development 130 A.2 Electromagnetic Properties of Sea Ice . 134 Appendix B Ice Chart Production at the Canadian Ice Service 139 Appendix C Ice Classification Using Single-Polarization SAR Imagery 142 C . l Canada 143 C.2 European Union 144 C.3 United States of America . 146 C. 4 Discussion of Current Systems 147 Appendix D Polarization Diversity in SAR Imaging 149 D . l Polarization 149 D.2 Matr ix Representation 151 D.3 Polarimetric Parameters . 153 v i D.4 Polarization Synthesis 154 D.5 Target Decomposition and Its Use for Classification 155 D.6 Classification Based on the Wishart Statistics . . . 166 D. 7 Multi-Polarization S A R Imagery 176 Appendix E Confusion Matrices 178 E . l Dual-Frequency Reference - Reference A 179 E.2 Classification Confidence and Band Restricted Dual-Frequency Refer-ence - Reference B . . . 184 E.3 Classification Confidence Restricted Dual-Frequency Reference - Reference C 188 E.4 Analyst Selected Validation Data -Reference D 192 Bibliography 196 vii List of Tables 2.1 Canadian Ice Service operational ice information requirements . . . . 14 2.2 Estimated R A D A R S A T - 2 N E S Z . : 20 3.1 C-band polarimetric parameters (i) 45 3.2 C-band polarimetric parameters (ii) . . : 45 3.3 L-band polarimetric parameters (i) 46 3.4 L-band polarimetric parameters (ii) 46 3.5 Overall classification accuracies for C-band imagery 51 4.1 Colour assignment for CV-580 classification result 67 4.2 Class average values for polarimetric parameters . . 71 4.3 Image parameters for CV-580 and R A D A R S A T - 2 S L C 72 4.4 Confusion Matr ix - 3x3 R A D A R S A T - 2 simulated imagery 77 4.5 Confusion Matr ix - filtered (n=5) R A D A R S A T - 2 simulated imagery . 78 4.6 E N V I S A T A S A R A P Mode: Number of echoes per burst 82 4.7 R A D A R S A T - 2 ScanSAR Narrow: Number of samples in matched filter 83 4.8 R A D A R S A T - 2 ScanSAR Wide: Number of samples in matched filter 83 4.9 E N V I S A T A S A R Processor Settings 85 vi i i 5.1 A S A R A P imagery availability for Resolute 96 5.2 A S A R Image Mode Swaths [34] 97 5.3 Classification evaluation (qualitative) 108 5.4 IS6 multi-temporal confusion matrices . I l l 5.5 Auxil iary data for the October 18 scene 112 5.6 Class average values for polarimetric parameters 119 A . l Sea ice stages of development 131 A.2 Dielectric constant e'r 136 E . l Validation data used in Chapter 3 178 E.2 Confusion Matr ix of C-band T P initialization (Reference A ) . . . . . 179 E.3 Confusion Matr ix of C-band Freeman-Durden initialization (Ref. A) . 179 E.4 Confusion Matr ix of L-band T P initialization (Reference A) 180 E.5 Confusion Matr ix of L-band Freeman-Durden initialization (Ref. A) . 180 E.6 Confusion Matr ix of dual channel (HH+HV) (Reference A) 181 E.7 Confusion Matrix of dual channel ( V V + V H ) (Reference A) 181 E.8 Confusion Matr ix of single channel (HH) (Reference A) 182 E.9 Confusion Matr ix of single channel (HV) (Reference A) 182 E.10 Confusion Matr ix of single channel (VV) (Reference A) 183 E . l l Confusion Matr ix of expectation based classification (Reference A) . . 183 E.12 Confusion Matr ix of C-band T P initialization (Reference B) 184 E.13 Confusion Matr ix of C-band Freeman-Durden initialization (Ref. B) . 184 E.14 Confusion Matr ix of dual channel (HH+HV) (Reference B) 185 ix E.15 Confusion Matr ix of dual channel ( V V + V H ) (Reference B) 185 E.16 Confusion Matr ix of single channel (HH) (Reference B) 186 E.17 Confusion Matr ix of single channel (HV) (Reference B) . 186 E.18 Confusion Matr ix of single channel (VV) (Reference B) . . . . . . . . 187 E . 19 Confusion Matr ix of C-band T P initialization (Reference C) . . . . . 188 E.20 Confusion Matr ix of C-band Freeman-Durden initialization (Ref. C) . 188 E.21 Confusion Matr ix of dual channel (HH+HV) (Reference C) 189 E.22 Confusion Matr ix of dual channel ( V V + V H ) (Reference C) . . . . . . 189 E.23 Confusion Matr ix of single channel (HH) (Reference C) 190 E.24 Confusion Matr ix of single channel (HV) (Reference C) . . . . . . . . 190 E.25 Confusion Matr ix of single channel ( V V ) (Reference C) . 191 E.26 Confusion Matr ix of C-band T P initialization (Reference D) 192 E.27 Confusion Matr ix of C-band Freeman-Durden initialization (Ref. D) . 192 E.28 Confusion Matr ix of dual channel (HH+HV) (Reference D) . 193 E.29 Confusion Matr ix of dual channel ( V V + V H ) (Reference D) . . . . . . 193 E.30 Confusion Matr ix of single channel (HH) (Reference D) 194 E.31 Confusion Matr ix of single channel (HV) (Reference D) 194 E.32 Confusion Matr ix of single channel (VV) (Reference D) 195 x List of Figures 1.1 Ice type examples 3 1.2 Canadian Ice Service areas of coverage 5 2.1 Example ice chart from the Canadian Ice Service 15 2.2 CV-580 C-band image from Resolute Passage 23 3.1 A I R S A R scene 1372 32 3.2 Freeman-Durden scattering mechanisms 35 3.3 C-band H / A - and H / a results , 38 3.4 L-band H / A - and H / a results . 39 3.5 Classification confidence 50 3.6 Single frequency polarimetric classification 53 3.7 Classification using expectation maximization 55 3.8 Single frequency dual-polarization classification 56 3.9 Single frequency single-polarization classification . 58 4.1 CV-580 polarimetric imagery 65 4.2 ' R G B false colour image and classification result 68 4.3 Backscatter strength histograms 69 xi 4.4 Comparison of CV580 imagery with R A D A R S A T - 2 simulations (i) . . 74 4.5 Comparison of CV580 imagery with R A D A R S A T - 2 simulations (ii) . 75 4.6 Classification of Simulated R A D A R S A T - 2 Imagery in Fine Mode . . 76 4.7 2-D scatter plots with colour coded density information 80 4.8 Simulated R A D A R S A T - 2 ScanSAR example 89 4.9 H V - H H scatter plot for simulated R A D A R S A T - 2 ScanSAR imagery . 90 5.1 Area of interest for E N V I S A T A S A R A P imagery analysis . . . . . . . 97 5.2 Temperature and snowfall records for Resolute 98 5.3 Dynamic ranges for co- and cross-polarization 99 5.4 A S A R A P M IS6 images 101 5.5 M Y I floe embedded in F Y I examples 102 5.6 M Y I and F Y I signal levels . . 103 5.7 Range profiles for the cross-polarization channel 105 5.8 Two-channel colour image and classification result (i) 107 5.9 Two-channel colour image and classification result (ii) 109 5.10 October 18 V H histogram 114 5.11 Two-step classification concept 116 5.12 V H exclusion region based on N E S Z estimate 117 5.13 October. 18 IS4 example results 118 D . l H / a and H / A planes 164 D.2 H/A/cv classification space . 166 D.3 Wishart-based classification concept. . 170 xi i D.4 Possible options for the Wishart based method 171 x i i i Acknowledgements This thesis evolved over many years and would not have been possible without the help of the following people: M y thesis supervisor, Ian Cumming, who accepted me as his student, shared his knowledge and provided guidance and encouragement. M y thesis co-supervisor, David Michelson, who offered his expertise and who challenged me to better organize my results. Ron Caves, my M D A supervisor, who initiated the project and introduced me to S A R polarimetry. Dean Flett and Roger De Abreu from CIS, who spent a considerable amount of time in support of this project and taught me a great deal about sea ice. Irena Hajnsek ( D L R ) , who offered her time for many fruitful discussions. Special thanks also go to Harold Zwick and Mark Scivier ( M D A ) , who provided me with an opportunity to work for M D A . John Yackel and Torsten Geldsetzer intro-duced me to the practical aspects of sea ice research during a field trip to Churchill. Ben Holt ( JPL) provided access to A I R S A R imagery and gave the vital advice that helped me to "finish my thesis". Helmut Rott (University of Innsbruck) changed my life by introducing me to remote sensing. I would also like to thank the many people from C S A and C C R S , who helped me at one point or another, my colleagues at M D A , and my fellow students in the radar remote sensing group at U B C , who offered their friendship. Sincere thanks go to my mother, who encouraged me to live my dream, and my wife for teaching me to always dream big. B E R N D S C H E U C H L The University of British Columbia April 2006 xiv Chapter 1 Introduction For the past decade, ice services around the world have become increasingly re-liant on spaceborne single-polarization Synthetic Aperture Radars (SAR) such as R A D A R S AT-1 as their primary information source. The pending launch of R A D A R -SAT-2, the most advanced of the second generation spaceborne SARs, has stimulated renewed interest in the use of multi-polarization and polarimetric imagery for sea ice monitoring. 1.1 Sea Ice and the Environment Oceans in and surrounding the polar regions of both hemispheres are, at least for part of the year, affected by sea ice. Ice coverage of the ocean's surface affects the interaction between the atmosphere and the water surface. The northern hemisphere contains human settlements and traffic routes that are, all year or seasonally, affected by sea ice. Thus, a significant economic component drives the need to study sea ice in the northern hemisphere in addition to the scientific reasons. In the southern hemisphere, no country is directly affected. Climatological and environmental aspects predominate the need for knowledge about the ice situation in the area. The large size of the areas affected makes remote sensing an important tool to gain' detailed knowledge about the extent, the movement and the condition of sea ice. 1 The formation of sea ice depends on a variety of environmental factors like tem-perature, wind, current, and water salinity. Sea ice undergoes a seasonal cycle and has several distinct stages of development, including: New Ice, Young Ice, First Year Ice (FYI) , Mu l t i Year Ice (MYI) , as well as deformation features like ridges and floes. Several examples of sea ice in various stages of development are shown in Figure 1.1. In Appendix A , a detailed description of the properties of sea ice, including its stages of development and its electromagnetic properties is presented. 1.2 Ice Information The World Meteorological Organization ( W M O ) is the specialized agency of the United Nations responsible for meteorology (weather and climate), operational hy-drology and related geophysical sciences. Among other responsibilities, W M O is tasked with defining and standardizing ice information [83]. Ice information is re-quired by a variety of users operating in ice-affected regions. These uses include [46]: • Marine transportation through the ice-affected regions, both to ship resources in or out of the polar regions, and as a more efficient route between ports in the mid-latitudes; • Development of resources, such as oil and gas, minerals, and fish; • Monitoring the polar regions to understand and protect the environment from both, the direct impact of development activities and the indirect impact of climate change caused by human activities elsewhere; • The design of structures and vessels to withstand ice conditions; • The day-to-day activities of northern residents. Many nations with ice areas within their national, territory have established a national ice service. A summary of Sea Ice Information Services around the world 2 (a) Frost flowers (b) Grease ice at ice edge (e) Consolidated pancake ice (f) FYI with ridges (g) Rafting of FYI (h) MYI with ridging Figure 1.1: Ice type examples. The images were taken between 1987 and 1992 in the Beaufort Sea and the Labrador Sea (see Appendix A for a detailed description of the properties of sea ice). Images used with permission from Ben Holt, N A S A JPL/Ca l t ech (http://polar.jpl.nasa.gov/polar/). 3 is given in [82]. While all ice services work relatively independently in providing ice information for their area of interest, there is co-operation between nations that share boundaries (e.g., the Baltic nations; Canada and U S A ) . Internationally, the Interna-tional Ice Charting Working Group (IICWG) is an effort to promote cooperation in the ice information sector [47]. In Canada, ice reconnaissance dates back to the 1920's when aircraft of the Royal Canadian A i r Force first flew over the ice covered Arctic. Starting in the 1960's, forward looking, non-imaging sweep radars were installed on reconnaissance aircraft. In 1972, the first side looking airborne radars (SLAR) were put in operation. In 1989, Canada introduced a routine ice reconnaissance service based upon airborne S A R . Aircraft continued to be the main data source until 1996, when R A D A R S A T - 1 S A R imagery became operational. N O A A A V H R R and D M S P S S M / I data are also utilized operationally. In 1959, responsibility for ice reconnaissance was transferred from the A i r Force to the Meteorological Branch of the Department of Transport. In 1971, the Mete-orological Branch became the Atmospheric Environment Service of the Department of the Environment [11]. A t present, the Canadian Ice Service (CIS) / Ice and M a -rine Services Branch of the Meteorological Service of Canada (MSC) is the leading authority for information about ice in Canada's navigable waters [40]. The areas of coverage are shown in Figure 1.2. The main objectives of the CIS are [40]: 1. To ensure the safety of Canadians, their property and their environment by warning them of hazardous ice conditions in Canadian territorial waters; and 2. To provide present and future generations of Canadians with sufficient knowl-edge about their ice environment to support sound environmental policies. To meet these objectives, the CIS collects and analyzes vast amounts of sea ice imagery in all Canadian regions affected. In Summer, the focus is on the Arctic and the Hudson Bay regions. In Winter and Spring, attention broadens to include the Labrador Coast and East Newfoundland waters, the Gulf of St. Lawrence, the Great 4 Figure 1.2: Canadian Ice Service areas of coverage. The main focus of ice chart production varies seasonally. Lakes, and the St. Lawrence River [40]. One of the products prepared by the CIS is the Daily Ice Analysis Chart. These charts provide information on ice edge, ice concentration, stage of development, as well as form of the ice or floe size in the area [112]. Since R A D A R S A T - 1 became operational in 1996, it became the primary data source for CIS. At present, human experts play a key role in data analysis for ice centers around he world. As the amount of remote sensing data increases, ice services must carefully consider what data should be used and processed. Automation of some tasks wil l enable an analyst to use additional information without increasing his or her workload. In Appendix B , the ice chart production process at the CIS is described. In Appendix C, a summary of both past and current research efforts on sea ice classification using single-polarization S A R imagery is presented. 5 1.3 Use of Multi-Polarization and Polarimetric SAR Imagery for Sea Ice Monitoring Polarization diversity has shown to increase the information content of S A R imagery significantly [12]. The benefit for sea ice monitoring was recognized early [28], but the amount of polarimetric S A R imagery of sea ice available to researchers remains limited. The first polarimetric sea ice image acquisitions were made in the late 1980's using airborne SARs. Ice missions have since been flown by the N A S A J P L A I R S A R , the Danish E M I S A R , the Canadian CV-580, and more recently, the Japanese P i - S A R . Two short SIR-C missions, flown in. 1994 with an orbit covering latitudes up to ± 60 degrees, were the first to acquire polarimetric spaceborne S A R imagery over sea ice. E N V I S A T A S A R , launched in early 2002, allows the collection of S A R imagery in two polarizations. To date, no repeat pass polarimetric sea ice imagery has ever been collected (at least to the best of our knowledge). Research on polarimetric signature evolution is based on scatterometer time series and laboratory experiments [79]. Many of the early studies of polarimetric S A R sea ice imagery focused on the physics of the scattering from sea ice and the resulting polarimetric signatures of ice types and their evolution as opposed to classification of ice types [29], [33]. While some authors have presented studies where polarimetric sea ice classifica-tion was mentioned in passing, the most detailed study to date on machine classifica-tion was presented in 1994 by Rignot and Drinkwater [92]. They used a Maximum A Posteriori ( M A P ) classifier. A I R S A R L- and C-band sea ice imagery in a variety of polarization and frequency combinations are investigated. Dual-frequency approaches are shown to improve the classification result. In 1995, Hara et al. used a Neural Network classifier on polarimetric sea dee imagery [50]. In 2003, Dierking et al. devel-oped a classifier for polarimetric imagery of sea ice, which is based on a hierarchical, 6 knowledge-based approach. Decision boundaries at individual levels in the hierar-chy can be determined using statistical methods. The approach can be "optimally, adapted" to a particular region and season [26]. In 2004, the same authors presented a comparative study between C- and L-band. They found no clear preference for a particular frequency band with the information and imagery available [27].. A l l second-generation spaceborne SARs have the capability to acquire imagery in more than one combination of transmit and receive polarizations. The very first ex-amples were two short SIR-C missions with dual-polarization and polarimetric acqui-sitions. The European E N V I S A T , launched in 2002, has an alternating polarization acquisition mode. The Japanese P A L S A R (launched on January 24, 2006), the Ger-man TerraSAR-X, and the Canadian R A D A R S A T - 2 wil l all have dual-polarization and polarimetric acquisition capabilities. 1.4 Motivation W i t h the launch of second-generation spaceborne SARs, interest in the utility of dual-polarization and polarimetric imagery for sea ice monitoring has increased. R A D A R S A T - 2 wil l have the most impact for operational sea ice monitoring, as it will.offer the entire suite of legacy beams of R A D A R S A T - 1 in addition to a number . of enhancements such as dual-polarization S A R [1]. In 1993, CIS reviewed their ice information requirements during preparation for the launch of R A D A R S A T - 1 [88]. In light of the success of the R A D A R S A T - 1 program for sea ice monitoring and the technical improvements for the R A D A R S A T - 2 program, an update of the ice information requirements is warranted. . Following the detailed analysis of the few available polarimetric images of sea ice, the focus shifted to operational use of single-polarization imagery in the mid 1990's. A number of new analysis methods for polarimetric imagery have been developed since then. Most notably, target decomposition algorithms were formulated for distributed targets, which allow an interpretation of scattering mechanisms [21], [44]. In addition, 7 new concepts for a K-means clustering algorithm which uses a Wishart classifier were developed [60]. We presented the first results obtained from target decomposition for sea ice monitoring in 2001 [97] and continued research in the field [106], [107] [108], [104] and [103]. The main conclusions of our research are presented in this thesis. Other authors using these methods reported their results later (2002, 2003 and 2005 respectively) [22], [93] and [75]. The main interest in most previous studies was the information content of the imagery available. Little attention was paid to operational constraints. The revisit time of polarimetric modes of real systems and its implications for operations were not addressed. Power constraints on spaceborne systems result in a higher system noise level (noise equivalent sigma zero (NESZ)) compared to airborne systems. Backscat-ter from sea ice is often lower than that from land with cross-polarized backscatter about 10 dB below co-polarized backscatter. A n increased N E S Z and the correspond-ing reduction in S N R wil l therefore affect the cross-polarized channel most severely. Antenna pattern correction will cause a variation of the S N R over the swath and affect data analysis. The availability of E N V I S A T A S A R A P imagery allows the analysis of the information content of multi-polarization imagery under various environmental conditions. This wil l provide a more complete picture on the potential of polarization diversity for operational sea ice monitoring than can be given by analyzing a single airborne scene. 1.5 Objectives The primary objective of this work is to assess the potential value of R A D A R S A T -2 multi-polarization and polarimetric C-band S A R imagery for classification of sea ice and to develop improved classifiers that account for the characteristics of such imagery. The specific objectives are: 1. To identify and summarize current thinking regarding the potential of R A D A R -SAT-2 imagery for operational sea ice monitoring (Chapter 2). 8 2. To evaluate the utility of the Freeman-Durden and Cloude-Pottier target de-composition methods for separating ice types based on scattering mechanisms (Chapter 3). 3. To assess the performance of a K-means clustering, algorithm which uses a Wishart classifier, including recently developed variants, for sea ice monitor-ing and to suggest possible improvements or simplifications (Chapter 3). 4. To evaluate the extent to which reduction of resolution, increase in N E S Z , and difference in number of looks affects the ability to differentiate between ice types using suitably processed CV-580 and E N V I S A T A S A R imagery (Chapter 4). 5. To further evaluate the utility of spaceborne multi-polarization S A R imagery for sea ice monitoring by identifying M Y I by means of classification using S A R imagery that covers a range of environmental conditions (Chapter 5). 6. To propose and evaluate a method for compensating for the effect of N E S Z vari-ation over the swath and thereby improve classification accuracy (Chapter 5). 1.6 Thesis Outline The remainder of this document is organized as follows: • In Chapter 2, we provide a comprehensive ice information requirements review, the first of its kind since pre-launch preparations for R A D A R S A T - 1 in the early 1990's and summarize the current thinking of the potential of R A D A R S A T - 2 imagery for operational sea ice monitoring. • In Chapter 3, we evaluate the classification performance of Freeman-Durden and Cloude-Pottier target decomposition methods and a K-means clustering algorithm which uses a Wishart classifier, when applied to sea ice. Our trials are conducted using airborne imagery of sea ice collected by N A S A / J P L over the Beaufort Sea in 1988. Based on this assessment, we propose a simpler method for initializing the K-means clustering algorithm which uses a Wishart classifier that accounts for the unique properties of sea ice. We also assess classifier performance on single- and dual-polarization s well as polarimetric imagery. In Chapter 4, we determine the degree to which different levels of multilooking and differences in N E S Z determine the difference in the classification potential of airborne and spaceborne polarimetric S A R imagery. We do so by comparing classification results and cluster separation for both airborne and simulated spaceborne imagery. We assess the value of spatial filtering to compensate for the different levels of multilooking applied to the two images. We also assess how well sea ice and open water can be distinguished in R A D A R S A T - 2 dual-polarization ScanSAR imagery. In Chapter 5, we present ice type classification results for the first full-season E N V I S A T A S A R A P imagery and seek to identify multi-year ice in a variety of environmental conditions. We assess the affect of N E S Z variations over the swath of spaceborne dual-polarization acquisition modes on classifier perfor-mance, and propose a method for compensating for such variations. In Chapter 6, we draw conclusions, discuss implementations and give recom-mendations for future work. In the Appendices, we provide the following information: — In Appendix A , we discuss properties of sea ice with relevance to microwave remote sensing. — In Appendix B , we describe the ice chart production process at the CIS. — In Appendix C, we summarize both past and current research on classifi-cation of single-polarization S A R imagery of sea. — In Appendix D, we discuss polarization diversity in S A R imaging with an emphasis on target decomposition and classification: — In Appendix E , we provide confusion matrices for the analysis in Chapter 3. 1 0 Chapter 2 Potential of RADARSAT-2 Imagery for Operational Sea Ice Monitoring1 2.1 Introduction R A D A R S A T - 1 has been a primary source of S A R imagery for sea ice monitoring since it became operational in 1996. This has revolutionized the way in which several na-tional ice monitoring agencies now operate. For example, the United States National Ice Center's (NIC) use of R A D A R S A T - 1 imagery has approached 6000 scenes annu-ally to meet their global ice monitoring objectives, while the Canadian Ice Service (CIS) has consumed on average between 3500-4500 scenes annually covering Canadian waters. W i t h the future launch of R A D A R S A T - 2 , the CIS is looking forward to conti-nuity of R A D A R S A T - 1 imagery to meet its operational needs. R A D A R S A T - 2 wil l offer the entire suite of legacy beams of R A D A R S A T - 1 . Besides data continuity, R A D A R S A T - 2 promises several enhancements to the payload and the space and 1 A version of this chapter has been published in [104]. Used with the permission . of the Canadian Aeronautics and Space Institute. 11 ground segments [1]. The potential benefits of a number of these enhancements for operational ice monitoring are summarized'in [87]. The space segment wil l of-fer higher resolution, multi-polarization, and polarimetric modes. The capability of R A D A R S A T - 2 to measure two or four polarizations respectively is expected to en-hance the measurement of parameters (e.g., ice edge location, ice concentration, stage of development), which are important for operational ice monitoring. The focus of this chapter is on the potential of some of these modes for meeting operational ice information requirements, with an emphasis on dual-polarization modes. Previous studies of the potential of multi-polarization and polarimetric S A R imagery for sea ice monitoring have been limited to a small number of data sets. The most authoritative studies are based on polarimetric imagery acquired by the Jet Propulsion Laboratory (JPL) A I R S A R airborne system over the Beaufort and Bering Seas during 1988 along with extensive ground truth [92], [78], and [91]. Several other studies have been conducted using imagery from the Danish E M I S A R [115] and [117] as well as the Canadian CV-580 S A R system [65], [108], [105], and [109]. The SIR-C mission in 1994 was the first to acquire polarimetric spaceborne S A R imagery over sea ice. These images have been studied in [33], [97], and [98]. In 2002, E N V I S A T -A S A R came into operation. First results using its alternating polarization imagery is reported in [4] and [104]. These results are not a fully representative indication of R A D A R S A T - 2 perfor-mance, however, because of the differing noise level, incidence angle, resolution and coverage area of these sensors. These differences are more pronounced in the case of airborne SARs, which generally have finer resolution and lower noise levels, but cover less area than spaceborne SARs . In this chapter, we first identify the key ice information requirements for opera-tions at the CIS. Second, we briefly review the current use of R A D A R S A T - 1 single-polarization imagery as the primary information source. Expected improvements for sea ice monitoring using R A D A R S A T - 2 are discussed, emphasizing the advantages of dual-polarization and polarimetric imagery. 12 2.2 Ice Information Requirements Ice information requirements span a variety of temporal and spatial scales and a range of geographical zones and seasons. They vary with the specific application and situation. A summary of sea ice information requirements from the operational perspective in the context of R A D A R S A T - 1 is given in [88], while [9] summarizes ice information requirements from a science viewpoint. The science and operational requirements overlap to a large extent, differing primarily in the spatial and temporal scales. In this section, the focus is primarily on the operational ice information requirements, as identified by the CIS. Sea ice information requirements for operations and navigation are often cate-gorized into two types, strategic and tactical, which differ in their scale and time-liness [88], [52]. Table 2.1 summarizes the parameters that are important for both types, which depend on the particular situation or application. For the purposes of this table, Strategic refers to the level of detail and information required for the preparation of a Canadian Ice Service Daily Ice Analysis Chart, an example of which is given in Figure 2.1. Tactical refers to the level of detail and information required to support daily operations and ship navigation in ice. As an example of the use of Strategic and Tactical levels of information, the CIS deploys Ice Service Specialist (ISS) field personnel onboard the Canadian Coast Guard icebreakers. The ISS can use the Daily Ice Analysis Chart to assist the Captain in preliminary route planning. In contrast, tactical information of finer detail is collected in the vicinity of the vessel, e.g., by helicopters, to support immediate operations. Some airborne and spaceborne imagery potentially meet tactical requirements and can be used if available. In ad-dition to information on sea ice, the detection and tracking of icebergs is part of the CIS mandate and is included as a separate requirement. Daily coverage of an area is important for many ice information requirements, in particular all primary requirements as outlined in Table 2.1. Generally, wide swath imaging is required o achieve near daily coverage. Such modes are provided 13 P r i m a r y Secondary Iceberg Ice in fo rmat ion requ i rement Ice Edge Loca t ion Ice concentra-t ion Stage of develo-pment (e.g. new, th in , f i rs t -year , and mul t i -year ice) Presence and location of leads (open water) Ice thickness Ice topography and roughness Ice decay state (or, more speci f ical ly , Ice strength) Snow propert ies (e.g. thickness; also wetness, density) Iceberg detect ion Spat ia l resolut ion strategic 5km +/- 10% requires resolution < 100 m 50-100 m 50-100 m similar to ice concentrati on 5 km for average thickness over an area to +/-20% of total thickness < 50 m to determine extent of ridging: need average ridge heights to within +/- 20 % 20 km for average strength over an area to +/-20% of total strength 5 km for average snow depth to +/- 20 % <50m tactical <] km +/- 5 % requires resolution <25 m <20m <20m < 100 m to determine average and maximum . thickness (including rafting) over an area to +/-10% of total thickness < 10 m to determine mean and maximum ridge heights to within+/- 10% 5 km for average strength over an area to +/-10% of total strength 1 km for average snow depth to+/- 10% <5 m T e m p o r a l strategic daily daily daily daily 2x/week daily weekly weekly daily resolut ion tactical 6 hours 6 hours 6 hours 6 hours daily 6 hours daily daily hourly Descr ip t ion In case of a diffuse ice edge, the CIS defines the ice/water boundry as 10% ice concentration Percentage of ice covered area.Ice concentratio n is a key parameter in the W M O egg code (MSC, 2002 ). Ice classes are defined by W M O ( M S C , 2002). Leads and polynyas Important for navigation and/or loads on structures (Remark: Rafts show double or more the undeformed floe thickness) Presence, location and height of ridges Roughness indicated by concentration of ridges (in % or in number per unit area) - ridge density plus average height (or total thickness) of ridges Identification of melt onset and ponding To determine hull friction for ship resistance -also •• important for ice strength Detection and tracking, possibly classificatio n of type ( M S C , 2002). Figure 2.1: Example of a Canadian Ice Service Daily Ice Analysis Chart, which covers approximately 900000 km2. Ice information is provided by the colour in the chart, and more detailed information is given in the egg codes. Colour and egg codes are described in detail in the "Manual of standard procedures for observing and reporting ice conditions" [112]. 15 by R A D A R S A T ScanSAR modes, whose spatial resolution of 50 to 100 m satisfies strategic, as well as many tactical requirements. ScanSAR Wide is more suited for regular wide-area monitoring, while ScanSAR Narrow is preferable for monitoring areas requiring a greater level of detail. Single-beam modes can be used for monitoring tasks requiring higher resolutions, for example, monitoring busy shipping routes and iceberg detection. The tradeoffs for increased detail are reduced area coverage and less frequent revisit times. The wider swath, single-beam modes (wide and standard) are preferred over the fine mode to optimize coverage. In terms of incidence angles (e.g., S3-S7, W2-W3) are more suited for detecting the ice edge and icebergs with co-polarization, particularly in rough seas. However, time-critical considerations wil l limit the incidence angle choice for an area of interest. 2.2.1 Primary Ice Information Requirements The first and primary task for sea ice monitoring is to identify the location of the boundary between ice and open water (the ice edge), with sufficient spatial coverage and temporal repetition. The backscatter contrast between the ice and open water determines the capability to accurately define the ice edge. For co-polarized imagery, the contrast varies greatly depending on incidence angle, ice type, sea state, and wind conditions. In relation to incidence angle, the ocean clutter level is highest at steeper angles. Increased surface roughness due to wind and waves wil l also increase the ocean clutter level. The most problematic situation for ice-edge delineation is when the ocean clutter conditions result in backscatter conditions similar to sea ice signatures. Accurate estimation of ice concentration is a derivative of the ability to discriminate between ice and open water, but requires finer spatial resolution. The ability to discriminate different ice stages of development is very dependent on the geographic region and season [88]. Based on CIS experience, H H can be used to discriminate between the two major categories of multi-year and first-year ice in 16 cold winter conditions [39]. Backscatter over multi-year ice is increased due to volume scattering. Under wet conditions, volume and surface scattering from overlying snow dominates the backscattered signal, thus masking the contrasts between the under-lying ice stages of development. Discrimination of ice stages of development is also aided by the identification of large-scale ice structures (e.g., floes) and deformation features (e.g., ridges, fractures, etc.). The separation of thin and new ice from open water and smooth landfast ice is a major problem with single-polarization SARs. 2.2.2 Secondary Ice Information Requirements Few of the secondary ice information requirements are met using existing imaging systems. Visual reconnaissance and modelling tools are used where possible, however, to provide some of the required information. The ability to reliably separate sea ice from open water also allows the detection of leads, which are openings in the ice. Aside from their thermodynamic relevance, leads are important for route planning and navigation. Presently, ice thickness cannot be measured directly using single-polarization S A R . Rather, it is inferred from analysis and interpretation of ice stages of develop-ment (e.g., grease, nilas, grey, grey-white, first-year, multi-year [112]), which are a proxy for ranges of ice thickness. Ice topography, structure, and ice deformation features, such as ridging, are im-portant parameters as these features pose a significant hazard and impediment to navigation. Several researchers [84], [74], [55], and [114] have examined the potential of using S A R imagery for identifying deformed ice and characterizing ice topography. The authors of [56] assessed the potential of using ScanSAR imagery for ridge detec-tion. Results for ridge detection were not very promising, due to the coarser resolution (100 m) relative to the lateral and longitudinal ridge dimensions and sail heights (3.5 to 4 m) observed in the study. While ScanSAR imagery can be used successfully to identify areas of deformed ice, it was concluded that identification and measurement 17 of specific ridges is not feasible. Higher resolution modes (e.g., Fine, Standard) wil l be more successful at ridge identification, but would not support daily revisits. In addition, modes with higher incidence angles approaching those of airborne SARs are preferable [74]. Currently, the CIS is exploring the use of R A D A R S AT-1 imagery, in combination with other sources of information, for assessing sea ice spring melt. The goal is to reliably identify and monitor the state of ice decay, and ultimately infer ice strength. Part of this relies on the ability to accurately determine when ponding begins on the ice surface and when the water begins to drain though the ice. The drainage is a precursor to fracture and breakup, drainage is also an indication of weakening or decay. ScanSAR imagery have proven effective at identifying "Winter" and "Snow Melt onset" stages of the sea ice evolution. However, the "Ponding" stage is more susceptible to misidentification due to the high sensitivity of water backscatter to wind-induced roughness over much of the ScanSAR swath [2]. Information on snow cover is important due to its effects on the decay and breakup of sea ice, and the associated impact on marine navigation. Snow plays an important role in the thermodynamic evolution of the snow and sea ice cover. The authors of [10] note that understanding the relationship between the thermodynamic state of the snow/ice system and S A R backscatter wil l lead to information on snow thickness classes over thick first-year ice. 2.2.3 Iceberg Detection The radar backscatter from an iceberg arises from several mechanisms, mostly surface and volume scattering [51]. Multi-bounce scattering is also a contributing factor, depending on the shape of the iceberg. A n overview of different iceberg shapes is given in [86]. As a result of these scattering mechanisms, icebergs, like ships, often manifest themselves in C-band radar images as bright point targets with intensities above the ocean backscatter (clutter). Iceberg validation studies on the detection 18 limits of R A D A R S A T - 1 indicate a high probability of detection for icebergs of size on the order of the resolution cell of the selected beam mode [86]. High wind conditions as well as high sea states limit iceberg detection, due to increased sea clutter. Icebergs larger than the resolution cell can be detected more reliably, even in rough sea states. Wide 3 is the recommended mode for operational use [14] to balance detection and discrimination capability with coverage (150 km swath). For this mode and resolution (nominally 30m), it has been shown that small, medium, and large class icebergs can be detected with a probability of at least 0.7 under wind conditions up to 20 knots [14]. Some discrimination of ships from icebergs is currently possible with single-channel systems, primarily using higher resolution (e.g., Fine) beam modes [13]. Airborne radars are frequently used for operational iceberg detection and tracking. 2.3 Expected Improvements Using R A D A R S A T - 2 Imagery Based on a review of the literature, observations from practical operational experience, and experiments with E N V I S A T - A S A R imagery, a number of areas can be identified where the capabilities of R A D A R S A T - 2 are expected to provide improvements. These improvements are discussed in this section with emphasis on dual- polarization and polarimetric imagery. The estimation of polarimetric parameters is affected by the strength of the signal relative to the system noise level, as defined by the noise-equivalent sigma zero (NESZ). There is a variation of up to 2.5 dB in the N E S Z of the different R A D A R S A T -2 standard beams, but no marked trend with look angle of the beam. The higher N E S Z for R A D A R S A T - 2 as compared to several airborne sensors can result in a low S N R for areas with low backscatter (e.g., calm water). This is particularly important for analysis involving the cross-polarized backscatter, which is generally lower than the backscatter of the two co-polarized channels. Table 2.2 summarizes the estimated R A D A R S A T - 2 N E S Z levels. 19 Table 2.2: Estimated performance values (dB) for R A D A R S A T - 2 noise-equivalent sigma zero (NESZ) . - ' Mode Standard Fine Wide ScanSAR ScanSAR Narrow Wide Polarization polarimetric polarimetric dual-pol dual-pol dual-pol N E S Z [dB] - 3 1 - 2 8 - 2 3 - 2 3 - 2 3 2.3.1 Dual-Polarization Modes The R A D A R S A T - 2 dual-polarization options include H H / H V or V V / V H modes. The additional information provided by the cross-pol channel can be very useful, as the cross-pol channel responds to different scattering mechanisms than the co-pol chan-nel. This dual-polarization imagery is available in all beam configurations, giving a wide choice in resolution, coverage, and incidence angle. Most importantly, dual-polarization wil l be available for ScanSAR modes, the modes most used at the CIS for R A D A R S A T - 1 imagery. Operational experience by CIS analysts based on several thousands of images over the last 10 years results in a preference for H H polariza-tion over V V . The increased sea clutter from the VV-polarization results in more interpretation confusion across the entire incidence angle range. A I R S A R imagery of the Beaufort Sea was analyzed with emphasis on the simula-tion and validation of E N V I S A T beam modes [76]. The authors of the study suggest the use of different dual-polarization combinations depending on incidence angle and wind speed (if open ocean/water is present). They conclude that dual-polarization wide swath modes, such as R A D A R S A T - 2 ScanSAR, wil l give better results for sea ice monitoring compared to single-polarization S A R . 20 Primary ice information requirements The H V backscatter response from water is generally low and is relatively indepen-dent of wind induced surface roughness conditions, whereas H V backscatter from sea ice is affected by surface roughness, volume scattering, and multi-bounce scattering. Thus, at steeper incidence angles, the ice-ocean contrast of H V can be expected to be greater than for either of the co-polarization channels, especially at high wind condi-tions. This is confirmed by results a study [78], which finds that surface scattering is dominant up to 30° incidence and volume scattering is dominant above this value. The combined use of co- and cross-pol channel gives better results across a wider range of incidence angles. Improved separation between multi-year ice and first-year ice has been demon-strated for H V [81], [38]. Thus, cross-polarization ScanSAR imagery would be advan-tageous if discrimination between these two primary stages of development is a high priority. Although there appears to be some incremental benefit with the availabil-ity of cross-polarization ScanSAR imagery for multi-year/first-year discrimination, use of cross-polarization alone for thin ice detection is not recommended as intensity differences due to surface scatter are suppressed [81]. Secondary ice information requirements Cross-polarization imagery can enhance the structural information of sea ice, and has demonstrated some utility for improving discrimination between smooth and deformed ice. This is a function of the combined volume scattering and multiple-reflection surface scattering in the ice ridges enhancing the cross-polarization radar returns [51]. Observations of C-band scatterometer measurements of Baltic sea ice [72] quantitatively illustrated that the backscatter contrast between level ice and ice ridges is larger at cross-polarization than co-polarization. Figure 2.2 illustrates this contrast with airborne C-band H H and H V imagery for cold winter ice conditions [38]. As a result, cross-polarization ScanSAR imagery from R A D A R S A T - 2 is expected to be an 21 improvement over the current co-polarization case for better detection of ice topogra-phy and structure. However, there wil l still be limitations as a function of resolution and incidence angle as noted earlier. Operationally, the effects discussed above can improve the detection of hazards to navigation in ice, specifically traces of M Y I in a matrix of F Y I and the detection of pressure ridges. However, both of these features are small compared to the resolution of the wide-swath radar modes generally used, so their detectability wil l likely remain resolution limited. Melt conditions wil l also limit the detectability of roughness fea-tures due to the masking of the ice surface as a result of the free water on the surface and/or in the snow pack. Cross-polarization imagery wil l assist in resolving some of the ambiguity in de-tecting the presence or absence of surface water on decaying ice now encountered with co-polarization imagery [2]. Similar to the case for open water vs ice discrimination, the cross-polarization imagery wil l assist in identifying melt ponds on the ice surface, particularly under wind-roughened conditions. Iceberg detection The reduced sensitivity of H V to sea state as compared to H H leads to the conclusion that cross-polarization imagery wil l be advantageous for detecting icebergs, particu-larly at steeper incidence angles below about 35° and in high sea states. The limiting factors for detection are radar resolution, iceberg size, sea state, and signal to clut-ter/noise ratio. R A D A R S A T - 2 wil l offer a high-resolution single-polarization mode (Ultra-Fine, H V not available), which is expected to enhance iceberg detection and classification capabilities but. only over 20 km swaths. 2.3.2 Polarimetric Modes The polarimetric mode of R A D A R S A T - 2 is expected to provide additional informa-tion compared to dual- and single-polarization imagery. However, the narrow coverage 22 •-look (a) HH 30° 37° (b) HV Figure 2.2: CV-580 C-band HH-polarization image (left) and HV-polarization image (right) acquired in May 1993 in Resolute Passage. The image dimensions are approx-imately 1 km in range and 2.5 km in azimuth, and the incidence angle range is about 30° — 37°. Data analysis was performed on 4 pixels x 4 pixels. The left hand two thirds of each image is first-year ice, with ridging and deformation throughout. The brighter fracture at the right hand side of each image is a multi-year ice floe. The cross-polarization image shows much greater ice surface topography and structural detail within the first-year ice than the co-polarization image. 23 (25 km swath) will limit its utility for operational sea ice monitoring. The potential for repeat acquisition of polarimetric imagery over the same area throughout an ice season wil l lead to an increased understanding of sea ice signatures and improved techniques for extracting information from polarimetric imagery. P r i m a r y ice i n f o r m a t i o n requi rements Using SIR-C imagery acquired near the Newfoundland Coast in A p r i l 1995, the im-proved ice/water separation potential of single-channel cross-polarization imagery was characterized and similar potential using the co-polarization channel ratio was deter-mined [97], [98]. The authors also found some utility in the anisotropy parameter for ice/water classification. Generalization of the results is difficult, as the imagery used in this study were acquired during melting conditions. Estimation of the ice stage of development wil l benefit from the availability of polarimetric imagery. Additional phase information embedded in the correlations between the different polarizations is important information for ice typing and clas-sification [29], [26]. The use of one co-polarization channel with the co-polarization ratio is suggested in [33]. In R A D A R S A T - 2 , this parameter is available only in the po-larimetric modes. The authors of [117] examine C-band polarimetric sea ice imagery in the Greenland Sea and note that no single polarimetric parameter can adequately discriminate between all ice types. However, they find that by combining various polarimetric parameters, better classification can be achieved. Detailed information of the scattering mechanism symmetries can be used to separate frazil and congela-tion ice in newly formed leads and open water. The co-polarization ratio is useful for separating multi-year ice from rough thin ice and open water. The ability to separate thin ice from open water is limited at C-band but it is better at longer wavelengths, such as L - and P-band [29]. Automation of classification is presented in [92], [50], [97], and [98]. 24 Secondary ice information requirements Polarimetric S A R imagery shows some potential for ice thickness measurement. Ice thickness up to 50 cm using E band was measured in [125]. C-band imagery was used to measure thin ice thickness [59], time series polarimetric C-band data was used in [113]. More recent studies on ice thickness using L-band imagery (in particular the V V - t o - H H ratio) were conducted in Japan [124], [75]. There might be some potential for snow cover information extraction using time series of polarimetric C-band imagery from R A D A R S A T - 2 coupled with a thermody-namic model [6]. Iceberg detection It is expected that polarimetric radar systems wil l further facilitate the discrimination of icebergs from ships and for iceberg detection in pack ice, which presently not possible with R A D A R S A T - 1 . The improvements in ship detection with polarimetric S A R (e.g., [119], [120]) provides some guidance for icebergs, which have similar (but not identical) scattering response. 25 2.4 Discussion In this chapter, we have provided a comprehensive ice information requirements re-view, the first of its kind since pre-launch preparations for R A D A R S A T - 1 in the early 1990's. We have also reviewed previous research concerning sea ice monitoring using polarization diversity in order to evaluate the potential of R A D A R S A T - 2 for operational sea ice monitoring. Our findings are as follows. First, the Canadian Ice Services' primary requirement is for information con-cerning ice edge location, ice concentration and ice stage of development. This is information provided on ice charts in form of egg codes. B y accurately classifying ice types in S A R imagery, all three can be addressed. However, the backscatter contrast between ice and open water is highly variable in single-polarization imagery. This contrast depends on incidence angle, ice type, sea state, and wind conditions thus making image interpretation difficult and potentially inconsistent. However, previous studies suggest that the use of polarization diversity wil l improve the separability of ice and open water. Second, daily revisit is the single most important criterion for using S A R imagery and classification maps derived from such imagery in Canadian Ice Service operations. Wide coverage modes are therefore the modes of choice for operations. Polarimetric modes offered by R A D A R S A T - 2 and all other second-generation spaceborne SARs have narrow swaths. The operational capability of polarimetric modes is therefore limited. R A D A R S A T - 2 , however, offers a dual-polarization ScanSAR mode that provides wide coverage. If it can be shown that dual-polarization imagery can offer acceptable classification accuracy, R A D A R S A T - 2 wil l be the only spaceborne S A R providing multi-polarization imagery suitable for operational sea-ice monitoring 26 Chapter 3 Information Content and Classification Potential of Airborne Polarimetric SAR Imagery of Sea Ice2 A key difference between the first-generation spaceborne S A R technology currently used for sea ice monitoring and the second-generation spaceborne SARs that are in the process of being deployed is the availability of polarimetric data acquisition modes. Such imagery contains information concerning the targets within a scene that is not available in single-polarization imagery. The availability of polarimetric imagery permits the use of a variety of data analysis methods that are not available for single-polarization imagery. In this chapter, we assess the potential for distinguishing between different types of sea ice using either: (1) various methods for target decomposition based upon particular scattering mechanisms or (2) a K-means clustering algorithm which uses a Wishart classifier. Based on this assessment, we develop an improved classifier 2 Several results presented in this chapter were published in [103]. Used with the permission of the Canadian Aeronautics and Space Institute. 27 that accounts for the unique characteristics of sea ice imagery. The imagery that we have chosen is airborne polarimetric S A R imagery of sea ice that was acquired by the N A S A / J P L A I R S A R over the Beaufort Sea in 1988. Almost twenty years later, it remains one of the best documented sets of such imagery that is generally available. Since the use of S A R polarimetry for sea ice monitoring was first studied in the early 1990's, polarimetric theory has evolved. Polarimetric target decomposition methods which allow one to interpret the mechanisms which contribute to scattering from distributed targets were developed by Freeman and Durden [44] and Cloude and Pottier [21], as described in Appendix D. Target decomposition also led to new concepts for the use of the K-means clustering algorithm which uses a Wishart classifier. Such classification methods show promising results for scenes containing urban areas, forest, and agricultural fields [60] but their utility in distinguishing between different types of sea ice was unproven until we began this study [97], [103]. Since then, other authors have helped to carry this work forward [93], [75]. One challenge when evaluating alternative methods for sea ice monitoring is the lack of reliable, accurate ground truth. This is due to the vastness of the regions affected, logistical issues in remote regions and mobility aspects when moving on the ice. For sea ice research, "ground truth" usually refers to an interpretation of a scene (and existing ancillary data) by an expert rather than information collected by re-searchers on the ground. This is different from agricultural or forestry applications, where fields or growth stands can generally be accurately delineated and a variety of measurements can relatively easily be collected. Estimating the classification confi-dence allows us to assess classification results when ground truth is not available. In general, the classification of an image is a process implying the identification of the different thematic classes present in it and their connection to some specific ground cover types [89]. In this thesis we attempt to distinguish ice types using polarimetric and multi-polarization S A R imagery. Our final result is a set of classes that can be interpreted as ice types. We wil l therefore use the terms "classification" and "classifier", even if the class interpretation is a manual process. 28 The specific objectives of this chapter are: 1. To evaluate the utility of the Freeman-Durden and Cloude-Pottier target de-composition methods for separating ice types based on scattering mechanisms. 2. To develop an improved classifier that compensates for the characteristics of sea ice imagery and to compare its performance with recently developed variants of the K-means clustering algorithm which uses a Wishart classifier initialized with target decomposition methods. 3. To suggest methods for assessing classification results when ground truth is uncertain. 4. To evaluate the classification potential of polarimetric and multi-polarization imagery of sea ice. The remainder of this chapter develops as follows. • In Section 3.1, we present the imagery that we use and describe the ice situation. • In Section 3.2, we evaluate the suitability of target decomposition methods to distinguish between ice types. • In Section 3.3, we present the K-means clustering algorithm which uses a Wishart classifier and propose a new variant for initializing the classifier that compensates for the characteristics of sea ice imagery. • In Section 3.4, we define a set of references for assessing classification accuracy. We discuss dual-frequency classification, define a classification confidence mea-sure to aid classifier evaluation in the absence of accurate ground truth, and base our references on dual-frequency classification and classification confidence. • In Section 3.5 we evaluate classification results for polarimetric, dual-polarization, and single-polarization imagery using the references defined in this chapter. • In Section 3.6, we discuss the results and draw conclusions. 29 3.1 Description of the Imagery and the Ice Situa-tion In 1988, National Aeronautics and Space Administration / Jet Propulsion Laboratory ( N A S A / J P L ) conducted an extensive sea ice research program in the Beaufort Sea. A floating ice camp was set up to obtain in situ measurements including ice profiles and meteorological data. In addition, the A I R S A R airborne S A R acquired imagery in the area. Although polarimetric S A R imagery in P-, L - , and C-band were acquired, this chapter focuses on the L- and C- band imagery only for relevance to upcoming spaceborne missions. The authors of a previous study suggest that P-band imagery is similar to L-band with a reduced contrast between first year ice and multi-year ice and is therefore not considered particularly useful for ice monitoring [92]. The sensor noise level (Noise Equivalent Sigma Zero, NESZ) is given as -40 dB at 40 degrees incidence for L - and P-band and -35 dB for C-band. Many researchers analyzed the imagery, their findings are published in [92], [78], [91], [125], [29], [59]. The scene used in this study is Scene 1372. It was acquired on March 11, 1988 with scene center coordinates of 73.048°N and 142.285°Vl / arid an incidence angle variation from 22 to 52 degrees. Figures 3.1 (a) and (b) show colour representations for the 4-look C- and L-band imagery, respectively. A number of ice types that can be identified in the scene are labelled in Figure 3.1 (b). The size of the test site is approximately 6.6 km (range) and 12.1 km (azimuth). In and around the camp, a mixture of First Year (FYI) and Mul t i Year (MYI) ice floes is reported. The average thickness of the F Y I near the ice camp was reported to be between 1.5 m and 2.4 m. M Y I floes were hummocked to heights up to 6 m and are surrounded by compressed F Y I . A layer of dry snow with an average thickness of 15 cm is reported. When the S A R imagery was acquired, the air temperature was below -10°C and the wind speed was below 5 m/s [92]. The ice camp location is not the same as the area covered by S A R imagery. However, the two locations were close enough to assume that similar ice conditions were present. 30 The large floes that appear bright in the C-band image (Figure 3.1 (a)) and purple in the L-band image (Figure 3.1 (b)) are M Y I . Compressed first year ice (CFYI) surrounds the M Y I floes and is best visible in the L-band image where it appears white (Figure 3.1 (b)). Here, the C F Y I is clearly visible and there is generally better contrast between compressed and ridged ice compared to ice with a smoother surface. In comparison, the C-band image does not show good contrast between C F Y I and M Y I . The web of bright lines in both images indicates rough and deformed F Y I (RFYI) and first year rubble and ridges ( F Y I R R ) . Dark purple areas are smooth F Y I (SFYI) and the black features indicate leads that opened in the ice due to dynamic forces. The classification is based on results from a previous study [92]. Other studies suggest that the leads are already covered with Th in Ice (Thi) [125], [59]. Figures 3.1 (c) and (d) are two dual-frequency classification results. They are discussed in detail in Section 3.4.1. 31 CFYI RFYI Thi SFYI FYIRR (a) R: C - V V ; G: C - H V ; B: C - H H (b) R: L - V V ; G: L - H V ; B: L - H H (c) M A P dual frequency result (d) Dual-frequency Reference Figure 3.1: A I R S A R scene 1372 acquired by N A S A / J P L A I R S A R on March 11, 1988 in the Beaufort sea. The image dimensions are approximately 6.4 km in range and 12 km in azimuth, the incidence angle range is about 22° to 54°. (a) C-band R G B false colour image; (b) L-band R G B false colour image; (c) Dual-frequency result using an M A P classifier [92]; (d) Dual-frequency result using a K-means clustering algorithm which uses a Wishart classifier, this result is used as reference throughout this chapter. 32 3.2 Target Decomposition for Sea Ice Monitoring Analysis of polarimetric S A R imagery allows the scattering mechanisms involved to be identified. This in itself is a form of classification, however, the result is interpreted based on scattering mechanisms rather than ice types. Early research on polarimet-ric imagery from sea ice did not include target decomposition. Such methods were developed later and so far are used mainly for land applications. 3.2.1 Dominant scattering Based On Preeman-Durden Freeman and Durden proposed a target decomposition method based on models for three scattering mechanisms: volume scattering, even bounce (i.e., double bounce) and scattering from a rough surface [44]. The method is described in detail in A p -pendix D. Figures 3.2 (a) (b) show R G B false colour images for C- and L-band respectively using the double bounce (red), volume (green) and surface (blue) scattering compo-nents as calculated using the Freeman-Durden method. Both images appear mostly blue indicating the dominance of surface scattering. The Freeman-Durden decomposition can be used to calculate the scattering mechanisms either pixel by pixel or over an entire class (derived using a maximum likelihood classifier for example). In the latter case, the decomposition is applied to the mean signature over all pixels of the class. This approach is discussed later in this chapter. Using the a pixel by pixel approach, the method can be interpreted as an unsu-pervised classifier. Figures 3.2 (c) to (f) show the dominant scattering mechanism for C- and L-band for the original 4-look image as well as filtered image. The 4-look im-age was filtered using a polarimetric filter [64] and a filter window size of five resulting in an equivalent number of looks (ENL) of 60. 33 Dihedral scattering (shown in red) dominates only a very small portion of the imagery, mostly in ridged areas. One exception is a lead region for C-band (see (A) in Figure 3.2 (c)). The red horizontal and green vertical linear features shown in Figure 3.2 (c) (see (B)) are the result of a processing or instrument error in the C-band H V channel. Surface scattering is shown in blue and dominates a large portion of both scenes. There is a correlation between ice types and dominant scattering mechanism. The volume scattering contribution (shown in green) increases for certain ice types, de-pending on the frequency. Volume scattering does not dominate an entire ice type in either frequency, the method is therefore not well suited for a pixel-based approach. For C-band imagery, volume scattering (green) is strongest in portions of M Y I areas, as well as in some areas of compressed ice (C) and leads (D). The latter could be an effect of the H V signal being masked in noise, although, a volume component can be expected in thin ice [78]. M Y I regions show predominantly surface scattering for L-band (see (E) in Figure 3.2 (d)). C F Y I shows dominant volume scattering in L-band (see (F) in Figure.3.2 (d)). The application of a polarimetric filter [64] further decreases the proportion of pixels with predominantly volume scattering in the image. Filtering results in spatial averaging and the backscatter of a larger area compared to a single pixel is estimated. Pixels with dominating volume or dihedral scattering are averaged with neighbouring pixels where surface scattering is dominant. This results in an increase of the area with predominantly surface scattering (see Figures 3.2 (e) and (f)) for the filtered image. In a recent publication, the use of the Freeman-Durden method is suggested with respect to ice type separation for L-band imagery [75]. Our results presented in Figure 3.2 do not support their conclusions. One likely difference is the geographical separation of the test sites leading to different ice regimes. The imagery used in [75] were acquired in the Sea of Okhotsk (Japan), whereas the A I R S A R imagery used in our study was acquired in the Beaufort sea. 34 C-band L-band (a) Freeman/Durden RGB composite (b) Freeman/Durden RGB composite R: Dihedral, G: Volume, B: Surface R: Dihedral, G: Volume, B: Surface flight,, look Dihedral ] Volume | | Surface A •• .• •  • Freeman-Durden dominant scattering mechanism - unfiltered data Freeman-Durden dominant scattering mechanism -filtered data Figure 3.2: Freeman-Durden scattering mechanisms. (a),(b): R G B false colour im-ages; (c)-(f): Classification results based on dominant scattering mechanism (surface scattering; volume scattering, or double bounce). 35 3.2.2 H/A/a Decomposition A set of three parameters, namely entropy, H , anisotropy, A , and a-angle, can be used to analyze distributed targets with respect to the scattering mechanisms involved [21]. The a-angle describes the type of scattering, whereas H and A describe the randomness of the scattering and the ratio between the minor scattering mechanisms, respectively. A more detailed description of these parameters, including their use for automated classification is given in Appendix D. Figures 3.3 (a) and (b) and 3.4 (a) and (b) show the C- and L-band H / a and H / A scatter plots respectively. For both frequencies the a-angle is generally low, limiting the utility of the standardized partitioning of the classification space as only few classes are used. Despite this limitation, the H / A / a classification results indicate some potential for ice class separation if applied pixel by pixel. Similar to the Freeman-Durden decomposition, the H / A / a decomposition can be applied to the mean signature over all pixels of the class. This application is discussed later in this chapter. C-band Result The corresponding C-band classification result is shown in Figure 3.3 (c). The general set of thresholds provides 16 classes as outlined in Appendix D. 15 of .these classes are populated but only four of those 15 classes are heavily populated, resulting in the separation of M Y I , smooth and rough F Y I and T h i . A l l other classes were com-bined and shown in gray. Classes are pre-interpreted with respect to their scattering properties due to the physical interpretation of the decomposition parameters. The classifier is therefore fully automated. The F Y I classes are present in the low entropy surface scattering class (see A p -pendix D) . Similar results are presented in [22]. Anisotropy provides some separa-tion between smooth (A > 0.5) and rough ice (A < 0.5). Relatively higher a-angles 36 (40° < a < 50°) are observed over the leads (Thi). This class, the medium entropy dipole scattering class, is also characterized by high entropy values (0.5 < H < 0.9). M Y I is found in the medium entropy surface scattering class (0.5 < H < 0.9; a < 40°), indicating the presence of a secondary scattering mechanism (here volume scattering). The low Anisotropy is consistent with higher surface roughness. The pixel-based result is rather noisy and therefore would not provide a good ice concentration estimate. Some of the classes sparsely populated do not correspond to ice types and are a result of the processing problems visible in the H V channel. L-band Result A three class L-band classification result is shown in Figure 3.4 (c). Here the thresh-olds were manually set in the H/a-plane to provide only three classes, M Y I , F Y I and T h i . Visual observation of the scatter plot suggests the presence of two distinct clusters that can be separated using a threshold of H=0.25. Further analysis reveals that a threshold at a = 30° (over all H values) results in T h i being identified. This example shows how M Y I , F Y I and T h i can be separated using L-band imagery by manually editing the thresholds of the H / a domain. 37 C-band (c) H / A / a Classification Figure 3 .3: C-band results, (a) H / A - scatter plot, (b) H / a - scatter plot, and (c) example automated classification result (4 of 15 classes shown). 38 L-band 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 Entropy H [1] Entropy H [1] (a) H/A scatterplot (b) H /a scatterplot (c) H /a Classification - manually modified thresholds Figure 3.4: L-band results, (a) H / A - scatter plot, (b) H / a - scatter plot, and (c) example manual classification result (3 classes only). 39 3.3 Maximum Likelihood Classification Based on the Wishart Statistics of the Covariance M a -trix Image classification in the maximum likelihood sense is not based on the physical event responsible for image generation (i.e., scattering) but on a tendency of the data of similar ground cover to form clusters in the feature space. Knowledge of the data statistics (i.e., the theoretical statistical distribution) allows the derivation of a classification approach that is optimal in the sense that, on average, its use yields the lowest probability for misclassification [45]. The Wishart classifier is a M L classification method for n-look polarimetric im-agery [61]. For each pixel, the distance d to a set of class means Vm can be calculated as d((Z),Vm) = l n | V m | +Trace (V^-^Z)) . (3.1) Z is the data sample for the pixel analyzed. A pixel is assigned to the class to which it has the minimum distance d d((Z).Vrn) < <l({Z).Vj) for all ^ • / ^ „ . (3.2) Initial classes are derived using a training data set in known terrain. The image is then classified according to the distance to the class means, which form clusters in the feature space. Each pixel is assigned to the cluster to which it has the mini-mum distance. The distance itself is calculated to provide that the Bayes maximum likelihood classification condition is satisfied [61]. The output of the algorithm is a set of clusters in the feature space, which form segments in the image space. Only, if these segments are connected to some ground cover type, should the term "classes" be used. In the supervised case, where user selected training data is used, the output is related to the training data and can be called classification. 40 The method is look independent and can be modified for use with multi-polarization imagery as well as multi-frequency polarimetric imagery. A more detailed description of the Wishart classifier is provided in Appendix D. A combination of the Wishart classifier with an unsupervised method based on target decomposition is suggested in [60] and [63]. The approaches show good results for land applications, where a variety of different ground cover is present (i.e., fields, forest, urban areas). The unsupervised method is used to initialize the Wishart classifier. Rather than using only one iteration of the Wishart classifier, a K-means clustering algorithm which uses a Wishart classifier is used. The combined method does not require user selected training data, and can therefore be interpreted as unsupervised method. The K-means approach wil l cause the cluster centroids to move in the feature space with each iteration. The output should therefore be treated as a segmentation as it requires some degree of interpretation. A more detailed description of the methods available is provided in Appendix D. For sea ice classification, we propose a simplification to approaches currently used. Target decomposition based classifiers are limited by the scattering properties of sea ice, where surface scattering dominates a significant portion of the imagery. We wil l show, that a simple backscatter strength-based initialization of the K-means clustering algorithm which uses'a Wishart classifier provides good ice type separation. Our classifier involves the following steps: 1. Mask any land areas in the imagery. Only marine areas (sea ice and open water) are to be used for classification. 2. Generate trial centroids for a number of major clusters based upon the total backscatter from the scene. The human analyst may decide on the number of clusters. For operations, we suggest the maximum number of classes to be four, as CIS ice chart egg codes provide information on up to four ice types. This step is only required once and is used to initialize the following step. 3. Apply a K-means clustering algorithm which uses a Wishart classifier. 41 4. If polarimetric imagery is available, assist human analysts in the identification of the class types associated with each of the four major clusters by apply-ing target decomposition techniques to their respective centroids and thereby permit interpretation of the results in terms of the total backscatter and the relative contribution of various scattering mechanisms. We have accounted for the unique properties of sea ice imagery in the following ways. First, we do not include land areas in the classification. None of the airborne scenes analyzed in this thesis contain land, this step is implemented for the E N V I S A T -A S A R scenes available. Second, because target decomposition reveals that only a few scattering mech-anisms contribute to backscatter from sea ice, existing methods for generating trial centroids are unnecessarily complicated when applied to such imagery. Generating trial centroids based upon total backscatter is easy to do, yields very good results, and is applicable to single-polarization, multi-polarization, and polarimetric imagery. Third, we recognize that human analysts play an important role in operational sea ice monitoring and that pixel by pixel machine classification has significant l imi-tations. Therefore, we seek to augment rather than replace human analysts. The output of the proposed algorithm is a segmentation and class interpretation by a human analyst is generally required. This is considered an integral part of the method, which we subsequently refer to as a classifier. The remainder of this chapter is organized as follows. First, dual-frequency data are classified and a classification confidence measure is defined and used to generate a set of reference solutions. Second, polarimetric, dual-polarization and single-polarization data are classified and the results are compared to the references. 42 3.4 Evaluation of Classification Results using L i m -ited Ground Truth For sea ice remote sensing, ground experiments are not generally available and, if they are, they only cover a small area. In contrast to agricultural applications, the collection of information on the ground is difficult. The assessment of classification results from sea ice imagery therefore remains a challenge. In later chapters expert interpretation of a scene (and existing ancillary data) is used in some cases to assess the ice situation in a scene. This "ground truth" is ambiguous and makes it difficult to provide a quantitative analysis of the classification result. In an effort to provide information on the classification accuracy, a number of methods is tested in this chapter. The methods include: 1. Comparison with full dual-frequency result [92]: Reference A . 2. Comparison of a selected sub area with the dual-frequency result. The sub area is selected based on a classification confidence measure. Two different approaches are tested: References B and C: 3. Use of validation data (User selected): Reference D 4. Use of confidence measure only (without reference classification). 3.4.1 Dual Frequency Classification Figures 3.1 (c) and (d) show two dual frequency classification results with six classes based on a previous result [92]. A result from a previous study using a Maximum A Posteriori ( M A P ) classifier is shown in Figure 3.1 (c) [92]. Figure 3.1 (d) shows the maximum likelihood classification based on the Wishart distribution of the dual frequency imagery. A detailed description of both classification methods used is 43 provided in Appendix D. Figure 3.1 (d) wil l be used to provide reference data in this chapter. Tables 3.1 to 3.4 show selected polarimetric parameters for each class of the reference result for C- and L-band respectively. A l l parameters are estimated from the class mean (i.e., the averaged coherency matrix, see Appendix D) . Polarimetric parameters are explained in detail in Appendix D. A l l class means show dominant surface scattering (see the surface to volume scattering ratio Ps/Pv as well as the a- angle). They represent the average of all pixels of one class. Spatial averaging (filtering) of the imagery shows a significant reduction of the proportion of volume scattering dominated pixels (see above); calculating the class means leads to the same effect. C-band For C-band, the parameters support the separation of C F Y I , M Y I , F Y I and T h i . The three F Y I classes show differences mostly in their backscatter strength. There is only a 0.5 dB difference between surface scattering Ps and volume scat-tering Pv for leads (Thi) and 0.8 dB for C F Y I indicating a strong volume scattering component. M Y I shows a 3 dB difference between Ps and Pv, the other three classes (all F Y I ) differ by about 6 dB. Volume scattering dominating in large portions of the leads is an indication that the leads are in fact covered with thin ice. This result is in agreement with previously published analysis results for this imagery. Based on the Freeman-Durden model, there is some potential for error since the system noise level is in the order of the backscatter signal for the cross-polarized channel. Freeman-Durden use the cross-polarization signal to estimate the volume scattering component and a noise level in the order of the cross-polarization signal backscatter is expected to negatively affect this estimation. 44 Table 3.1: C-band polarimetric parameters for reference result estimated from the class means - backscatter related TP Ps/Pv Ps/Pd HH HV VV Ice type [dB] [dB] [dB] [dB] [dB] [dB] C F Y I -6 .5 0.8 9.2 -10.2 -19.2 -9 .9 M Y I -7 .7 3.1 10.7 -11.3 -21.8 -10.8 F Y I R R -13.3 5.2 9.6 -16.6 -29.1 -16.5 R F Y I -16.3 6.1 9.4 -19.4 -32.7 -19.5 S F Y I -18.6 6.1 8.8 -21.6 -35.1 -22.0 T h i -23.5 0.5 5.4 -27.8 -36.4 -26.3 Table 3.2: C-band polarimetric parameters for reference result estimated from the class means - continued . H A a \P\ Ice type • [°] C F Y I 0.69 0.19 28.3 0.65 M Y I 0.58 0.23 22.5 0.73 F Y I R R 0.52 0.39 20.7 0.74 R F Y I 0.50 0.47 19.1 0.74 S F Y I 0.52 0.51 19.6 0.72. T h i 0.76 0.39 34.8 0.46 45 Table 3.3: L-band polarimetric parameters for reference result estimated from the class means - backscatter related TP Ps/Pv Ps/Pd EE EV VV Ice type [dB] [dB] [dB] [dB] [dB] [dB] C F Y I -8 .9 3.7 22.2 -14.4 -23.2 -10.8 M Y I -15.1 11.2 20.1 -20.6 -35.6 -16.6 F Y I R R -13.6 5.6 13.4 -19.0 -29.4 -15.4 R F Y I -19.5 6.9 12.1 -25.0 -36.4 -21.1 S F Y I -25.1 7.5 11.4 -30.9 -42.4 -26.7 T h i -30.4 5.0 11.9 -37.4 -45.8 -31 .7 Table 3.4: L-band polarimetric parameters for reference result estimated from the class means - continued H A a I P I . Ice type [°] C F Y I 0.49 0.06 24.8 0.79 M Y I 0.19 0.21 16.6 0.93 F Y I R R 0.45 0.25 23.5 0.78 R F Y I 0.42 0.37 23.8 0.77 S F Y I 0.41 0.44 24.7 0.76 T h i 0,50 0.30 32.2 0.63 46 Differences for H , A , a are more subtle but support the general interpretation. A higher entropy indicates more mix in scattering mechanisms ( C F Y I , Th i ) . The corresponding a-angle is also higher, supporting an increase in volume scattering for these ice types. The anisotropy provides for a differentiation of the three F Y I types. L-band The L-band parameters show a more significant dominance of surface scattering. Here, the M Y I floes have the relative weakest volume backscatter contribution (11.2 dB difference to Ps). Pv is relatively strong for C F Y I (3.7 dB difference to Ps). H , A , a and correlation magnitude are also useful in class interpretation, they support the information obtained from the Freeman-Durden model. Classes with a Ps to Pv ratio close to zero also show the highest entropy values thus indicating the presence of several backscatter mechanisms of similar strength. The generally low a-angle (< 35°) also shows the dominance of surface scattering. 3.4.2 Classification Confidence for the Wishart Classifier The Wishart distance measure, d, as defined in Chapter 3.3, can range from —oo to oo. Using d2 = e~d (3.3) to transform d, the modified distance d2 now ranges from 0 to oo. In addition, the maximum value of d2 is relevant for classification. A classification entropy measure, Hclass, can now be derived, describing the information content of the distance mea-sures for n classes Haass = y^-Pilo^iPi) where Pt = ^ . . (3.4) Hciass ranges from 0 to 1. A low classification entropy indicates a dominant maximum d2 and hence high confidence in the classification. The measure depends 47 on the number of classes used and is less suitable for problems where this number is high. In sea ice monitoring, the number of ice types under investigation is generally low. The information of one egg-code symbol is restricted to three ice types and open water for example. A scene with a high variability of ice types is divided in regions with different egg codes. Based on all distances for one pixel, a confidence measure can therefore be derived as conf = 1 - Hdass . (3.5) The confidence measure, conf, ranges from 0 to 1 with high values indicating high classification confidence. It can be calculated for multi-polarization and multi-frequency polarimetric imagery, in this case only the classifier needs to be modified as discussed in Appendix D. Figures 3.5 (a) and (b) show images of the confidence measure for a C-band classification and the dual-frequency classification presented in the previous section. The distribution of the confidence measure for each class is shown in Figures 3.5 (c) and (d). Both results show a dependence of the confidence on the class. The dual-frequency result shows an increase of 29% in overall confidence level (0.82 vs. 0.53 for the C-band), thus indicating a higher quality of the classification. This result also supports the decision to use the dual-frequency result (Reference A) or parts of it (Reference B ,C) as reference solution in this chapter. Using the confidence measure and the knowledge. about its class dependency, measures can be taken to utilize only areas where the classification shows high confi-dence. Here, two options are tested: • Reference B : For each class use to top 50%, based on conf. In addition, only those pixels are included in the reference area, where a C-band only classification in the last iteration results in the same class as the dual-frequency classification. Less than 50% of the image are used as reference. 48 • Reference C: For each class use to top 10%, based on conf. This is the most restrictive of the references based on the dual-frequency result as only about 10% are used as reference. A fourth option for a reference was tested. In this case (Reference D), several regions were manually selected. Only these regions, two for each class, were then used for comparison. This reference is independent from the dual-frequency result. Confusion matrices comparing various classification results to the references derived in this section are given in Appendix E . A n overview of the overall classification accuracies for variants of the K-means clustering algorithm which uses a Wishart classifier is provided in Table 3.5. The concept of confidence thresholding can even be used to assess a classifi-cation result without the availability of a reference. In this case, all areas with a confidence below a threshold (a different threshold for each class is recommended) are not considered for the result. The excluded regions require a different approach to obtain classification information. Manual interpretation is one option, however, the increased interpretation effort makes this approach an unlikely choice. 3.4.3 Adaptive Post-Classification Filter We presented an alternative use of classification confidence in [102]. A post-processing operation was developed to reduce the variability of the classification of individual pixels. The proposed method is applied on the classification result. Considering the distance measure from the ML-classifier as a confidence measure on the initial result, the classification of neighborhood pixels can be used to alter the result if the confidence is low. The neighborhood is selected in the same way as in the polarimetric filter presented in [64]. This post-classification filter can be applied to classification results of unfiltered and filtered single- and multi-polarization imagery alike. . 49 (a) C-band Confidence [0..1] (b) Dual-freq. Confidence [0..1] o u 0.5 -J 0.0 J 0.72 (±0.12) 0.59 (±0.05) 0.46 (±0.03) i 0.47 0.39 ° - 4 1 (±0.07) (±0.02) (±0-04) CFYI MYI FYIRR RFYI SFYI Thi (c) Confidence distribution per class (C-band) 1.0-, B C U 0.5 0.0 J 0.92 0.91 0.78 (±0.14)(±0.11)(±0.15) 0.68 (±0.10) 0.70 (±0.08) 0.77 (±0.12) CFYI MYI FYIRR RFYI SFYI Thi (d) Confidence distribution per class (dual-frequency) Figure 3.5: Classification confidence: (a) C-band classification; (b) Dual frequency classification; (c) Confidence distribution per class (C-band); (d) Confidence distribution per class (Dual-frequency). The colour in the scatter plots indicates the population density, ranging from (blue) to high (red). Mean values and standard deviations are provided. 50 Table 3.5: Overall classification accuracies for C-band imagery Polarimetric Dual-Polarization Single-Polarization C T P C F D H H + H V V H + V V H H V V H V Reference A Full scene 65.9 66.1 66.3 66.2 56.3 57.7 64.4 Reference B Top 50% 89.2 81.0 89.6 89.4 71.2 73.8 82.4 Reference C Top 10% 83.9 67.3 81.2 81.8 71.4 71.2 83.0 Reference D Manual selec-tion 82.0 84.6 80.4 74.6 67.1 62.0 78.4 Note: C T P : Total power initialization C F D : Freeman-Durd'en initialization 3.5 Single Frequency Classification Results 3.5.1 Classification of Polarimetric Imagery of Sea Ice W i t h few dual-frequency sensors available to date, much interest lies in the capability of single-frequency imagery classification. Operational constraints discussed in Chap-ter 2 wil l limit the utility of spaceborne polarimetric imagery. A discussion of the classification potential of such imagery is important for the following reasons: • Narrow swath polarimetric imagery can be useful in high traffic areas. • Classification potential of multi-polarization imagery can be compared against polarimetric imagery. • Future SARs may provide wide swath polarimetric modes. 51 K-means clustering algorithm which uses a Wishart classifier Figures 3.6 (b) and (d) show the classification result using the Freeman-Durden ini-tialization [63]. W i t h little dominant dihedral scattering available only two scattering mechanisms, surface and volume scattering, were used to separate the data. Six sur-face scattering classes were used as all classes of the reference result show strong surface scattering. Two volume scattering classes only were used as dominant volume scattering was found for only few ice types. While the classes are interpreted with respect to scattering mechanisms, they still have to be manually assigned to ice types. Our proposed method as described in Section 3.3, is also evaluated against the dual frequency reference result (Figure 3.1 (d)). The image was divided into initial classes based on total power (TP) to initialize the K-means clustering algorithm which uses a Wishart classifier. This approach was chosen because the imagery shows predominantly surface scattering. Figures 3.6 (a) and (c) show the classification result based .on the total power initialization. Six classes were chosen in accordance with the reference result. Similar to the Freeman-Durden initialization method, result interpretation is a required manual process. For both approaches, the initial class separation is based on non-uniformly sep-arated signal strength levels. First, the data were divided in half using the median value for T P , Ps and Pv respectively. For T P and Pg, six classes were needed, which was implemented by dividing each of the two class again using the median value for the class. The four resulting classes approximately cover the same number of pixels. A t last, two more classes were created containing all pixels with the lowest 1% and 5% backscatter strength respectively. The low backscatter class approach is necessary to ensure that leads are identified. During the iterations, classes with low backscatter show a trend to grow with each iteration. Leads were not correctly identified using C-band imagery if the initialization did not contain low backscatter classes. Both frequencies show some limitations in their potential for ice type separation, the use of dual-frequency imagery clearly leads to a better result. A l l single frequency 52 results show the influence of the incidence angle on the classification, which is not the case for the dual-frequency result. (c) L-band: 6 class - TP (d) L-band: 8 class - Freeman/Durden Figure 3.6: Single frequency polarimetric classification results using different ini-tialization options. Quantitative comparisons to the various reference results are presented in Appendix E . C-band Both C-band results (Figures 3.6 (a) and (b) show some confusion be-tween M Y I and C F Y I (see the confusion matrices in Appendix E) . Visual inspection of the R G B image (Figure 3.1 (a)) shows that there is rather little contrast between the two ice types in C-band. The confusion is most severe when the full scene is used as reference (Reference A , see Section 3.4). The separation of the two ice types is 53 generally better for the T P initialization compared to the Freeman-Durden initializa-tion. The ridged and rough F Y I classes ( F Y I R R and R F Y I ) as well as S F Y I also show some confusion depending on the reference used. Leads are underestimated, when compared to Reference A , a portion of the leads are then classified as S F Y I . Applying the Freeman-Durden initialization results in two additional classes, which carry some information. These classes do not represent separate ice types and need to be merged with other classes during interpretation. No improvement of overall classification accuracy can be achieved, mainly due to the dominance of surface scattering. For C-band imagery both, Freeman-Durden and total power initialization show better results than the H / A / Q initialized result (not shown here). In addition to leads being identified, the K-means clustering algorithm which uses a Wishart clas-sifier converges much faster thus requiring fewer iterations (n=3 used here) and less processing time. The additional effort required in classifier setup and result inter-pretation for the Freeman-Durden initialization with little to no gain in classification accuracy makes the T P initialization the method of choice. L-band The L-band results (Figures 3.6 (c) and (d)) provide good information on the C F Y I which is also apparent from visual inspection of the R G B image (Figure 3.1 (b)). Leads are overestimated in the total power initialization approach, the Freeman-Durden initialization result shows lead underestimation similar to the C-band results. A n overestimation of M Y I can be reported for both results. Multiple classes had to be merged to a M Y I class (Details on the classes that needed to be merged are provided in Appendix E) . The estimation of ridged and rough F Y I is generally low (no R F Y I was assigned). The overall classification accuracy is significantly lower. This is an indication that the initialization was optimized for C-band imagery and would need to be modified in the L-band case. Here, for example the Freeman-Durden initialization provides the more accurate classification result. 54 Expectation Based Classification A C-band classification result based on expectation methods (see Appendix D for a description of the method) is shown in Figure 3.7, the confusion matrix to the full scene dual-frequency result is shown in Appendix E . The initial number of classes had to be set to seven to achieve a classification of T h i . Reducing the seven classes to six is a manual step. The poor overall classification performance for this method can be attributed to the failure to classify C F Y I and R F Y I . The reference to which the result was compared to is based on filtered imagery, which may influence the accuracy measure as well. The main focus of this work is on the issues arising when spaceborne imagery is available. This wil l be addressed using our proposed method as described in Section 3.3. No attempt to improve the expectation method for sea ice classification was therefore made. RFYI SFYI Polarimetric Expectation Method Figure 3.7: Single frequency polarimetric classification result using expectation maxi-mization. A quantitative comparison to the full scene dual-frequency reference result is presented in Appendix E 55 3.5.2 Classification of Dual-Polarization Imagery of Sea Ice W i t h respect to dual-polarization S A R acquisition modes on upcoming missions and their potential for sea ice monitoring, the availability of one co- and one cross-polarization channel was assumed. The correlation between the two channels is ex-pected to be small and was assumed to be zero. The resulting partial covariance matrix is diagonal, thus simplifying the classification process. Figures 3.8 (a) and (b) show the classification results based on two polarizations only. The corresponding confusion matrices are provided in Appendix E . Both results reach almost the same level of accuracy as the results based on polarimetric imagery. This is the case for all reference options (Reference A - D ) . Table 3.5 shows that there is less than 3% difference in overall classification accuracy Similar observations are reported for the M A P classifier [92]. Using two channels only results in faster algorithms and a smaller data volume to be handled. Given the operational constraint discussed in Chapter 2, this result is particularly promising for the upcoming generation of spaceborne SARs . | CFYI ___] MYI | RFYI I g h J | FYIRR | Thi | SFYI 1 l ook a) Dual-pol HH+HV b) Dual-pol V V + V H Figure 3.8: Single frequency dual-polarization classification results. Quantitative comparisons to the various reference results are presented in Appendix E . 56 3.5.3 Classification of Single-Polarization Imagery of Sea Ice Figures 3.9 (a), (b), and (c) show the classification results based on a single-polarization The corresponding confusion matrices are provided in Appendix E , a summary of overall classification accuracy is provided in Table 3.5. The co-polarization results are generally worse compared to multi-polarization, they show between 8% and 20% less overall classification accuracy. The overall accuracy of the classification using the cross-polarization channel is surprisingly high. Here much of the misclassification happens in areas with low backscatter (Thi , S F Y I ) . Less than 20% of T h i is correctly identified. Given the cross-polarization signal level of T h i and the posted noise level of the system, the misclassification is likely related to the N E S Z of the system. It is expected that higher noise levels of spaceborne sensors wil l mask a significantly larger portion of the cross-polarization signatures thus further reducing classification accuracy. This issue wil l be further discussed in Chapters 4 and 5. 57 (a) Single-pol HH Figure 3.9: Single frequency single-polarization classification results. Quantitative comparisons to the various reference results are presented in Appendix E . 5 8 3.6 Discussion of Results In this chapter, we have assessed the potential for distinguishing between different types of sea ice using either: (1) various methods for target decomposition based upon particular scattering mechanisms or (2) variants of a K-means clustering algorithm which uses a Wishart classifier, including a new approach for initializing the K-means clustering algorithm which uses a Wishart classifier developed specifically for sea ice monitoring. Our findings are as follows. First, we have shown that both the Freeman-Durden and Cloude-Pottier target decomposition methods provide a consistent physical description of the scattering processes and allow the identification of ice types in airborne A I R S A R imagery. Our analysis of the C-band A I R S A R imagery using both methods suggests that surface scattering dominates the backscatter of first-year ice. It is also a strong component for all other ice types observed. The volume scattering component dominates only, for some portions of compressed first-year ice, multi-year ice and thin ice. Dihedral scattering is of little value for classification, as its contribution is low. Pixel by pixel classification using the dominant scattering mechanism according to Freeman-Durden leads to a salt and pepper appearance of the three class results with the density of vol-ume scattering pixels increasing several ice types. Spatial filtering the data reduced the number of pixels with predominantly volume scattering and decreases the classi-fication potential of the data. Pixel by pixel classification based on a partitioning of the H/A/a-space allows thin ice, first-year ice, and multi-year ice to be distinguished with reasonable success. Second, we have used our insights concerning the scattering mechanisms that contribute to backscatter from sea ice to develop a simpler method for initializing un-supervised clustering algorithms, and demonstrated this using our K-means clustering algorithm which uses a Wishart classifier. We reduce the complexity for the initial-ization from target decomposition to a simple backscatter strength-based method. The leads to faster execution without apparent penalty in classification accuracy. 59 Unlike target decomposition based initializations, this method can be used for po-larimetric, dual-polarization, and single-polarization imagery. The approach can be used to identify ice types with a minimum of user interaction. Parameters like the surface-to-volume ratio from the Freeman-Durden method and entropy and a-angle from the Cloude-Pottier method prove particularly useful for identifying ice types by analyzing class averages. For example, the C-band surface-to-volume ratio of C F Y I and T h i is less than 1 dB, distinguishing them from M Y I (3 dB) and F Y I (approx. 6 dB). The entropy for both is around 0.7 (compared to less than 0.6 for other ice types). The a-angle is about 10° higher (30° vs. 20°) and confirms the higher volume scattering component. C F Y I and T h i have vastly different backscatter strength and can therefore easily be distinguished. Third, we have proposed a classification confidence measure that allows an ob-jective evaluation of the classification result even in cases where ground truth is not available. The classification confidence is high, where class-separation is high. For the Wishart classifier the confidence measure is based on the distances calculated during the classification process. Results shown indicate that the classifier is slightly biased towards classes with higher backscatter where higher confidence values are achieved. The overall confidence for the dual-frequency classification 80% compared to 50% for a single-frequency result. Fourth, we have shown that the use of dual-polarization imagery leads to sim-ilar overall classification accuracy compared to the use of polarimetric imagery of the same quality (NESZ, resolution). Using the dual-frequency (L-band and C-band) classification result and the corresponding confidence measure, we created four dif-ferent reference solutions in order to evaluate the single frequency results: (1) the full scene, (2) the pixels with the top 50% confidence per class, (3) pixels with the top 10% confidence per class are used as reference, and (4) manually selecting ice types in the scene. For the scene analyzed, the difference between classification re-sults obtained using either dual-polarization and polarimetric imagery is less than 3%, at overall classification accuracies above 80% for references (2) to (4) and 66% for reference (1). The accuracy of the classification based on the cross-polarization 60 imagery only is surprisingly high; the reduction compared to polarimetric imagery is only between 1% and 7%. For this channel problems arise for classes in which the signal levels are close to the N E S Z of the system. For example, less than 20% of thin ice is correctly identified. For results based on single-polarization imagery alone, overall classification accuracy falls to about 56% for reference (1), to about 70% for references (2) and (3), and less than 70% for reference (4). These results clearly show that there is a benefit of using polarization diversity for sea ice monitoring. Coverage, which is usually larger for dual-polarization imagery, needs to be traded off with information on the scattering, which is only available for polarimetric imagery. Results suggest that for both cases a backscatter based initialization of a K-means clustering algorithm which uses a Wishart classifier wil l lead to similar classification results. 61 Chapter 4 Evaluation of Simulated RADARSAT-2 Imagery In Chapter 3, we used N A S A J P L A I R S A R imagery to demonstrate the potential for using polarimetric and multi-polarization imagery to distinguish between types of sea ice in S A R imagery. Spaceborne scenes usually cover larger areas making the imagery more suitable for operations. However, differences in coverage, resolution, incidence angle range, and system noise level between airborne and spaceborne sensors must be taken into account before conclusions regarding the utility of spaceborne S A R imagery can be drawn. In particular, spaceborne images usually have a lower resolution and a lower SNR, both of which wil l negatively affect classification performance when compared to airborne imagery. In this chapter, we evaluate the extent to which reduction of resolution, in-crease in N E S Z , and difference in number of looks affects the ability to differentiate between ice types. We simulate R A D A R S A T - 2 polarimetric and dual-polarization imagery by suitably processing CV-580 and E N V I S A T A S A R imagery. A similar approach was used by the Canada Centre for Remote Sensing (CCRS) prior to the launch of R A D A R S AT-1 to provide researchers with simulated R A D A R S AT-1 single-polarization imagery for applications development [16]. 62 The differences between spaceborne and airborne imagery are more pronounced for R A D A R S A T dual-polarization ScanSAR imagery than for R A D A R S A T - 2 polari-metric stripmap imagery. In both cases, it is desirable to reduce speckle using either multilooking or a more sophisticated adaptive filter [64]. The difference in resolution for spaceborne imagery compared to the airborne case limits the level of speckle re-duction available. This wi l l result in reduced class separation due to the amount of speckle present in the scene. The specific objectives of this chapter are: 1. To evaluate the effect of speckle reduction on ice type separation and classifi-cation accuracy. 2. To evaluate the effect of N E S Z levels of R A D A R S A T - 2 polarimetric stripmap imagery on ice type separation and classification accuracy. 3. To evaluate the classification potential of R A D A R S A T - 2 dual-polarization Scan-S A R imagery for the separation of sea ice and open water. The remainder of this chapter develops as follows. • In Section 4.1, we simulate R A D A R S A T - 2 polarimetric stripmap imagery by degrading the resolution and S N R of high-resolution airborne CV-580 imagery. Classification results of airborne and simulated imagery are compared for similar pixel sizes. • In Section 4.2, we simulate R A D A R S A T - 2 dual-polarization ScanSAR imagery by processing E N V I S A T A S A R A P imagery using different range look band-width, range look overlap, number of range looks, and sample spacing in range and azimuth. The simulated imagery is classified to analyze how easily sea ice and open water can be distinguished using the classifier proposed in Chapter 3. • In Section 4.3, we discuss the results and draw conclusions. 63 4.1 Simulation of RADARSAT-2 Polarimetric Im-agery In an effort to prepare for the availability of R A D A R S A T - 2 , M D A has developed a R A D A R S A T - 2 imagery simulator that transforms airborne CV-580 S A R imagery to a form that more closely resembles R A D A R S A T - 2 imagery [118]. While some differences to actual spaceborne imagery remain, the output of this tool gives some insight concerning the potential of R A D A R S A T - 2 imagery. A discussion of the tool and the remaining differences is provided in this chapter. 4.1.1 CV-580 Imagery from the Northumberland Strait On March 8, 2001, the Environment Canada CV-580 airborne S A R acquired polari-metric S A R imagery in the Northumberland Strait off the Coast of New Brunswick. The acquisition campaign was part of an ongoing effort by the CIS to study the bene-fits of multi-polarization and polarimetric S A R imagery of sea ice for their operational needs. Figure 4.1 shows part of a R A D A R S A T - 1 ScanSAR wide image acquired on the same day over the Canadian East Coast, overlaid with a magnified R G B false colour images using the CV-580 channel information H H , H V , and V V . The airborne image covers 6.4 km in slant range and approximately 8 km in azimuth. S L C imagery with a pixel size of 4 m in slant range and 0.43 m in azimuth are available. Because of the high resolution and to reduce the effect of speckle noise, the scene multi-looked (40 looks in azimuth and 4 looks in range) resulting in a pixel size of approximately 17 m in azimuth and 16 m in slant range. Both images shown in Figure 4.1 were acquired in freezing conditions. Climate reports for Moncton (New Brunswick) and Charlottetown (Prince Edward Island) from March 7, 8, and 9, 2001 indicate an average temperature in the area of approx-imately -4.2°C. Temperature records also indicate an extended melt-free period prior 64 Figure 4.1: CV-580 polarimetric imagery (Red: W ; Green: H V ; Blue: HH) on top of a 150 x 210 km subset of a R A D A R S A T - 1 ScanSAR image. Flight line and test area (approx. 6.4 x 8 km) of the CV-580 imagery is shown in red and yellow respectively. Both images were acquired on March 8, 2001 over the Canadian East Coast. The bright areas in the ScanSAR image are landmasses; the black areas are open water. The rest of the image shows sea ice coverage. The vertical streak is the nadir return that is unavoidable in some ScanSAR configurations. 65 to data acquisition. Analysis of the R A D A R S A T - 1 imagery by the CIS shows an ice concentration >90% for the area of the airborne acquisition. The ice is generally classified as medium First Year Ice (FYI) (70-120 cm thick) with floe sizes between 500 m and 2000 m (as observed from R A D A R S A T - 1 imagery). Using additional information like N O A A data, ice breaker and coastal observation, CIS analysts produce a more detailed ice analysis chart. The ice analysis chart shows a mix of ice types in the area of interest [40% thick F Y I (> 120 cm thick), 40% medium F Y I (70-120 cm thick), 20% thin F Y I (30-70 cm thick)]. Based on the more detailed information, floe sizes between 20 m and 500 m are reported. The corresponding CIS ice chart for this date is shown in Figure 2.1. Figure 4.1 illustrates the differences in coverage between the airborne and the spaceborne sensor. The airborne scene covers much less area than the subset of the ScanSAR scene. Little of the rich information content visible in the R G B false colour image is also visible in the ScanSAR image (see Figure 4.1). Resolution is the main issue here. On the other hand, due to the limited coverage, the airborne scene is not a representative example for variations in sea ice signatures present in the region at the time. The scene is not as rich in ice signatures as, for example, the A I R S A R scene analyzed in Chapter 3. The imagery was investigated because of its availability in a format required by the R A D A R S A T - 2 simulator. Also, CIS support for the analysis was available. We tested the calibration of the CV-580 data due to concerns about calibration quality. Following this test, we excluded the near range section of the image was excluded from further analysis. This resulted in a reduction of the incidence angle range from 22° to 64° to 41° to 64°. It was later confirmed by the Canada Centre for Remote Sensing (CCRS) that data specifications are not met in near range and that this area should not be used for detailed analysis [95]. 66 4.1.2 Classification Result - Airborne Visual analysis of the polarimetric R G B composite image reveals six different ice types in the image, four of which are F Y I . Leads and new forming ice are also present in the scene (see Table 4.1). Unfortunately, no additional validation information is available for the acquisition. A n assessment of the ice types is therefore only possible based on the CV-580 imagery available. Expert ice analysts from the CIS aided in the interpretation of the imagery. Table 4.1: Colour assignment for CV-580 classification result. The class number is arbitrary and used only in the classification process. Class number Colour Ice Type 1 grey Young Ice (Yl) 2 magenta F Y I floes, far range (FYIfr) 3 white rough F Y I strong H V (FYIro) 4 green ridged F Y I (FYIri) 5 blue Leads 6 orange F Y I floes, near range (FYInr) In difference to the classifier we proposed in Chapter 3, we classified this scene using the K-means clustering algorithm which uses a Wishart classifier, which was initialized with a result from a modified H / a classification. W i t h only six classes desired, the H / a classification was modified from the standard approach by removing one threshold (H=0.9) as we suggested in [109]. The high number of looks used for the airborne imagery allows the accurate estimation of H and a-angle [68]. Figure 4.2 shows the classification result; the corresponding colour assignment is described in Table 4.1. A comparison to our method proposed in Chapter 3 shows a similar result. . - • 67 (a) R: VV, G: HV, B: HH mm range (b) Classification Result Figure 4.2: (a) CV-580 polarimetric R G B composite and (b) classification result. Test site: Northumberland Strait. (Image L2P5 R l ) . The colour assignment is shown in Table 4.1. Figure 4.3 shows the histograms for the polarimetric channels as well as the total power for each class. The F Y I classes (see Table 4.1) are partially divided by the reflected power. The young ice class has significantly different levels for H H and W . Using one co-polarized channel only, young ice could not be uniquely identified and would be confused with a F Y I type, although with different F Y I types for H H and W . Young ice is well classified using the algorithm described above. The signature of leads is generally weakest and also shows a significant difference in H H and W . The average H V signature is close to -45 dB, the measured noise level for the scene. The maximum H V response is generally below about -20 dB. Several parameters derived from the class means are summarized in Table 4.2. A number of parameters show differences in the classes rather well, while others have only subtle differences. The signatures of leads and young ice are similar. In par-68 2000 1000 2000 1000 Histograms I I I I I I , I / A \ % \ HH I I 1-—,—^Au\y, \ -50 2000 1000 2000 1000 -45 -40 -35 -30 -25 -20 -15 -10 l I I I I [ A A Y\ I I HV —m^Sf^*^ r^—> x i — i — — 0 -50 -45 -40 -35 -30 -25 -20 -15 -10 Figure 4 . 3 : Backscatter strength histograms of the polarimetric channels and the total power. The colour assignment is given in Table 4 . 1 (Rough F Y I is shown in black). 69 ticular H H / V V (< -5 dB) as well as the ratio between the Freeman-Durden surface and volume scattering components (> 7 dB) separate the two from F Y I classes. The polarimetric entropy (<0.5) also provides a clear separation. The magnitude of the complex correlation, the Anisotropy and the a-angle do not provide such clear class separation. A n entropy value of < 0.5 in combination with a low a-angle (< 40°) indi-cates dominant surface scattering, which is supported by the surface to volume ratio of the Freeman-Durden components. A n increased contribution of volume scattering can be noted for the F Y I classes; this contribution is relatively small compared to the surface scattering component except for the rough F Y I class, where Ps/Pv reaches 3 d B . 4.1.3 Simulation of RADARSAT-2 Polarimetric Imagery The transformation from CV-580 S L C imagery to R A D A R S A T - 2 S L C data involves three steps [105]: 1. A d d zero mean complex Gaussian noise at the R A D A R S A T - 2 N E S Z level. 2. Filter the complex airborne imagery in slant range and azimuth directions to reduce the bandwidth to equal the bandwidth of R A D A R S A T - 2 imagery. This filtering represents a loss of information, applied to achieve the desired resolution and spectral shape of the simulated imagery. The filter inverts the assumed spectral shape of the input data and applies a Kaiser window to simulate the spectral shape due to R A D A R S A T - 2 processing. The filter is normalized to. preserve statistics of the intensity values. 3. Interpolation in the slant range and azimuth directions to the desired sample spacing. After these operations, the simulated image has the same noise level, slant range, and azimuth resolution and sample spacing as the anticipated R A D A R S A T - 2 S L C 70 Table 4.2: Class average values for polarimetric parameters (TP: Total Power; \PHHVV\- magnitude of the complex correlation of H H and V V ; PS,PV: Freeman-Durden scattering components for surface and volume ; H,A,a: Cloude-Pottier de-composition parameters) T P H H / W [Ps/Pv] \pHHVv\ H A a Ice Type [dB] [dB] [dB] - [°] Leads -24.5 -5.7 8.4 0.70 0.40 0.51 29.0 Young Ice (YI) -16.1 -5.7 10.0 0.84 0.28 0.41 25.7 F Y I floes, far range (RYIfr) -17.8 -0.2 5.3 0.65 0.58 0.54 25.1 F Y I floes, near range (RYInr) -15.0 -0.2 5.6 0.69 0.54 0.50 23.8 ridged F Y I (RYIri) -12.9 -0.4 4.9 0.68 0.57 0.46 22.9 rough F Y I strong H V (RYIro) -10.3 -0.5 3.0 0.70 0.60 0.27 22.2 71 Table 4.3: Image parameters for CV-580 and R A D A R S A T - 2 S L C (polarimetric modes) CV-580 R A D A R S A T - 2 Parameter standard mode fine mode Slant range sample spacing 4.0 m 11.8 m 4.7 m Slant range resolution 5.0 m 12.9 m 5.0 m N E S Z -40 dB -31 dB -28 dB Azimuth sample spacing 0.434 m 5.1 m Azimuth resolution 0.6 m 8.0 m Incidence angle range 41° - 64° 20° - 49° images. If the input image is calibrated and the filters are properly normalized, then the output image is also calibrated. Table 4.3 shows the relevant image parameters. Several differences from a real R A D A R S A T - 2 S L C image remain in this ap-proach. Firstly, the simulated image wil l be smaller than a real R A D A R S A T - 2 image because of the limited range swath width of the CV-580 image (6.4 km vs. 25 km for R A D A R S A T - 2 ) . Also, the ground range sample spacing wil l differ as the simulated data are subsampled to the same slant range sample spacing. The incidence angle variation of the CV-580 imagery is large because of the low altitude and causes a geometric distortion of the simulated image compared to what a real R A D A R S A T - 2 image would show. The difference in incidence angle coverage (see Table 4.3) wil l also have an impact on the backscatter observed. The difference in incidence angle range 72 to the R A D A R S A T - 2 incidence angle range wil l therefore affect polarimetric analysis results. For classification purposes, the S L C imagery is multi-looked. Figure 4.4 shows four H H images for: (a) CV-580 imagery, (b) R A D A R S A T - 2 standard simulation, (c) fine mode simulation and (d) approximate coverage of the area from the actual R A D A R S A T ScanSAR image. Figure 4.5 shows the corresponding R G B composite images where appropriate. A target pixel size of approximately 15 m in each dimension was chosen for each case, resulting in different levels of multi-looking for each scene. This in combination with the higher noise level and lower resolution of R A D A R S A T - 2 imagery results in an increase in speckle from Figure 4.5 (a) to Figure 4.5 (c). Due to the significantly higher resolution of the airborne imagery, the ratio of looks between the airborne imagery and the R A D A R S A T - 2 standard imagery is about 18, which affects the distribution of the measured parameters. Despite an increase in noise, the R G B false colour images of the R A D A R S A T - 2 simulated imagery appear to retain much of the spatial information. Two different classification strategies were employed for the original and the simulated imagery. The highly multilooked CV-580 imagery was classified using a modified H / a classifier to initialize a K-means clustering algorithm which uses a Wishart classifier. The simulated R A D A R S A T - 2 imagery was classified using the Backscatter-based initialization method we propose in Chapter 3. The results com-pare visually well, though the result based on the simulated R A D A R S A T - 2 imagery is clearly noisier as less speckle reduction was applied. Figures 4.6 (b) to (d) show classification results for simulated R A D A R S A T - 2 fine imagery. A l l images are multilooked (3x3), for images (c) and (d) additional speckle filter is used. The modified H / a classifier was tested but fails to separate leads from F Y I . A low level of multilooking for the simulated scene wil l affect H / a parameter estimation. 73 (a) CV-580 (40x4 looks) (b) R A D A R S A T - 2 fine (3x3 looks) (c) R A D A R S A T - 2 standard (3x1 looks) (d) R A D A R S A T - 1 ScanSAR Wide Figure 4.4: H H Channel of CV-580 Imagery (40x4 looks) in Comparison with Sim-ulated R A D A R S A T - 2 Imagery in Standard Mode (3x1 looks) and Fine Mode (3x3 looks). The pixel dimension varies between 12 and 17 m. The 50 m resolution R A D A R S A T - 1 ScanSAR imagery has only approximate coverage. 74 (a) CV-580 (40x4 looks) (b) R A D A R S A T - 2 fine (3x3 looks) (c) R A D A R S A T - 2 standard (3x1 looks) (d) R A D A R S AT-1 ScanSAR Wide Figure 4.5: CV-580 Polarimetric Imagery in Comparison with Simulated R A D A R S A T - 2 Imagery in Standard Fine Mode ( R = V V , G = H V , B = H H ) The grey scale image is a 50 m resolution R A D A R S AT-1 ScanSAR image. 75 (c) R A D A R S A T - 2 filtered (n=5) (d) R A D A R S A T - 2 filtered (n=7) Figure 4.6: Classification of Simulated R A D A R S A T - 2 Imagery in Fine Mode (3x3 looks and filtered). See text for details. 76 Table 4.4: Confusion Matr ix of 3x3 R A D A R S A T - 2 simulated imagery classification compared with CV-580 reference given in % of the reference class (see text for more details on the reference). FYI ro F Y I r i FYIn r FYIfr Y I Leads FYIro 71 12 1 0 0 0 F Y I r i 24 51 26 8 1 2 FYIn r 4 26 51 25 6 5 FYIfr 1 9 19 58 9 9 Y I 0 2 2 5 78 7 Leads 0 0 1 4 6 77 Overall classification accuracy: 59.2% The classification performance of the simulated imagery can be assessed by com-paring it with the classification derived from the highly multilooked, high-resolution airborne data. The reference was derived from the airborne result with only the top 50% of pixels used based on the classification confidence described in Chapter 3 (similar to Reference B in Chapter 3). Confusion matrices for two of the simulated R A D A R S A T - 2 cases confirm that the use of an additional filter (in this case n=5) improves on the overall classification accuracy (see Tables 4.4 and 4.5). While for visual classification the 3x3 look result is sufficient, the application of a filter is recommended if the imagery is to be used for ice and ice type concentration estimates. The filter size needs to be chosen carefully. For this example, an increase to n=7 results in a slight decrease in overall classification accuracy (68.6% vs. 69.5%) compared to the result derived with n=5. 4.1.4 Effect of Noise and Speckle on Classification Capability To evaluate class separability, the distribution of some of the key features between the CV-580 and the simulated R A D A R S A T - 2 imagery can be examined using 2-D scatter plots of the entire scene. The distribution of three parameters vs. H H are 77 Table 4.5: Confusion Matr ix of filtered (n=5) R A D A R S A T - 2 simulated imagery clas-sification compared with CV-580 reference given in % of the reference class (see text for more details on the reference). FYI ro F Y I r i FYIn r FYIfr Y I Leads FYI ro 72 3 0 0 0 0 F Y I r i 25 62 11 3 1 2 FYInr 2 29 78 25 7 6 FYIfr 1 6 11 72 11 15 Y I 0 0 0 0 79 7 Leads 0 0 0 0 2 70 Overall classification accuracy: 69.5% shown in Figure 4.7. The CV-580 classification result was used for both scenes to define classes and calculate the mean and standard deviation for each class (shown as crosses in Figure 4.7). For the R A D A R S A T - 2 simulated imagery, each sample was assigned the class of the nearest sample in the CV-580 image. Using 2-D scatter plots, class separation for a single feature or a feature pair can be investigated. While the CV-580 scatter plots show good class separation, the R A D A R S A T - 2 scatter plots are affected by the increase in speckle due to less multilooking and the increase in N E S Z . Higher dispersion of all classes can be observed in the R A D A R S A T - 2 case owing to fewer looks applied. Due to increased speckle, the pixels do not form separate clusters for the simulated imagery (right plot) as they do for the airborne imagery (left plot) and the error bars increase in size thus causing significant overlap between classes. The co-polarized scatter plot shows that the increased noise level affects both H H and V V channels in the R A D A R S A T - 2 case. The mean of Class 5 (leads), which is well separated from the rest for the airborne imagery, is now higher, more so for H H than for V V . The smaller increase in V V indicates that information that was picked up in the airborne case is now masked in the noise. This wil l certainly affect the ability for ice-water separation. 78 The co-polarization ratio distribution in Row 2 shows the same problem. Where a simple ratio threshold was sufficient to separate Classes 1 and 5 ( Y l and leads) from F Y I for the airborne imagery, this threshold is no longer obvious in the simulation case (i.e., no separation of clusters and overlap of error bars). Error bar overlap happens not only between neighbouring classes but affects now up to four classes. Class 5 is most affected by the increased noise level and shows a higher average ratio than Class 1. The magnitude of the complex correlation coefficient in Row 3 also shows an increase in the standard deviation, although, the mean values appear relatively un-changed. Again, no separated clusters can be seen in the simulated imagery. Leads (Class 5) appear to be severely affected by the increased noise level, which shows in practically all parameters presented here. Backscatter from leads is largely masked by the increased noise of the R A D A R S A T - 2 imagery. Because the increase in the average V V is smaller than for H H , the airborne imagery indeed picked up a signal from leads. This information is not available for the simulated R A D A R S A T - 2 imagery. In spite of the effect of the increased N E S Z on the backscatter, the confusion matrices (Tables 4.4 and 4.5) do not indicate a reduction in class separability that exceeds that of other classes. For the imagery observed, the reduction in resolution between the high-resolution airborne data and the simulated spaceborne data has more effect on class separability than the increase in N E S Z . 79 F i gure 4.7: 2-D scatter plots with colour coded density information. Left side: C V -580 imagery, 40x4 looks; Right side: simulated R A D A R S A T - 2 fine mode imagery, 3x3 looks The crosses indicate the class means and standard deviations based on the reference solution derived from the CV-580 imagery. 80 4.2 Simulation of RADARSAT-2 ScanSAR Imagery In the previous section, polarimetric R A D A R S A T - 2 imagery was simulated from sin-gle look complex CV-580 polarimetric imagery. For this purpose, the image was filtered to reduce the resolution, complex Gaussian noise was added to reduce the SNR, and finally the data were interpolated to ensure the desired sample spacing. No attempt was made to simulate R A D A R S A T - 2 ScanSAR imagery from the CV-580 imagery. Several reasons can be given for this decision: 1. ScanSAR imagery cover a much wider area than the airborne imagery at hand, as shown in Figure 4.1. 2. Resolution differences are significant; the simulated imagery set would cover only few pixels in range. 3. ScanSAR data acquisition differs substantially from stripmap acquisition. 4. No swath stitching would be included in the simulation. 5. E N V I S A T A S A R A P imagery is available. A full-season data set is analyzed in the next Chapter (Chapter 5). A P imagery is acquired similar to ScanSAR imagery, though no swath stitching is required. The availability of E N V I S A T A S A R alternating polarization imagery opens an-other possible way to solve the problem. Data acquisition is very similar to ScanSAR with the difference that the bursts are used to switch between polarizations rather than between beams [34]. This results in swath widths equal to strip map mode imagery, which are still significantly wider than CV-580 swaths. Also, the incidence angle ranges correspond much better to those of R A D A R S A T - 2 . This section describes the simulation of R A D A R S A T - 2 dual-polarization ScanSAR imagery from E N V I S A T alternating polarization imagery and discusses an example result. 81 4.2.1 EN VIS AT ASAR Alternating Polarization Burst E N V I S A T - A S A R has a single receiver. To acquire imagery in two polarizations, the following approach is used: During co-polarized alternating polarization mode the radar operates in horizontally polarized transmit and receive blocks interleaved by vertically polarized transmit and receive blocks. In order to achieve this, a ScanSAR technique is applied where there are two "subswaths", namely the horizontally and vertically polarized blocks. The size of the blocks is selected to ensure continuous coverage for each polarization and hence information from both polarizations is avail-able to the user for each imaged area. There is no change in antenna swath; the entire information is collected in one incidence angle range [110]. In the cross-polarized modes, the transmit pulses are all H or all V polarization, with the receive chain operating alternately in H and V as in the co-polarized mode [110]. Table 4.6 shows the number of echoes per burst. Following the transition to alternating polarization mode, a stabilization period and a transmit "off" period, noise measurements are made in the transmit polariza-tion. Then each polarization is passed through an initial calibration sequence. The actual data collection is periodically interrupted for one cycle for periodic calibra-tion. These interruptions of the time line are small and wil l not affect the focused result [110]. Table4.6: E N V T S A T A S A R A P Mode: Number of echoes per .burst . Swath IS1 IS2 IS3 IS4 IS5 IS6 IS7-Echoes per burst 194 196 257 218 277 : 238 297 82 4.2.2 RADARSAT-2 ScanSAR Burst ScanSAR technology is used to provide wide coverage; the concept differs from al-ternating polarization data acquisition. The data are acquired in bursts to allow the combination of different swaths as opposed to collection of two polarizations in one swath. Tables 4.7 and 4.8 give an overview of the number of echoes per burst, which is equivalent to the number of samples in the matched filter [70]. While the num-ber of echoes per burst is lower for R A D A R S A T - 2 the real comparison between the two sensors is the length of the burst in time. Here the differences in Pulse Rep-etition Frequency (PRF) causes the actual difference in burst length in time to be less significant. A S A R has a P R F range from 1580-2150 Hz, depending on the swath. R A D A R S A T - 2 uses P R F ' s between 1200 Hz and 1400 Hz. For example, 218 pulses at E N V I S A T rate corresponds to about 160-170 at the R A D A R S A T - 2 P R F . In addition to shorter bursts compared to E N V I S A T - A S A R alternating polarization mode, the gaps between bursts are relatively longer if more than two swaths are combined [69]. Table 4.7: R A D A R S A T - 2 ScanSAR Narrow: Number of samples in matched filter Beam Combinations W l W 2 W 2 S5 S6 Number of samples in matched filter 112 85 Table 4.8: R A D A R S A T - 2 ScanSAR Wide: Number of samples in matched filter Beam Combinations W l W 2 W 3 S7 W l W 2 S5 S6 Number of samples in matched filter 58 58 . 83 4.2.3 ENVISAT A S A R Processor Settings The similarity of the data acquisition allows the simulation of R A D A R S A T - 2 dual-polarization ScanSAR imagery by processing E N V I S A T alternating polarization im-agery with modified processing parameters. The parameters that were changed to give the processed image a more R A D A R -SAT-2 ScanSAR narrow mode appearance are summarized in Table 4.9. The differ-ences in the processing are confined to window type and some range look processing parameters (bandwidth, look overlap and number of looks): For spectral weighting in range and azimuth, a Hamming window is used for E N -V I S A T processing, whereas a Kaiser window is chosen for R A D A R S A T - 2 processing. The Hamming window is designed to minimize the peak sidelobe level. The Kaiser window is more flexible and usually has the largest energy in the mainlobe for a given peak sidelobe level. The range look processing parameters differ as for R A D A R S A T - 2 two looks are used whereas for E N V I S A T - A S A R the number of looks used depends on the swath. Also, a 15% look overlap is used for R A D A R S A T - 2 compared to no overlap for E N V I S A T - A S A R . The lower range look bandwidth set for R A D A R S A T - 2 results in the reduction of information compared to E N V I S A T - A S A R and therefore reduced resolution. The pixel spacing for R A D A R S A T - 2 is set to 25 meters (ScanSAR narrow). One important aspect that is the same for both sensors is that two azimuth looks are used. R A D A R S A T - 2 uses the looks from different bursts; there is no difference between the two sensors expected due to this approach. 84 Table 4.9: E N V I S A T A S A R Processor Settings Comment Variable E N V I S A T A S A R R A D A R S A T - 2 Range processing parameters Range matched fil-ter window type ( K A I S E R , H A M -M I N G , N O N E ) R N G - M F -W I N D O W H A M M I N G K A I S E R Range matched filter window coefficient R N G - W I N D -C O E F 0.75 2.9 Azimuth processing parameters Azimuth matched filter window type ( K A I S E R , H A M -M I N G , N O N E ) A Z I - M F -W I N D O W H A M M I N G K A I S E R Azimuth matched filter window coefficient A Z I - W I N D -C O E F 0.75 2.9 Range interpolation parameters Output range pixel spacing; Can be slant or ground range (float) O U T - R N G -S P A C I N G 12.5 25 Azimuth interpolation parameters Nominal output az-imut pixel spacing (meters) O U T - A Z I -S P A C I N G 12.5 25 85 Comment Variable E N V I S A T A S A R R A D A R S A T - 2 Range look processing parameters 1 2 1 2 Number of range looks array 2 2 to generate for each N U M - R N G - 3. 2 beam (IS1..IS7) L O O K S 3 3 3 2 2 2 0.0 0.15 Range look overlap (IS1..IS7) array R N G - L O O K -O V E R L A P 0.0 0.0 0.0 0.0 0.0 0.0 0.15 0.15 0.15 0.15 0.15 0.15 15.86e6 5.792e6 Range look bandwidth for each beam (in Hz) (IS1..IS7; SS1,SS3,SS5,SS2,SS4) array R N G - L O O K -B W 15.87e6 12.54e6 10.70e6 9.44e6 8.77e6 8.14e6 5.792e6 5.792e6 5.792e6 5.792e6 5.792e6 5.792e6 86 While there are many similarities in data acquisition that allow the approach to simulate R A D A R S A T - 2 imagery from E N V I S A T - A S A R alternating polarization imagery, a few differences need to be taken into account: • The sensor noise levels are slightly higher for E N V I S A T - A S A R , an improvement of up to 3 dB is possible for R A D A R S A T - 2 . This may be important for the cross-polarized channel in particular with respect to sea ice monitoring where cross-polarized return is generally lower than over land. • The availability of two receivers in R A D A R S A T - 2 provides more flexibility for the optimization of one of the receivers for the cross-polarized channel. There seemed to be issues with the E N V I S A T - A S A R A P cross-polarization quanti-zation (not utilizing the dynamic range of the data optimally), which is not expected for R A D A R S A T - 2 , • In contrast to R A D A R S A T - 2 , E N V I S A T A S A R imagery does not provide the relative phase between the two channels, although this information is not ex-pected to be useful. • Image formation for ScanSAR imagery includes the combination of several swaths resulting in difficulties for the calibration of the beam transition area. As only one swath was used for the simulation, no such beam transition is present in the imagery. • The combination of swaths results in R A D A R S A T - 2 having a wider swath that is possible with a single swath used for. the simulation. The main goal of the simulation was to provide R A D A R S A T - 2 dual-polarization ScanSAR imagery with emphasis on the two polarizations rather than the swath width. 4.2.4 Example Simulation Figure 4.8 shows an E N V I S A T - A S A R IS4 alternating polarization (HH+HV) scene acquired on February 7, 2003 in the Gulf of St Lawrence south of Anticosti Island. 87 Covering an incidence angle range from 31° to 35°, the image dimensions are approxi-mately 67.5 km in range and 100 km in azimuth. W i n d speed and direction are shown in Figure 4.8 (b) [4]. Just south of Anticosti Island is an area of open water; the rest of the area is covered in new and young ice (see labels in Figure 4.8 (b)). The E N -V I S A T imagery was available in R A W format. It was processed using the parameters described above to more closely resemble R A D A R S A T - 2 ScanSAR narrow imagery. The H H image (Figure 4.8 (a)) shows high backscatter for the open water area due to wind, thus making delineation of the ice edge difficult. The H V image (Figure 4.8 (b)) provides better contrast between open water and new ice, but not necessarily young ice. The backscatter of H V is significantly lower (10 dB difference in the grey level scaling). Figure 4.8 (c) shows a colour representation of the imagery using the H V / H H channel ratio, in addition to the two individual channels. Visual image interpretation plays a significant role in the CIS operational environment and the use of colour wil l aid ice analysts in their work. Figure 4.8 (d) shows a classification result using our proposed classifier for dual-pblarization imagery (more specifics are provided in Chapter 3). A land mask was used to exclude land areas from the classification. Spatial averaging of 4x4 pixels was performed to reduce the effect of speckle. Resulting classes were manually assigned to new ice (white), young ice (grey) and open water (blue). The result, while still noisy, illustrates the potential of dual-polarization ScanSAR imagery for machine classification, and even ice concentration estimates. A H V vs. H H scatter plot based on the classification result is shown in Figure 4.9. The crosses indicate class mean values, with a spread of ± one standard deviation in each dimension. The colouring of the scatter plot is used to show the density of the plot, with brighter colours representing higher density. While ice and open water classes appear to come from a single large cluster of data points making simple visual interpretation of the scatter plot difficult, the classes are reasonably well separated in the two dimensional space. Neither of the channels alone would be sufficient to classify the scene. Young ice and open water show the same average H V level, whereas 88 (c) R : H V / H H G : H V B : H H (d) 3-class classification Figure 4.8: Simulated R A D A R S A T - 2 ScanSAR example. E N V I S A T A S A R (IS4) acquisition from February 7, 2003 in the Gulf of St. Lawrence. Image dimensions: 67.5 km (range) x 100 km (azimuth). Wind speed and ice types are indicated. 89 -16 -14 -12 -10 -8 -6 -4 HiHfdB] Figure 4.9: H V - H H scatter plot for simulated R A D A R S A T - 2 ScanSAR imagery. Brighter colours indicate higher data density. The cross centers denote class mean values ± one standard deviation indicated by the crossbars. 90 new ice and open water show the same average H H level. H V is close to or at the noise level in the open water and young ice areas. Land is shown here for comparison only. 4.3 Discussion of Results In this chapter, we have evaluated the extent to which reduction of resolution, increase in N E S Z , and difference in number of looks affects the ability to differentiate between ice types. We have done so by simulating R A D A R S A T - 2 polarimetric and dual-polarization imagery by suitably processing CV-580 and E N V I S A T A S A R imagery. Our findings are as follows. First, we have shown that class separability for all classes is affected by the different levels of multilooking applied to the high-resolution airborne imagery and the simulated R A D A R S A T - 2 polarimetric imagery. However, we have shown that spatial filtering of the spaceborne data can help to compensate for the lower speckle reduction. The filter window size needs to be carefully chosen, though. We have assessed classification accuracy of the simulated spaceborne data by using the result based upon the highly multilooked airborne data as a reference. Four first-year ice types, young ice and leads were present in the imagery. • • Second, we have shown that the increased N E S Z of R A D A R S A T - 2 polarimetric imagery reduces class separability for low backscatter classes. While speckle and speckle reduction affects all classes, we have shown that the increase in N E S Z (-40 dB for the airborne imagery versus -28 dB for the simulated spaceborne imagery) only degrades the signatures of classes with low backscatter. Backscatter at or below N E S Z level were measured at N E S Z level and class separation was reduced. For our imagery, one class (leads) showed less than -40 dB signal level for the airborne cross-polarization channel. The simulation reduced the contrast to other classes by increasing the N E S Z . In our example, however, the contrast reduction was not severe enough to significantly degrade class separability. 91 Third, we have shown that R A D A R S A T - 2 dual-polarization ScanSAR modes wil l easily separate sea ice and open water in spite of an even higher N E S Z level compared to its polarimetric modes. A simulated scene based on E N V I S A T A S A R A P imagery contains two different ice types and open water that can only be separated by using both channels. The N E S Z affects the cross-polarization channel, both open water and young ice show levels just below the posted N E S Z but can be separated using the co-polarization channel. The cross-polarization channel shows different levels of open water and new ice, which is not the case for the co-polarization channel due to wind roughening of the water surface. 92 Chapter 5 Analysis of a Full-Season E N V I S A T A S A R A P Imagery 3 In the previous two chapters, we drew conclusions concerning the classification poten-tial of polarimetric and dual-polarization S A R imagery based upon just three images. While a variety of ice types are available in the imagery, these are not fully represen-tative for all ice conditions encountered throughout an ice season. Also, changes in environmental conditions like temperature and snowfall cannot be investigated using the imagery available. In this chapter, we determine the extent to which multi year ice can be distin-guished in spaceborne multi-polarization S A R imagery that covers a full ice season. E N V I S A T A S A R with.its 35 day repeat orbit and seven A P imaging modes provides the opportunity to continuously collect imagery over an area and build a represen-tative test data set to further evaluate the utility of multi-polarization imagery for sea ice monitoring. Between Apr i l and November 2003, the first full-season, multi-polarization S A R imagery set of sea ice was acquired and temperature and snowfall were recorded. This provides a unique test environment for the classification method we proposed in Chapter 3. 3 Results presented in this chapter were published in [99], [100], and [101]. 93 In Chapter 4, we have shown that the higher N E S Z on spaceborne systems com-pared to airborne systems limits the utility of the cross-polarization channel as ice signatures can be at or below the noise level. Following antenna pattern correction, the N E S Z is not longer constant over the swath. Such a variation was not considered during a simulation of polarimetric R A D A R S A T - 2 imagery. The variation is sensor and beam specific and wil l affect classification results if the S N R is low or negative. A mitigation strategy is therefore required. The specific objectives of this chapter are: 1. To further evaluate the utility of spaceborne multi-polarization S A R imagery for sea ice monitoring by identifying M Y I by means of classification using S A R imagery that covers a range of environmental conditions. 2. To propose and evaluate a classification method that wil l mitigate the effect of N E S Z variation over the swath and improve classification. The remainder of this chapter develops as follows. • In Section 5.1, we present the available imagery and describe the environmental situation. • In Section 5.2, we compare ice signatures acquired in different environmental conditions and assess classification accuracy for M Y I . • In Section 5.3, we present example classification results and introduce a strategy for mitigating large N E S Z variations. • In Section 5.4, we discuss the results and draw conclusions. 94 5.1 Resolute Test Area Data Set Between A p r i l and November 2003, E S A regularly acquired A S A R A P imagery over Resolute, Nunavut (74°42' N , 94°54' W ) . Resolute is one of the official E S A calibration sites and acquired imagery is being used for ongoing A S A R instrument calibration. As part of a research agreement with E S A , the CIS obtained all imagery acquired in that period. The selection of modes was made solely for calibration test purposes, sea ice research was only a secondary use of the imagery. 5.1.1 A S A R A P Imagery Table 5.1 summarizes the 16 A S A R A P M acquisitions available. Figure 5.1 shows a sketch map of the area including the approximate IS6 A P M coverage. Combina-tions of co- and cross-polarization imagery was collected over 13 passes (+ 3 single-polarization scenes, see Table 5.1). Table 5.2 shows the available swaths for A S A R A P imagery. The medium resolution product ( A P M ) was selected to approximate Scan-SAR type resolution as well as to allow for longer acquisition lines. This product features a pixel spacing of 75 m and an equivalent number of looks (ENL) of approximately 50 [5]. No additional filter was applied. 5.1.2 Auxiliary Data As part of their regular operation, the CIS acquired R A D A R S A T - 1 imagery and produced ice charts for the period in question (see Table 5.1). A description of the production of ice charts at the CIS including a listing of data and information used is given in Appendix B . Mostly thick first year ice and old ice are identified in the region of interest. Several scenes contain areas of open water. In addition to the ice charts, daily tem-95 Table 5.1: A S A R A P imagery availability for Resolute Date A S A R Imagery A u x Data Beam Pol. Ice chart type R A D A R S A T Apr . 3 IS6 H H + H V Monthly — May 8 IS6 H H + H V Weekly May 6 June 12 IS6 H H + H V Weekly Same day June 20 IS1 V H Weekly Same day July 17 IS6 H H + H V Weekly Same day July 25 IS1 V V + V H Daily July 24 July 27 IS5 V H Daily Same day Aug. 3 IS3 V V + V H Daily Same day Aug. 21 IS6 H H + H V Daily Same day Aug. 31 IS5 H H + H V Daily Same day Sept. 25 IS6 H H + H V Weekly — Oct. 5 IS5 V V + V H Daily Same day Oct. 18 IS4 V V + V H Daily Same day Oct. 30 IS6 H H + H V Weekly Oct. 31 Nov. 5 IS4 H H Weekly Nov. 4 Nov. 9 IS5 V V + V H Weekly Same day 96 Figure 5.1: Area of interest for E N V I S A T A S A R A P imagery analysis. The approx-imate IS6 A P M swath coverage is shown by the dotted rectangle). Table 5.2: A S A R Image Mode Swaths [34] Swath Swath width Incidence angle range Worst case N E S Z [km] [°] [dB] IS l 105 15.0-22.9 -20.4 IS2 105 19.2-26.7 -20.6 IS3 82 26.0-31.4 -20.6 IS4 84 31.0-36.3 -19.4 IS5 64 35.8-39.4 -20.2 IS6 70 39.1-42.8 -22.0 IS7 56 42.5-45.2 -21.9 9 7 perature and snowfall data (see Figure 5.2) as well as wind data from Resolute are available. The imagery covers a wide range of environmental conditions. S A R acquisitions were made in both melting and freezing conditions. The daily mean temperature ranges from -30°C to +10°C (see Figure 5.2). Calm wind (< 10 km/h) are reported for May 8, June 20, August 3 and 31, September 25, and October 30. Wind speeds above 30 k m / h are reported for Apr i l 3, October 5 and 18, November 5 and 9. Few significant snowfall events are reported, mostly in the fall with temperatures well below zero. U 0!) u O c u E— re P c u 15 10 5 0 -5 -10 -15 -20 -25 -30 15 r 1 0 I § 5 CO • 1— C3 P 0 c re 2 Apr ! Mav ' Jun ' Jul ' Aim ' Sen ' Oct Nov 3 ! 8 1 12,20 1 17,25,2713, 21,311 2515, 18,30!5,9 1, II 1 111 1  LL A I , 1 , 1 f , f ' i n T 3 1 8 ; 12,20 ; 17,25,27;3, 21,31; Apr ! May ! Jun ! Jul ! Aug 1 25; 5, 18,30;. Sep ! Oct !l >,9 Calendar Day (2003) Figure 5.2: Temperature and snowfall records for Resolute with A S A R acquisition dates, the latter are shown in gray. 98 5.2 Data Analysis CIS operations are heavily based on human expert analysis. Data visualization is therefore an important aspect. In the analysis, the land area was masked out and the dynamic range of the H H and H V channels of the marine areas was evaluated. -22 dB -10 dB 10 HV, VH [dB] Figure 5.3: Dynamic ranges for marine areas for co- and cross-polarization backscat-ter. Histograms of different swaths cover different areas on the ground. The dashed lines indicate thresholds used for data visualization. Figure 5.3 illustrates the difference in dynamic range for co- and cross-polarization imagery. The histograms cover slightly different areas for different swaths. Also, a wide range of sea ice stages is represented. The examples show that for the imagery at hand, the co-polarization backscatter shows a dynamic range of about 14 dB, compared to approximately 8 dB for cross-polarized imagery. The sensor noise level restricts the lower limit of the dynamic range of the cross-polarization imagery. V V backscatter is shown to be on the higher end of the data interval. This is largely because they were acquired at steep incidence or late in the season with a 99 larger portion of brightly scattering deformed and old ice present in the area. To ensure colour consistency and allow visual comparison of the various scenes, thresholds were applied to all scenes (see Figure 5.3 for threshold values used). Exam-ples for colour and grey level representations of the IS6 imagery available are shown in Figure 5.4. The top row of Figure 5.4 shows all cross-polarized images acquired in beam IS6 (39.1° to 42.8° incidence angle). Old ice can easily be identified due to relatively high cross-polarized backscatter levels, except for images where wet snow obscures the signature (June 12, July 17). The bottom row of Figure 5.4 shows colour images formed from the co- and cross-polarized channels. Representing both channels in a colour image reveals the additional information content of alternating polarization imagery compared to single-polarization imagery. Relative differences in co- and cross-polarization backscatter re-sult in colours (blue for H V , and yellow for HH) . Common attributes in both channels appear in grey levels (and white). The Apr i l and May images (see Figure 5.4) are very similar, with old ice present in the upper part of the images. Both scenes were acquired during stable, late winter ice conditions (though rising temperatures and 7.1 cm snow accumulation were observed). A more unified signature over the entire ice region can be observed for the June 12 image, associated with surface melt (i.e., a.wet snow layer). Temperature data for the day confirm this assumption. The breakup of the ice can be seen in the July 17 image. The imagery acquired in August, September and October show various stages of freeze-up. The change in ice type signatures over time is illustrated in Figures 5.5 and 5.6. The multi-temporal signatures of M Y I and F Y I are shown for the same region (as outlined in Figure 5.4). Three scenes (Apri l 3, May 8 and June 12) show stable ice conditions but significant variations in environmental conditions. The reduction in backscatter difference between the two ice types due to wet snow (June 12) is significant (>50%) for both channels. The backscatter levels between A p r i l and May 100 April 3 May 8 June 12 July 17 Aug 21 Sep 25 Oct 30 April 3 May 8 June 12 July 17 Aug 21 Sep 25 Oct 30 Figure 5.4: A S A R A P M IS6 images acquired in 2003. Top row: H V scaled from -25 dB to -19 dB. Bottom row: R , G (yellow): H H scaled from -22 dB to -10 dB; B : H V (scaled as above). Old or compressed ice shows as bright grey, open water or thin ice shows in blue. Note the variation of H V over the swath in the upper row (especially in the early season images), which creates a yellow band in the colour images below, adversely affecting the interpretation. The red rectangle shown in the upper row indicates an M Y I floe embedded in F Y I . 101 April 3 May 8 June 12 Figure 5.5: M Y I floe embedded in F Y I for three acquisition dates with little change in the ice situation in between acquisitions. Wet snow affects the signature for the June 12 scene. The mask shown is used to estimate signal levels shown in Figure 5.6. 102 m -12 re o </> ° -14 re m -16 -18 U -20 h --22 \— • HH - MYI • HH - FYI ® HV - MYI • HV-FYI J Acquisition date: April 3 May 8 June 12 Air temperature: -21.6 °C -14.3 °C +0.2 °C Snow accumulation (since April 3): 0 cm 7.1cm 15.7 cm Figure 5.6: M Y I and F Y I signal levels for three acquisitions shown in Figure 5.5. Environmental information is provided for each acquisition date. 103 are within 0.55 dB. The co-polarization channel appears to provide the slightly better separation of F Y I and M Y I in this case even under wet snow conditions. A l l IS6 colour images (see bottom row of Figure 5.4) show a banding effect, which is most visible in near range. The left side of each image shows a yellow colouring indicating a co-polarization signature that is relative stronger than the cross-polarization signature. The top row of Figure 5.4 shows that it is in fact the cross-polarization signature that is lower in near range than in far range. This observation is not expected and the effect is assumed to be system related. The variation of the cross-polarization signature over the swath is not restricted to IS6. Cross sections across the swath in homogeneous areas (e.g., June 12, October 18) show a range variation of the cross-polarization component in homogeneous areas (see Figure 5.7). As the cross-polarized channel is expected to have a small dependency on the incidence angle, little variation over the image swath should be present [76]. The observed variation is approximately 1.5 dB for IS6 First Year Ice (FYI) data and about 4 dB for IS4 open water (data from other beams did not provide a homogeneous area over the full swath). This variation in H V over the swath can be attributed to the variation of the signal/noise ratio (SNR). This noise limitation is not expected to be an issue for land applications. Low cross-polarized backscatter from smooth first year ice and open water compared to the noise level wil l affect both visual analysis and automated classification. A n approach to address the issue for the IS4 swath is presented later in this chapter. 104 I 1 . 1 I I 1 I 1 I I I I 1 I I 1 1 , • I . . . Near range Far range (a) Open Water Range Profile, October 18, IS4 (b) First Year Ice Range Profile, June 12, IS6 Figure 5.7: Range profiles for the cross-polarization channel. The two examples show system related variations of the signal level over homogeneous areas. 105 5.3 Classification Results The imagery available represent a unique test data set for the classifier presented in Chapter 3. In this section we wil l propose an improvement to the method that mitigates the effect of N E S Z variation over the swath, which affects classification accuracy. 5.3.1 Standard Classification The imagery is classified using a K-means clustering algorithm which uses a Wishart classifier as described in Chapter 3. Four classes are used, corresponding to the maximum number of classes available in a standard egg-code. The initial classes are set up by using the median values of the marine areas for the two channels to separate low and high backscatter pixels. The classifier is run with three iterations, where the class means are updated after each iteration. The interpretation of the final classes is a manual task. Land areas are excluded from the classification by means of a land mask. Fo-cusing on the marine area only is a crucial step as backscatter from land is generally higher then backscatter for sea ice. Any classification result including land areas wil l require a significantly higher number of classes, thus increasing interpretation effort [98]. Coastline data are generally available in vector format, land masks can therefore be created for operations. One specific test includes a test of how the classifier works for separating sea ice and open water. A reduction to two classes (ice and water) proved unsuccessful, the two types (sea ice and open water) were not well separated [96]: This is likely because of the variation of the ice signatures present, which is best accounted for by using more than one ice class in the process. Nevertheless, the number of classes needs to-be traded off with the class interpretation effort, to minimize the amount of human interaction. In a semi-supervised scenario the number of classes could be decided by 106 the analyst. The analyst also identifies the region to be analyzed. A l l scenes with two polarizations available were classified using the method de-scribed. Table 5.3 provides a qualitative overview of the results. The opinions given in the table were agreed on after a discussion of the results with experts from the CIS [42]. W i t h no reference result or accurate ground truth available, a quantitative analysis of the results is not possible. (a) R, G : H H ; B : H V (b) 4 class classification result Figure 5.8: Two-channel colour image (left) and classification result (right) of the September 25 scene (IS6). The colour assignment for the right image is: white: deformed ice, gray and orange: first year ice (slight differences in roughness), blue: thin ice or possibly open water. Land is shown in black. Figure 5.8 shows a colour representation of both channels for September 25, 2003 (IS6, H H + H V ) , and a four-class classification result. The classification result preserves the spatial detail of the R G B false colour image and identifies ice types with a common radar signature. However, it is not obvious if the blue class is open water or thin ice; these two scatterers cannot easily be separated. 107 Table 5.3: Classification evaluation (qualitative) Swath Date Performance Criteria Old ice detected Ice vs. open water Leads detected Comments IS6 Apr . 3 + • - - Inc. angle artifact IS6 May 8 + - - Inc. angle artifact IS6 Jun. 12 + - - Inc. angle artifact + surface melt IS6 Jul . 17 - + Surface melt IS6 Aug. 21 . + + + Large area of open water IS6 Sep. 25 N / A +. N / A Unsure if thin ice or open water IS6 Oct. 30 - • N / A - Overestimation of leads and M Y I IS5 Aug. 31 + + . + IS5 Oct. 5 + + + IS5 Nov. 9 + • + + IS4 Oct. 18 + • + + Incidence angle ef-fect in far range IS3 Aug. 3 N / A + + I S l Jul . 25 N / A + + 108 Figure 5.9: Two-channel colour image and classification result (right) of the July 25 scene (ISl). The colour assignment for the right image is different from Figure 5.8: white, gray: first year ice (differences in roughness), dark blue: thin ice, light blue: open water. Land is shown in black. 109 Figure 5.9 shows an colour representation and a classification result for the July 25 scene, which was acquired in IS1 ( V V + V H ) . Classes visible in the R G B composite are identified as such in the classification result. Note the generally low backscatter for thin ice (dark blue in classification result, black in the colour representation) compared to low cross-polarization and high co-polarization backscatter for open water (light blue in classification result, bright yellow in the colour representation). This signature inversion happens due to the low incidence angle for IS1 (15° — 22.9°) and is not necessarily wind related. For the IS6 scenes many classification results are affected by the low level and the variation of the cross-polarization return over the swath. The Apr i l , May and June acquisitions show relatively stable ice conditions but environmental conditions change significantly. Table 5.4 shows confusion matrices evaluating the capability to classify M Y I , the only class not affected by the H V variation. Between A pr i l and May, the classification performance is stable. Between Apr i l and June, a reduction in classification accuracy due to the wet snow layer and the resulting reduction in contrast can be observed. The rate of 67% M Y I (April) classified as M Y I (June) is underestimated partly due to a breakup of part of the ice in the June 12 scene (upper far range section in Figure 5.4). A l l other swaths show generally better classification results. Due to the limited number of scenes available per swath and different ice situations in the various scenes, the result is not fully representative. W i t h the exception of the October 30 scene, performance issues can mostly be attributed to the low level of cross-polarization backscatter and to the system related variation of the S N R over the swath. This issue is addressed below, but ultimately, more power is required to get better results at higher incidence angles. 110 Table 5.4: IS6 multi-temporal confusion matrices given in % relative to the A p r i l 3 scene. A p r i l 3 A pr i l 3 M Y I Other May 8 M Y I 89 3 Other 11 97 June12 M Y I 67 9 Other 33 91 Three scenes with stable ice conditions are compared. MYI classification accuracy only is analyzed due to unsatisfactory classification performance for other classes. See text for more detail. 5.3.2 Consideration of NESZ Variation Over the Swath The Noise Equivalent Sigma Zero (NESZ) of a spaceborne system is generally higher than that of airborne systems. It may reach or even exceed the cross-polarization backscatter level. In this section, an attempt is made to utilize the known potential of the cross-polarization channel for ice type classification. A classification scheme is suggested that allows the utilization of dual-polarization information in areas where the cross-polarization channel is expected to contain useful information. As shown before (Figure 5.7), the October 18 V V + V H imagery, acquired in IS4 mode (31° — 36.3° incidence) show significant V H backscatter variation across the swath. The two channels are shown in Figure 5.13 (a) and (b). Available auxiliary information is summarized in Table 5.5. Figure 5.13 (a) and (b) clearly show the different information contained in the two channels. A n area of open water shown in the lower half of the image is wind roughened, which affects the V V backscatter, but not V H . The latter is mainly dark for open water areas. Sea ice in the scene is mostly ice that has survived one or more summers melt (see Table 5.5). In near range, the area just above Somerset Island 111 Table 5.5: Auxil iary data for the October 18 scene Auxil iary Data Value Comment Temperature -4°C Temperature increased com-pared to previous and following days Wind speed 40 k m / h Higher wind speeds compared to previous and following days Snowfall 7 cm Snowfall above 5 cm on October 5 and 18 Ice chart Mostly SYI and M Y I ; some Grey ice The ice chart for October 18 does not cover the area; Octo-ber 5 ice chart information used instead. Note: SYI : Second Year Ice; M Y I : Mul t i Year Ice appears dark in V V , which is likely new or young (grey) ice under these conditions [43]. A variation of V H in range can be observed for the open water area (Figures 5.13 (b) and 5.7 (a)). In far range, V H increases visibly. A similar but less pronounced increase is present in the first third of near range. These variations are not expected for V H backscatter of open water. They are rather the result of N E S Z variations due to antenna pattern correction combined with a S N R < 0 dB. The N E S Z and its variation in range are analyzed using a cross section of the scene in the open water area. 87 azimuth lines were averaged to reduce the effect of speckle in the cross sections shown in Figure 5.7. Spikes in the plot are mainly caused by ice signatures inadvertently included in the average. Cross-polarized backscatter of water and ice should have a small dependency on the incidence angle [76]. The variation observed for the October 18 IS4 scene is approximately 4 dB over the swath, with the highest values in far range. A drop-off in near range is present and probably caused by a change in S N R due to a signal level variation. 112 The likely reason for the variation of the N E S Z is the high receiver noise level relative to the signal level in combination with the antenna elevation pattern correc-tion. Elevation pattern correction is applied to compensate for low antenna gain on the beam edges, which results in raising the receiver noise. A S A R specifications state the worst case N E S Z for IS4 as -19.4 dB (see Table 5.2). These levels are not expected to be an issue for land applications. Low cross-polarized backscatter from smooth first year ice and open water may result in an S N R < 0 dB. Any variations in the N E S Z wil l affect both visual analysis and automated classification. The potential of the V H channel for sea ice / open water separation justifies the utilization of this information. To reduce the effect of the N E S Z variation over range, the utilization of the V H channel should be carefully implemented. Modified classification Assuming the worst case N E S Z for the IS4 mode of -19.4 dB, there would be little gain in using the cross-polarized channel as an additional source of information. Figure 5.13 (b) shows the potential of the cross-polarization information and an effort should be made to utilize this information. The variation of the N E S Z can actually be used to include V H in areas of low N E S Z (i.e., high SNR) in the classification scheme. Figure 5.10 shows the histogram of V H for the open water area. Based on the averaged range line, it is recommended that V H be excluded from classification for incidence angles larger than 34.75° when the signal level is low. The resulting N E S Z threshold is approximately -22.5 dB thus restricting N E S Z variations to about 2 dB. Using this framework, we propose a novel classification scheme to compensate for N E S Z variations over range. Our classifier involves the following steps: • Identify areas with known high N E S Z (in this case far range) and low signal level (less than 3 dB above the estimated noise level) in the imagery. 113 2500 HV [dB] Figure 5.10: October 18 acquisition. V H histogram of the open water area. Use two K-means clustering algorithms which use a Wishart classifier in parallel. The first algorithm is applied on the co-polarization channel only, the second one utilizes both available channels. A more detailed description of the K-means approach is provided in Section 3.3. Merge the two results after each iteration by utilizing the dual channel result in all areas where the cross-polarization channel is not restricted. In areas with known high N E S Z (in this case far range) and low signal level (less than 3 dB above the estimated noise level) use only the result from the co-polarization only classifier. Use the merged result for calculating the cluster centroids for the iterations. The centroids are calculated separately for the two channels. For the co-polarization 114 channel the entire marine area is used, whereas the cross-polarization channel is restricted. A graphical illustration of the classification scheme is presented in Figure 5.11. The method described above corresponds to Step 1. The noise level across the swath is estimated using the open water area in the scene. Operationally, open water areas may be available from nearby scenes. Also, the sensor characteristics are not expected to vary significantly over time so that only infrequent N E S Z estimates are required for each mode. The exclusion region (see Figure 5.12) is selected based on a V H threshold es-timated from the histogram (near range)and a 3 dB S N R (far range). In far range the 3 dB threshold is applied in an attempt to utilize the V H channel as much as possible, while considering the increase in N E S Z . A second, scene-specific step is added for this particular scene to identify the new or young (grey) ice in the image. This ice type is indicated by the low V V backscatter in near range. Ice charts suggest only two old ice types with some grey ice present in the scene. For Step 1, a four class approach does not work appropriately. The initialization of the classifier with three classes is therefore suggested in this case (analyst interaction). The initial classes are set up by first using the median value of the marine area for the cross-polarization channel to separate low and high backscatter pixels. The pixels with high cross-polarization level are further separated using the median co-polarization value. The K-means clustering algorithm which uses a Wishart classifier is run with three iterations. Both classifiers (single-polarization and dual-polarization) use the same class information (the merged result) for the update. Only the marine areas (open water and sea ice) are included in the process. The fourth class is determined in a second step, which is scene specific and described in more detail below. The interpretation of the final classes is a manual 115 Land mask ASAR AP (W+VH) Wonly fclassificationl NESZ estimate W & VH (classification! Wishart (1 channel) Cross-pol valid area Wishart (2 channels) Water + Gray Ice W & VH Iclassificationl Wishart (2 channels) Step 2 Result compilation Final Result (4 classes) Figure 5.11: Two-step classification concept. The second step is scene specific and separates new or young (grey) ice from open water. task and requires the expertise of an ice analyst. Classification results for the two-step method Figure 5.13 shows a number of classification results, all of which were derived using different methods. Co-polarization information only was used in the case of Figure 5.13 (c). Wind roughening of the water surface is the main cause for confusion of sea ice and open water. The cross-polarization information only was used to generate Figure 5.13 (d). The N E S Z variation clearly affects the result in far range, particularly in the lower half of the image. Figure 5.13 (e) uses the dual-polarization information for classification. It is apparent that the extra noise in far range ( V H level increase) has a negative effect on the result. 116 Incidence Ang le [°] Figure 5.12: V H exclusion region based on N E S Z estimate. 117 (c) VV (d) VH (e) VH+VV (f) Step 1 ; (VH)+VV (g) Step 2 Figure 5.13: October 18 IS4 acquisition. Backscatter intensities (a) V V ; (b) V H . Classification results (step 1) based on different inputs: (c) V V only; (d) V H only; (e) V V + V H , (f) V V + spatially restricted V H ; (g) Final result (step 2) 118 The result of the new approach (Step 1) is shown in Figure 5.13 (f). In far range the result is visually improved, although some class confusion remains. The reason for the latter is that both V H and V V show potential for confusion in far range. In the case of V H , the effect is system related, whereas V V is sensitive to a wind roughened water surface. None of the results discussed above identifies the new or young (grey) ice in the scene. A second step in the classification is therefore applied. In this step only the blue class from Figure 5.13 (g) (lowest backscatter) is considered and divided into two classes thus adding an additional class to the final result. Table 5.6 summarizes the backscatter statistics for the classes of the final result shown in Figure 5.13 (g). The mean cross-polarization level for all four classes is close to or below the worst case N E S Z of -19.4 dB posted for this swath. Compared to the N E S Z estimate shown in Figure 5.12, all classes are above this level for an incidence angle range between 31° and 34.5°. Table 5.6: Class average values for polarimetric parameters (mean: mean value; std: standard deviation) Class V V mean V V s t d V H mean V H std colour assignment •. . — v [dB] - • ' [ d B ] ' [dB] [dB] • —' , . 1 -9.28 0.90 -18.48 1.15 white 2 -10.99 0.77 -20.40 1.00 orange 3 -12.62 0.89 -22.95 1.26 blue 4 -18.47 . 2.87 • -23.89 0.80 grey. Average cross-polarization backscatter levels for all classes are reported at or below the posted worst case N E S Z for the swath. New or young (grey) ice (class 4) shows little difference to open water (class 3) for V H but is markedly different in . V V . V H provides more contrast than V V between Classes 2 and 3. The proposed 119 method allows the utilization of the cross-polarization channel in a large portion of the image. Detailed N E S Z information, in particular its variation over range, is required for successful classification. This information can be obtained from ocean scenes and is expected to need only infrequent updates. 120 5 . 4 D i s c u s s i o n o f R e s u l t s In this chapter, we have determined the extent to which multi year ice can be distin-guished in spaceborne multi-polarization S A R imagery that covers a full ice season. We have also proposed a method to mitigate a large variation of the N E S Z over the swath and thereby improve classification accuracy. Our findings are as follows. First, we have shown that spaceborne dual-polarization imagery can be used with some success to detect multi-year ice over a season in spite of N E S Z levels that reduce the cross-polarization contrast between multi-year ice and first-year ice. We have successfully identified multi-year ice in eight out of ten images acquired over a full season. In one case, multi-year ice is overestimated. In another scene, multi-year ice identification failed because breakup occurred and surface melt severely affects the backscatter. The contrast between M Y I and F Y I is about 0.5 dB higher for H H than for H V , as the latter is affected by the N E S Z . A comparison of the classification results for three scenes with very similar ice types but varying air temperature and snow layer shows that 89% of M Y I classified in one scene are classified as M Y I in the second scene if both scenes were acquired in freezing conditions. This level drops to 67% if one scene has a wet snow layer. The latter value is underestimated due to a breakup of part of the ice in the scene. Second, we have assessed the impact of N E S Z variation due to antenna gain correction and developed an adaptive classification scheme that compensates for such variation. We have developed an adaptive classifier that compensates for high N E S Z areas in the image. Based on a N E S Z estimate, we switch between two K-means clustering algorithms both of which use a Wishart classifier. One algorithm uses dual-polarization imagery; it is applied in areas with acceptable N E S Z . Wherever the N E S Z exceeds an acceptable range, classification is based on the second algorithm which uses single-polarization data only. The method was tested on A S A R A P IS4 imagery with about 4 dB N E S Z variation. - It increases the utilization of the cross-polarization channel and provides improved classification over the standard dual-121 polarization classification approach we proposed. Three out of four classes of the final result show cross-polarization levels that are between 1 dB and 4 dB below the posted worst-case N E S Z for the scene. 122 Chapter 6 Conclusions and Recommendations 6 . 1 C o n c l u s i o n s The imminent launch of R A D A R S A T - 2 , the most advanced of the second-generation spaceborne SARs, has stimulated renewed interest in the use of multi-polarization and polarimetric S A R imagery for sea ice monitoring. The primary objective of this work is to assess the potential value of R A D A R S A T - 2 multi-polarization and polarimetric C-band S A R imagery for classification of sea ice and to develop improved classifiers that account for the characteristics of such imagery. The outcomes of this thesis are as follows: Our comprehensive ice information requirements review, the first of its kind since pre-launch preparations for R A D A R S A T - 1 in the early 1990's, shows that: (1) the Canadian Ice Services' primary requirement is for information concerning ice edge location, ice concentration and ice stage of development, (2) accurate classification of ice types in S A R imagery wil l address the need for such information and (3) daily revisit is the single most important criterion for using S A R imagery and classification maps derived from such imagery in Canadian Ice Service operations. Second, we have shown that both the Freeman-Durden and Cloude-Pottier tar-get decomposition methods provide a consistent physical description of the scattering 123 processes and allow the identification of ice types in airborne A I R S A R imagery. Our analysis of the C-band A I R S A R imagery using the Freeman-Durden decomposition suggests that surface scattering dominates the backscatter of first-year ice. It is also a strong component for all other ice types observed. This is supported by our analysis of the data in the H/A/a-space, where we observe a single cluster in the low entropy (H<0.5) and medium entropy (0.5<H<0.9) surface scattering classes (a < 40°). Our estimate of the volume scattering component indicates that it dominates only for some portions of compressed first-year ice, multi-year ice and thin ice. The contribution of dihedral scattering is low and of little value for classification. Pixel by pixel classifi-cation using the dominant scattering mechanism according to Freeman-Durden leads to a salt and pepper appearance of the three class results with the density of volume scattering pixels increasing for compressed first-year ice, multi-year ice and thin ice. Spatial filtering the data reduced the number of pixels with predominantly volume scattering. This decreases the classification potential of the data. Pixel by pixel clas-sification based on a partitioning of the H/A/a-space allows thin ice, .first-year ice, and multi-year ice to be distinguished with reasonable success. Third, we have used our insights concerning the scattering mechanisms that con-tribute to backscatter from sea ice to develop a simpler method for initializing unsu-pervised clustering algorithms, and demonstrated this using our K-means clustering algorithm which uses a Wishart classifier. Following our observation that surface scattering in S A R sea ice imagery is generally strong, we reduce the complexity for the initialization from target decomposition to a simple backscatter strength-based method. The leads to faster execution without apparent penalty in classification ac-curacy. Unlike target decomposition based initializations, this method can be used for polarimetric, dual-polarization, and single-polarization imagery. A comparison of classification results based upon polarimetric and dual-polarization imagery with a reference classification based upon dual-frequency polarimetric imagery yields overall classification accuracies of about 66% for both. We have defined a classification con-fidence measure for the Wishart classifier to eliminate classification uncertainties in the reference solution. Using this refined reference, the overall classification accuracy 124 increased to more than 80% for both polarimetric and dual-polarization imagery. The overall difference between the two classification results is less than 3%. Polarimetric imagery allows target decomposition algorithms to be applied to cluster centroids to help identify ice types. Dual-polarization imagery has a reduced data volume and wil l be available in wide swath modes, which are important for operations. The classifica-tion accuracy achieved using cross-polarization imagery alone is surprisingly high with only about 7% reduction compared to polarimetric imagery. The N E S Z affects low backscatter classes for this channel, only about 20% of thin ice is correctly identified. For results based on single-polarization imagery alone, overall classification accuracy falls to about 56% (full dual-frequency result) or about 70% (refined reference). Fourth, we have shown that the difference in resolution between high-resolution airborne and simulated spaceborne polarimetric imagery caused more degradation in ice type separability for the spaceborne imagery than the increase in N E S Z . Eighteen times more multilooking is applied to the airborne data for the same output pixel size. However, we have shown that spatial filtering of the spaceborne data can help to compensate for the lower speckle reduction. We have assessed classification ac-curacy of the simulated spaceborne data by using the result based upon the highly multilooked airborne data as a reference. Four first-year ice types, young ice and leads were present in the imagery. Data averaged to nine looks yielded a classification ac-curacy of about 60%. B y applying an edge preserving spatial filter of window size five, the classification accuracy increased to about 70%. No further improvement in classification accuracy was achieved by increasing the filter window to seven. While speckle and speckle reduction affects all classes, we have shown that the increase in N E S Z (-40 dB for the airborne imagery versus-28 dB for the simulated spaceborne imagery) only degrades the signatures of classes with low backscatter. Backscatter at or below N E S Z level were measured at N E S Z level and class separation was reduced. For our imagery, one class (leads) showed less than -40 dB signal level for the airborne cross-polarization channel. The simulation reduced the contrast to other classes by increasing the N E S Z . In our example, however,, the contrast reduction was not severe enough to significantly degrade class separability. 125 Fifth, we have shown that spaceborne dual-polarization imagery can be used with some success to detect multi-year ice over a season in spite of N E S Z levels that reduce the cross-polarization contrast between multi-year ice and first-year ice. We have analyzed the first full-season E N V I S A T A S A R A P imagery covering a full ice season. CIS experts had identified 10 scenes out of 16 acquisitions where multi-year ice is present. We have analyzed these scenes and, according to CIS feedback, correctly identified multi-year ice in eight of them. In one failed case, multi-year ice was overestimated to a degree that CIS feedback was negative. For the second failed case, severe surface melt was present with breakup occurring during acquisition. Melt and breakup are known problems for ice type discrimination. We have analyzed three scenes of the set in more detail. These scenes, acquired under varying environmental conditions, show stable ice conditions, which allows a comparison of classification results. For two scenes acquired in freezing conditions, multi-year ice is consistently estimated with almost 90% accuracy. This level drops to about 70% when a wet snow layer is present. Our analysis of the signature contrast between first-year ice and multi-year ice shows that a wet snow layer causes a contrast reduction for both channels. In all three scenes analyzed, the cross-polarization channel contrast is degraded by the N E S Z . Previous research indicates higher contrast for the cross-polarization channel compared to the co-polarized channel. This is not the case for these images analyzed and suggests some limitation for spaceborne dual-polarization acquisition modes with N E S Z levels of around -20 dB. Sixth, we have assessed the impact of N E S Z variation due to antenna gain cor-rection and developed an adaptive classification scheme that compensates for such variation. As we have shown that, the cross-polarization backscatter sea ice mea-sured with spaceborne dual-polarization acquisition modes is affected by N E S Z . The N E S Z variation over the swath due to antenna pattern correction is then an issue. Because posted N E S Z values usually refer to the worst case N E S Z , which only occurs for a small portion of the image, the imagery wil l therefore have a better S N R than anticipated for a large portion of the swath. If the cross-polarization signature is at or below N E S Z , any variation of the N E S Z wil l have a negative impact on data 126 analysis. We have developed an adaptive classifier that compensates for high N E S Z areas in the image. Based on a N E S Z estimate, we switch between two K-means clustering algorithms both of which use a Wishart classifier. One algorithm uses dual-polarization imagery; it is applied in areas with acceptable N E S Z . Wherever the N E S Z exceeds an acceptable range, classification is based on the second algorithm which uses single-polarization data only. Our test case was an E N V I S A T A S A R IS4 image with about 4 dB N E S Z variation. Using our method we show improved ice water separation. Three out of the four classes of the result have cross-polarization backscatter levels below the posted N E S Z of -19 dB. The example also illustrates the utility of cross-polarization information for ice water separation, when a wind roughened water surface reduces the ice water contrast for the co-polarized channel. 6.2 Implications of Our Results The results of this study have important implications for operational use of multi-polarization and polarimetric S A R imagery in sea ice monitoring. 1. Our two improvements to sea ice classification algorithms account for the unique characteristics of sea ice imagery: • The first simplifies initialization of unsupervised classifiers and thereby simplifies classifier use • The second accounts for N E S Z variations over the swath and thereby im-proves classifier accuracy 2. We have shown that for sea ice monitoring, both dual-polarization and po-larimetric S A R imagery offer similar improvements in classification accuracy compared to single-polarization imagery. However: • The significantly smaller data volume associated with dual-polarization imagery reduces both transmission bandwidth requirements and the pro-cessing load. 127 • Dual-polarization is supported in ScanSAR mode which allows the require-ments for higher classification accuracy and frequent revisits to be met simultaneously. Thus, we are able to recommend with confidence that the Canadian fee Service use dual-polarization ScanSAR data from R A D A R S A T - 2 as their standard data prod-uct and that they adopt our three improvements to sea ice classification algorithms in their R A D A R S A T - 2 data processing stream. 6.3 Recommendations for Future Work Following the success in the research phase, the next logical step is an operational trial at the Canadian ice Service. This wil l require that: • Existing CIS workstations be upgraded for storing and processing multi-polarization and polarimetric imagery. • A work flow be developed that integrates the methods in an existing operational scenario. • Classification algorithms be optimized for fast execution and handling of large images. • Information on the N E S Z and its variation over the swath be obtained for each acquisition mode used. Although estimating N E S Z levels from the data is not practical for operations, the data provider should be able to provide this information as part of the data product. B y processing a large amount of imagery covering several geographic areas, the CIS wil l have the opportunity to comprehensively test the accuracy of the algorithm 128 and its robustness against seasonal changes. Such a large data collection wil l facil-itate the further development of improved sea ice classifiers. For example, accurate knowledge of the N E S Z characteristics in combination with a representative test data set wil l allow the development of an improved mitigation strategy. 129 Appendix A Properties of Sea Ice Sea ice is part of the cryosphere, which interacts with the underlying oceans and the overlaying atmosphere. Ice plays an important role in the global climate system. Firstly, sea ice is an effective insulator between ocean and atmosphere, restricting exchange of heat, mass, momentum, and chemical constituents. During winter, the flux of oceanic heat to the atmosphere from open water can be two orders of magnitude larger than the heat flux through an adjacent thick ice cover. Sea ice also affects the surface albedo (i.e., the fraction of incident electromagnetic radiation reflected by a surface). A n ice-free ocean generally has an albedo below 10-15%. Sea ice albedo, on the other hand, can vary between 20% (melting ice) to as high as 98% (fresh snow layer). A high albedo reduces the amount of solar radiation absorbed at the earth's surface; the effect is most significant in summer, when the solar heating is high [94]. A . l Sea Ice Stages of Development The formation of sea ice is a complex process and begins at -1.8°C in typical sea water. The freezing process is influenced by thermodynamic processes; the effect of salt contents as well as mechanical forces caused by wind, waves, and ocean currents. In addition to the frozen sea water, snow layers are likely to be present [51]. Examples. of sea ice in various stages of development are shown in Figure 1.1. 130 Table A . l : Sea ice stages of development. Several examples of sea ice in various stages of development are shown in Figure 1.1 \ Main ice type Sub Types Thickness Remarks New Ice Frazil Ice crystals Grease Soupy layer Nilas < 10 cm Elastic crust Young Ice Grey 10 — 15 cm Ridging and Grey-White 15 — 30 cm Rafting due to waves First Year Ice Th in F Y I 30 — 70 cm One winter's growth Medium F Y I 70 - 120 cm Thick F Y I > 120 cm Old Ice — meters More variety in topography New Ice When the temperature of both air and ocean is low enough, small ice crystals are formed at the ocean surface, called frazil ice. As the freezing continues, more and more ice crystals are forming a soupy layer called grease ice. Under calm conditions, this layer grows to an elastic crust, called nilas if it is less than 10 cm thick. Depending on the thickness of the layer, nilas can be further sub categorized into dark nilas (<5 cm) and light nilas (5 - 10 cm) [94], [112].Young, high salinity ice has a thin film of very high salinity surface brine. Under cold conditions (air temperatures below -5°C), this film supports the growth of so called frost flowers. These structures can reach a few centimeters in height and significantly change the surface roughness and electromagnetic properties of the ice surface [51]. 131 Young Ice Nilas loses its elasticity with increasing thickness and tends to break under swell. Ice with a thickness between 10 cm and 30 cm is generally referred to as young ice. The subcategory grey ice means ice in the range between 10 - 15 cm, grey-white ice stands for 15 - 30 cm thickness. W i n d and waves often break up the ice layer and form small rounded floes with diameters ranging from 30 cm to 3 m. These floes are known as pancakes. Continuous collisions between floes generate raised rims with a typical height of 5 - 20 cm. The rims of the pancakes are efficient scatters of radar waves. The process of brine drainage gradually replaces the brine pockets with voids of air, which change the visual appearance of the ice from almost black nilas to bright grey and grey-white ice. Ridging of the grey ice and rafting of the grey-white ice under pressure is responsible for topographic variation [94], [112]. First Year Ice (FYI) Sea ice with a thickness of more than 30 cm is called First Year Ice as long as it is only of one winter's growth. Depending on the thickness, it too can be sub categorised: Th in F Y I (White Ice) (30 - 70 cm), Medium F Y I (70 - 120 cm) and Thick F Y I (>120 cm) [112]. Old Ice Old ice, also referred to as Mul t i Year Ice (MYI) , has survived one or more summers melt. The melting process causes the surface of old ice to be smoother than for F Y I ; however, melt ponds and hummocks are also responsible for much more variety in the topography. Ice formed during one winter season (FYI) typically contains from 4 to 20%o salt (lower salt content in the bulk than in the ice-air and ice-water interfaces) if the ice is formed from normal sea water with normal salinity of about 35%o. The brine gradually drains from the upper part of the ice, causing multi-year ice to have a salinity of near 0 in the surface layer, and about 2-3%o in the bulk [51]. 132 Deformation Features The ice pack is usually in constant motion because of wind and currents. This can result in considerable deformation of the ice cover as well as the formation of floes surrounded by water or brash ice (a mixture of pulverized floes and new ice, which is interspersed with broken blocks). Floe sizes can range from several kilometers in diameter in the central ice pack to less than a meter in the marginal ice zone [54]. Ridges are formed when the ice sheet is compressed and sheared causing blocks to pile up above (sail) and below (keel) the surface. Most of the ridge ice is submerged; however, sail heights can exceed 10 meters. Repeated ridging causes rubble fields. Rafting of ice sheets is also a common occurrence, thus multiplying the ice thickness in affected areas. Ridges, rubble and rafts are obstacles for transportation, be it on the ice surface or by ship [54]. Snow Pack The snow cover, which is present for most of the year, is an important surface feature of sea ice. The snow acts as insulator in the winter and its high albedo reduces the impact of sunlight on the ice surface. During ice decay, snow is also a key factor. Snow drifts are surface features that can reach heights of more than one meter [54]. Ice Decay After the winter season, when temperatures rise, an increase in air temperature and solar input causes the decay of the ice cover. Three stages of decay are defined, "early melt", "melt onset", and "advanced melt" [51]. The snow layer on the ice plays an important role in all three phases. In the early melt season, the amount of free water within the snow cover increases. For melt onset, the free water content exceeds 3% with an accumulation of the free water at the snow-ice boundary. As a result, the snow surface becomes increasingly 133 rough. In the advanced melt stage, ice decomposition occurs. On the surface, melt ponds start building, in many cases, the water wil l eventually be drained through drainage channels, which weaken the strength of the ice sheet. [51]. A . 2 E lectromagnet ic Proper t ies of Sea Ice Sea ice is an inhomogeneous medium composed of an ice background, brine inclu-sions, air bubbles and solid salt. Physical properties are further complicated by surface effects such as roughness, snow, slush, and brine covers. Wave propagation, attenuation, and scattering are interrelated and linked to the physical and structural properties of sea ice [77], A n investigation of sea ice using microwaves therefore re-quires an understanding of the electrical properties of ice and snow. These properties in combination with the surface roughness wil l significantly influence the signal and are key for correctly interpreting the measurements. Dielectric Properties of Sea Ice The electrical properties of a medium are characterized by the relative permittivity er = 4 + je'l ( A . l ) where e'r is the dielectric constant and e" is the dielectric loss [51]. The relative permittivity is defined with respect to the permittivity of free space, eo-The dielectric properties of seawater and brine are functions of salinity, tem-perature, and microwave frequency. Scattering from open water takes place on the air-water boundary. The backscatter strength increases with increasing roughness of the water surface due to wind. New, gray and pancake ice are characterized by a high salinity, because in the ini-tial formation brine is retained. A surface scattering component can be expected from the ice-air (or ice-snow) boundary. The surface layer of new ice consists of randomly 134 oriented crystals. The layers below generally show a preferred orientation of crystal growth resulting in a columnar ice layer [51]. Brine inclusions are inhomogeneities in the material resulting in volume scattering. These inclusions are ellipsoidal and preferably vertically oriented. Due to the shape and orientation of the brine inclu-sions, the effective permittivity of sea ice is uniaxially anisotropic. The high salinity of young and first year ice causes the attenuation of secondary scattering events. As the ice ages, brine is drained resulting in air filled voids and drainage channels. W i t h a decreased salinity, the attenuation of secondary scattering processes is also decreased. The lowest salinity levels are present in old ice resulting in the lowest dielectric loss [51]. A i r bubbles are also more rounded that brine inclusions, old ice is therefore more isotropic [77]. The dielectric constant e'r for freshwater ice has been extensively studied for frequencies around 10 GHz . It is constant at 3.17. e'r for saline first year ice is higher, typical values are 4.6, it is also highly temperature dependent. For old ice e'r is lower and much less temperature dependent, typical values of 3.7 are reported [51]. In a simplified numerical model for sea ice, e'r is proportional to the relative brine volume, where H < 70%o . This model is an approximation at 4 G H z [49]. The dielectric loss e" for sea ice depends on the.temperature as well as on the salinity [51]. Due to this variability, no typical values were found in the literature, rather, empirical relations and graphs are provided [49], [51], [71]. For sea ice, e" can be modelled proportional to the relative brine volume, Vj, where VJ, < 70%o . This model is an approximation at 4 G H z [49]. For saline snow, empirical results support the relationship e ; = 3.05 + 0.0072 • Yb (A.2) e" = 0.02 + 0.0033 • Vb (A.3) 4(S,T,f) = e"(T,f) + D(T,f)-S (A.4) 135 Table A.2: Example values for the dielectric constant e'r. Example relations for the dielectric loss are provided in the text. Example < Comment Ai r 1 constant Water 80 constant Freshwater ice 3.17 constant (measured at around 10 GHz) Saline F Y I 4.6 highly temperature dependent Old Ice 3.7 less temperature dependent where e" (T, / ) is the dielectric loss of pure ice. D(T,f) is a constant of proportionality, which may depend on temperature, T and frequency, f. S is the salinity [71]. Spatial and temporal variability of the layer composition makes modelling diffi-cult as few observable parameters are available to model a multi-layer environment. A composite model to calculate effective permittivities and backscatter covariance matrices at microwave frequencies is described in [77]. The model was subsequently compared with real data; generally, good agreement between measurement and model was achieved [78]. Using this model the potential of E N V I S A T A S A R imagery for ice type separation and ice-water discrimination was analyzed [76]. Propagation in Lossy Media A n electromagnetic wave travelling in z-direction can be described as ' E(z) = E0 • . (A.5) Eq is the field intensity at z = 0. The propagation constant for the medium, 7, is given by 7 = 'a + j0. (A.6) 136 The absorption constant a describes the transformation of energy (=loss). The phase constant 8 is equal to the wave number in a lossless medium (k = 2tt/\). a and 3 are related to e r a = kG\lm{^Tr}\ (A.7) 8 = k0Re{^Tr}. (A.8) k0 is the wave number of free space [49]. A n alternative relationship between a and e can be stated as [51] a = 868.6*0^ j ! ^ l + t a n 2 ( 5 ) - l ) } 1 / 2 (A.9) where the loss tangent tan 6 is given as tan <5 = ( A - 1 0 ) The total electromagnetic loss in a scattering medium is described by the extinc-tion coefficient Ke, which is the sum of the power absorption coefficient Ka and the scattering loss ks Ke = Ka + ks . (A.11) The scattering loss is caused by inhomogeneities in the medium [49]. Part of the incident wave is transmitted into the medium, the penetration depth 5P indicates the maximum depth of the medium that contributes to the backscatter. Sp is defined as the depth at which ^ 2 L _ l . ( A 1 2 ) Backscatter from Sea Ice In microwave remote sensing, three different scattering types can be differentiated: 1. surface scattering - occurring at a boundary of two materials with different electromagnetic properties. For backscatter to occur, the surface has to be rough with respect to the radar wavelength. Rough surface scattering models are generally regarded as single bounce approximations [49]. 137 2. volume scattering - upon entry in a material of different electromagnetic proper-ties an incident wave wil l be refracted. Inhomogeneities in the material such as brine pockets and air bubbles wil l cause direct and secondary scattering events forming the resulting backscatter. Volume scattering has a depolarizing effect. 3. double bounce - in the presence of certain geometrical shapes it is possible for a radar signal to be scattered back to the antenna after two scattering events on the ground (e.g., dihedral). Ridges or rubble may provide the geometry required for double bounce. Sea water and brine are electrolytic solutions, sea ice and snow are heterogeneous mixtures. Sea ice is a mixture of ice crystals, air pockets, and brine inclusions. Snow is a mixture of air, ice particles, and liquid water inclusions. Together, they form a complex multi-layered target. Depending on the environmental conditions and the radar frequency used, all layers and the layer boundaries may contribute to the reflected signal [49]. 138 Appendix B Ice Chart Production at the Canadian Ice Service The Canadian Ice Service (CIS) / Ice and Marine Services Branch of the Meteorolog-ical Service of Canada (MSC), is the leading authority for information about ice in Canada's navigable waters [40].. It provides a number of services to clients and sub-scribers; these include ice bulletins and ice analysis charts (http://www.cis.ec.gc.ca/). Figure 2.1 shows a sample ice analysis chart. Information on the ice situation is pro-vided twofold; firstly encoded in colour for easy and fast readability and secondly in form of egg codes for more detailed information. Both codes are described in the Manual of standard procedures for observing and reporting ice conditions [112]. Of all sources used at the CIS for ice chart production, two are most important: 1. The visible and infrared N O A A Advanced Very High Resolution Radiometer ( A V H R R ) and 2. R A D A R S A T - 1 A V H R R provides ice forecasters with daily visible and infrared imagery over the O S ' s areas of interest. Cloud cover frequently reduces the portion of the imagery useful operations. R A D A R S A T - l ' s ability to provide useful imagery regardless of 139 solar illumination or cloud cover at a relatively high resolution is why it has become the primary sensor for ice detection at the CIS ever since imagery has become available operationally [40]. At present, human expertise is key for ice chart production at the CIS. Trained image analysts use the most recently acquired satellite imagery and all other infor-mation available (this includes previous day ice analysis charts and model outputs) to generate the daily ice analysis chart. Automated methods have not yet proven to be consistent and reliable enough to be used operationally [39]. W o r k F l o w for Ice A n a l y s i s C h a r t P r o d u c t i o n The following is a general timeline and flow of the nominal daily data analysis, product generation, and distribution scenario [40]: • CIS Numerical Ice Models are run overnight using output from the 0000 U T C run of the Canadian Meteorological Centre ( C M C ) numerical weather prediction runs, plus gridded ice data extracted from the previous days' Ice Analysis chart (a compilation chart of all available data sources valid for 1800 U T C ) ; ice model output fields (ice concentration, trajectory, etc.) forecast over the next 48 hours are available at around 1200 U T C . • The Image Analysis charts (a detailed analysis chart of a particular image valid for the time of image acquisition) from the previous nights' passes, plus the previous days Ice Analysis chart, the latest C M C model outputs, current mete-orological information, and observed charts and reports (from aircraft or heli-copter reconnaissance, ship or shore reports) are available to the Ice Forecaster when they arrive in the morning (1130-1230 U T C ) to begin their analysis in preparation for issuing the daily bulletins products • R A D A R S A T - 1 imagery from the morning pass is processed and available in the ISIS Catalogue (depending on region) from 1200-1500 U T C . 140 • As R A D A R S A T - 1 imagery (and other if no R A D A R S A T - 1 ) become available on ISIS, the Image Analyst performs a detailed analysis of the latest imagery in conjunction with the previous daily Ice Analysis chart, meteorological data, and the observed charts and reports; completed Image Analysis charts are automat-ically catalogued and made available and/or distributed directly to clients and subscribers (1300-1800 U T C ) as well as made available to the Ice Forecaster. • A t 1500 U T C , with all data available at that point, the Ice Forecaster issue text ice bulletins and forecasts • R A D A R S A T image analysis completed by 1800 U T C are used by the Ice Fore-caster in conjunction with all other available imagery and data to prepare the daily Ice Analysis chart which are issued from 2000 to 2300 U T C (depending on the region) Products resulting from the daily production flow described above are auto-matically made available or distributed electronically via the Product Distribution System. 141 Appendix C Ice Classification Using Single-Polarization S A R Imagery Remote sensing plays a key role in sea ice monitoring. Airborne remote sensing was followed by spaceborne remote sensing when polar orbiting satellites became available in the early 1970s. Initially, optical images in the visible and near infrared spectrum were acquired. While optical imagery is still in use, data analysis is limited by cloud cover during the summer and the lack of solar illumination during the polar winter. Microwave remote sensing, which had already been used for airborne reconnais-sance, was the natural next step i n sensor development. Firstly, passive microwave radiometers were deployed; current examples are the series of Special Sensor M i -crowave Imager (SSM/I) instruments operated by the US Defense Meteorological Satellite Program. These instruments are still affected by water vapor and cloud water content [121.]. Active microwave sensors include Side Looking Radars (SLR) and Synthetic Aperture Radars (SAR). The latter provides high resolution imagery and has been successfully deployed on spaceborne platforms. Examples for spaceborne SARs are S E A S A T (USA), E R S l / 2 (European Space Agency, E S A ) , J E R S (National Space De-velopment Agency of Japan, N A S D A ) , KOS-MOS-1870 and A L M A Z - 1 (USSR), and 142 R A D A R S A T - 1 (Canadian Space Agency, C S A ) . A l l afore mentioned sensors are capa-ble of acquiring images in one single-polarization. They provide imagery independent of daylight and cloud cover. Several of these sensors are not longer in operation. The formation of S A R images is not subject of this work, a comprehensive overview on this topic is given in [24]. The two most important first generation SARs were ERS-1 and R A D A R S A T - 1 . ERS-1 was the first spaceborne S A R to reliably provide large amounts of imagery. Despite limited coverage (100 km swath width), an important milestone was set for large scale applications. In 1995, Canada launched R A D A R S A T - 1 , the most advanced spaceborne S A R of its time. Unlike ERS-1 , with its multiple scientific goals, one specific objective of R A D A R S A T - 1 was to monitor ice covered Canadian waters. Several data acquisition modes are available, including narrow swath, high resolution modes, as well as a Scan-SAR mode providing 100 m resolution over a 500 km swath. The latter ensures a high repeat imaging capability in arctic regions. Single-polarization spaceborne S A R imagery is the primary data source for sea ice monitoring to date. Although imagery has been available for more than 10 years, no fully automated system is in place to generate ice charts for navigation. Research is underway around the world to improve and modernize existing ways of ice chart production. C . l C a n a d a A t present, no automated classification used for operations at the Canadian Ice Ser-vice [41]. A number of research projects is underway to develop machine-based meth-ods in support of ice analysts. Computer vision is one area of interest for researchers. Several methods are under investigation, for example, a three step sea ice boundary detection algorithm was developed. While the method shows some promise, it is com-putationally expensive and has therefore not been tested extensively [17]. Another method uses texture and tonal features to separate open water from sea ice and even 143 to separate sea ice types [18]. Gray-level cooccurrence probabilities and Gaussian Markov random fields are under investigation for texture analysis [19].' While the methods have not yet been tested in operations, latest results are well received by the CIS [3]. These efforts are preceded by a number of research projects carried out in Canada. Sea ice is an integral part of the Canadian environment, the decision to launch R A D A R S A T was partly made because of the need for continuous observa-tion. Much of the research conducted in the late '80, early '90 was in preparation for R A D A R S A T - 1 . The following references are some representative examples showing how Canada became one of the leading nations in S A R sea ice monitoring; the list is by no means complete [51] [67] [66] [8] [7]. G.2 E u r o p e a n U n i o n Several European states have interest in sea ice monitoring, there is a multitude of national and Europe wide projects underway. The Brit ish National Space Centre funded the development of an Ice Pilot A p -plication Project (IPAP) which was developed by industry in collaboration with the Scott Polar Research Institute. The project includes ice classification, ice motion estimation and iceberg detection based on S A R imagery. Data classification is based on either a K-nearest neighbor (K-NN) or a neural network multi-layer perceptron ( M L P ) . The process is supervised, an analyst is required to select the classification method and set parameters. A demonstrator was developed in 1994. No evidence was found that an operational system was ever implemented [111]. The Finnish Institute of Marine Research (FIMR) provides sea ice information for navigational purposes in the Baltic Sea. Only First Year Ice is present in the area of interest and most of the ice is drift ice. Ice conditions in the region can change rapidly, depending on environmental conditions. Spaceborne S A R images are the most important remote sensing source for F I M R . A n operational sea ice 144 classification system is currently under development. The classification is based on a modified version of the pulse-coupled neural network ( P C N N ) . The system uses 100 m R A D A R S A T - 1 ScanSAR Wide mode imagery, which are not absolutely calibrated but have an incidence angle correction applied. Six classes are used for classification, these are based on an extensive training data set. The classification was built for and trained with data from the Baltic Sea only. A l l training scenes were acquired in freezing conditions, thus limiting the applicability of the classifier. While comparisons of automatically generated results with ice charts show some promise, the method is not error free and has yet to be tested for wet snow or melt conditions. Generally, the method uses S A R imagery only as input, the use of ice history information is being investigated [57]. The European Union funded the I C E R O U T E S project, a cooperation between researchers and end users of several E U member countries as well as Norway and Russia. The overall goal of the research project was to demonstrate the feasibility of an ice routing tool for safer and more efficient ship transport in ice-covered waters. A series of classification methods was tested, including K - N N as well as Growing Neural Gas (GNG) and Gabor Wavelet features. The study included S S M / I and wind data in addition to spaceborne S A R imagery. Study results show classification accuracy at about 70%, with the best results achieved using all three data sources rather than S A R only. Classifiers were optimized for training data but did not perform well on validation data independent from the training data [32]. The latest European effort in high latitude oceanography in general and sea ice monitoring in particular is I C E M O N . "The overall objective I C E M O N is to imple-ment a coherent European operational oceanography system for the high latitudes, consisting of sea ice, meteorological and oceanographic services" [53]. This program has a much larger scope than looking into the exploitation of S A R imagery; a multi-tude of spaceborne imagery is considered as data sources and infrastructure as well as data standards are being considered. I C E M O N is an effort to co-ordinate all national efforts in Europe and provide standardized data and products. Ice analysts are used in ice chart production based on S A R imagery [53]. 145 C.3 United States of America Another ambitious development was conducted for the National Ice Center (NIC) in the United States. The "Advanced Reasoning using Knowledge for Typing Of Sea ice" ( A R K T O S ) image analysis system was developed over several years. A R K T O S is a feature based, fully automated intelligent classifier that is based on a rule-based system in combination with a modified Dempster-Shafer belief system. The system is limited to four classes; open water, new ice, first year ice, and multi year ice [116]. The system uses multiple data sources. In addition to R A D A R S A T - 1 images, ice concentration maps based on S S M / I , land masks and ice climatology data (based on a 19-year history, the probability of the presence of sea ice and the median ice concentration are generated) are also used. Data classification is a three step process: (i) Image processing: Following a preprocessing step where all data are georefer-enced and the S A R imagery is filtered, the S A R image is segmented using a watershed merging algorithm, which identifies intensity based homogeneous regions as features. For each feature, a total of 25 attribute measurements are calculated. The attributes can themselves be divided into four sub-categories (intrinsic, boundary related, tex-ture based, and geometric). Based on the feature attribute measurements, which are real-valued numbers, 16 discrete facts are generated on which the rule based classifi-cation is based. (ii) Multi-Source data fusion: Ancillary data are used in the reasoning process, the fusion framework is attribute and fact based. A l l information is geo-located. (iii) Rule-based classification: Finally, a set of 100 rules was developed based on extensive interviews with ice analysts followed by refinements in the testing phase. A modified Dempster-Shafer Belief system is used.to calculate the most credible ice class. The authors of [116] state that their system provides evidence that an automated system can be developed for near real time sea ice classification. For a brief period in the early 2000's A R K T O S was evaluated by ice analysts at NIC in parallel with 146 operations. More recent information indicates that this is not longer the case. The CIS assisted in the evaluation of A R K T O S but did not consider the system as suitable for use in their operations [3]. C.4 Discussion of Current Systems At present, a significant amount of research effort is spent on the automated analysis of S A R imagery for sea ice monitoring. The ultimate goal of fully automated ice chart generation may not be reached in the near term; however, automated classifiers are expected to at least assist human experts in operations in the future. Operational requirements need algorithms to be efficient. In most cases, this has lead to the development of simple algorithms that use a limited number of classes. Shortcomings that should be addressed in future research include: • A l l methods appear to focus on the image aspect of S A R (i.e., texture, edges) only. A Radar is an active measuring device, the backscatter, if properly cal-ibrated, does tell much about the target. A tremendous effort goes into the investigation of microwave signatures of sea ice [81]. The backscatter magni-tude information is currently under-utilized. A combination of computer vision and backscatter information wil l increase the information content available for classification. • S A R imagery, while proven extremely useful for sea ice monitoring, show short-comings that can only be addressed by including data from other sources. Ice analysts at the CIS utilize ancillary data when generating Ice Charts. A R K -T O S , I C E R O U T E S , and I C E M O N utilize ancillary data. W i t h the exception of I C E R O U T E S and subsequently I C E M O N , all methods use only S A R im-agery directly. A l l other information is derived from remote sensing data. In addition to an increasing number of SARs, other spaceborne missions provide information on sea ice. Progress in data fusion methods wil l allow the combi-147 nation of multiple data sources with different spatial resolutions in the future. The I C E M O N effort is going exactly in this direction. • In an operational system, the use of previous results as well as ice modelling results is possible and should be considered. Ice motion products can also provide valuable information. • The human aspect (i.e., the availability of an ice analyst) is largely ignored. Semi-automated methods developed in support of a human expert based ice chart generation system wil l be more useful in the short and medium term. • Multi-polarization and polarimetric S A R imagery wil l increase the information available. One aspect that has significance for the development of semi-automated systems is the increasing number of spaceborne SARs , all with increased capabilities. This naturally leads to an increase in the number of data acquisitions for a particular area. Frequent revisit is one key requirement for the CIS, for example. Having several sensors available, it is also possible to plan acquisitions in a way that full coverage of an area with a ScanSAR mode is guaranteed while high resolution imagery (or other new modes like multi-polarization imagery) can be planned for a specific, smaller area of interest. 148 Appendix D Polarization Diversity in S A R Imaging Polarization diversity in radar systems has been recognized as producing a more com-plete inference of natural surface parameters than is possible with a single channel radar system. Conventional imaging radars operate with a single, fixed-polarization antenna for both reception and transmission of radio frequency signals. In this way a single scattering coefficient is measured (for a specific receive and transmit po-larization combination) for each resolution element in a scene. As a result of this implementation only one component of the scattered wave, itself a vector quantity, is measured resulting in a scalar characterization of the wave. Any additional infor-mation about the target surface contained in the polarization is therefore lost. To ensure that all information in the scattered wave is retrieved, the polarization of the scattered wave must be measured through a vector measurement process [126]. D . l P o l a r i z a t i o n For a uniform plane wave travelling in positive fc-direction, the electric field vector E(t) must lie in a plane perpendicular to the k-axis [123]. For problems involving 149 the earth surface the electric field is best described using a horizontal/vertical base. This base, h and v, is perpendicular to k and h is parallel to the earth's surface (D. l ) The horizontal and vertical components Eh and Ev are complex with amplitudes ah, av and phases Sh, Sv. The angle u is the angular frequency and t denotes time. For simplicity the phase S is defined as difference between the phases Sh and Sv [122]. Polarization describes the locus of the electric field vector in a plane orthogonal to the propagation direction k in time. For the general case the shape used to describe an electromagnetic wave polarization is an ellipse, which can be determined by two geometrical parameters: Firstly, the ellipticity angle, x, which specifies the shape of the ellipse and the direction of rotation sin (2x) = Q t , a / l g m f l _ s i n (2a e) sin<5 . (D-2) The tangent of the angle ae describes the ratio between horizontal and vertical component of the electric field vector t a n a e = — . (D-3) The tangent of the ellipticity angle determines the ratio between major (a )^ and minor (av) axis of the ellipse and its sign defines the direction of rotation, x > 0° means clockwise rotation (left handedness) whereas x < 0° indicates anti-clockwise rotation (right handedness) t anx = ± — • (D.4) The range of values for x spans [—45°, 45°] where the maximum values indicate circular polarization and x = 0° stands for linear polarization. For the latter case, the ellipse reduces to a line and the polarization is referred to as linear. 150 Secondly, the orientation angle ip, which specifies the orientation of the ellipse's major axis relative to the vertical polarization tan (2^) = a , , a f c cosS = tan (2ae) cos 5 . (D-5) The range of values for ip. spans [—90°, 90°] where 0° indicates vertical polariza-tion and ±90° stands for horizontal polarization. D.2 M a t r i x Representat ion A polarimeter is a radar transmitting in two orthogonal polarizations and receiving all four available channels, two co-polarized and two cross-polarized. The basic data measured by a polarimeter is a complex (amplitude and phase) scattering matrix, S, for each resolution element of the radar. The scattering matrix models the scattering behavior of the target area. It relates the transverse components of the received wave electric field, Er, to those of the incident wave, Ef [36] Er = SE" where S = rs denotes the distance between sensor and target and kw is the wave number of the transmitted signal. Note that throughout this work the Backscatter Alignment convention [122] is used. The elements Srt are known as the complex scattering am-plitudes and the subscript rt refers to receive and transmit polarization respectively. In general the scattering amplitudes can be a function of frequency, illuminating and scattering angles as well as the target aspect angle which is the orientation of the scatterer relative to the coordinate system. A single channel (single-polarization) S A R measures only one of the scattering amplitudes which is generally one of the main diagonal elements (the co-polarized responses Shh or Svv). While the scattering amplitude relates to a single scatterer, in reality the received signal is a sum of backscattering from a random natural surface. Therefore, the observed matrix must be interpreted as a random variable [15]. 151 Shh Shv $vh Svv (D.6) In most cases, the two cross-polarized components Shv a n < f Svh are identical. While there are known targets, for which the reciprocity theorem does not hold, the identity can be assumed for most natural targets resulting in the reduction of the amount of data. This assumption, however, relies strongly on thorough calibration of the data, as channel imbalance, receiver noise, and crosstalk may severely affect the measurement. A vectorized version of the scattering matrix S is the covariance vector kc- For the case of reciprocity, kc can be written as (D.7) The outer product of this vector with itself (i.e., vector multiplication with its conjugate transposed) yields the covariance matrix C , which also fully describes the scatterer C = k c k c = The symbol + denotes conjugate transposed. The covariance matrix C is positive semidefinite Hermitian (i.e., it has real eigenvalues) and consists of nine independent real parameters, three real backscattering coefficients in the main diagonal and three complex correlations. It not only fully describes the scatterer but also immediately shows all measurable properties of the target [80]. C is a power representation, in difference to S, which is a voltage representation. While the absolute phase is lost in the power representation, inter-channel phases are preserved. In addition^ the matrix can be averaged thus allowing filtering of the data to reduce the effect of speckle. Another way of vectorizing the scattering matrix is the coherency vector kx, which is based on. the Pauli-spin elements [23]. For the case of reciprocity kx writes kn -Shh VZShy hh^hv 2 \Shhf V^ShhSj V2ShvSlh 2 \ Shv\ V2SvvShv q q* o v v o h h Shh.S*v VzShvS* (D.8) 152 as Shh + Shh ~ Svv 2Shv (D.9) The corresponding coherency matrix, T, is then defined as | S h h | 2 + 2 f i e ( S h h S ; j + | S „ „ | 2 \shh\2 \Shhf + j 2 / m ( S h h S ; u ) - | S „ „ | 2 | S h h | 2 2ShvS'hh +2Sh„S'vv 1 • j1Ira(ShhS"vv) - | S „ „ | 2 - 2 R e ( S h h S ; v ) + | S „ V | 2 4 | S h J 2 (D.10) Like C, T is positive semidefinite Hermitian. In fact, both matrices share the same eigenvalues since T relates to C by a unitary similarity transformation. The span (i.e., sum of the diagonal elements) of both matrices yields the total power (see below). D.3 Polarimetric Parameters S A R polarimetry provides a vast amount of information. A number of parameters can be measured directly, others can be derived. Relations between parameters also contain useful information. Backscatter Magnitude Target backscatter for distributed targets is often expressed in terms of normalized radar cross section (an) and given in dB. For all natural targets the backscatter magnitudes are decreasing functions of incidence angle. In polarimetry, the total power is also available and frequently used for data analysis TP = (\Shh\2) + (\Shv\2) + (\Svh\2) + (\SVV\2) . ( D . l l ) 153 Channel Ratios Channel ratios also contain information, most important for the analysis of polari-metric imagery is the co-polarization ratio |2\ Rcopol - • (D-12) Complex Channel Correlation The complex correlation of the channels contains twofold information (ShhS*v) . Pcopol = , • (DAS) (\Shh\2) (\SVV\2) Firstly, the correlation magnitude |p| indicating the similarity of the channels. The channel phase ip = Z {ShhSyV) is the relative phase between channels. This mea-sure is preserved during averaging of scattering matrices while the absolute channel phases are lost. The,correlation magnitude between co- and cross-polarized channels is generally low and the phase is random. D . 4 P o l a r i z a t i o n S y n t h e s i s Knowledge of the scattering matrix allows the calculation of the received power for any possible combination of receive and transmit antenna. This process, called polariza-tion synthesis, distinguishes a polarimetric R A D A R greatly from a single-polarization R A D A R . It shows that much more information about the target is available than a mere tripling of the number of channels B = \x%Sxt\ (D.14) 154 where , cos (ip)rt cos ( r ) r j t - i sin (ib)rj sin ( r ) r ) t / - ^ i r x av,t = I | = I | (D.15) sin (VOr.t cos ( r ) r i t - z cos (^) r i t sin ( r ) r > t Polarization synthesis can be applied to power representations (i.e., Stokes, co-variance and coherency matrix) thus providing the advantage that the collective prop-erties of a group of resolution elements can be investigated [126] B = I^Cy*! (D.16) where y ( hthr \ h t V r + V t h r \ VtVr j (D.17) Polarization Response The polarization response plot is the graphical representation of the received power for all co-polarized (or cross-polarized) transmit/receive combinations. The plot is generated as function of ellipticity and orientation and often used to illustrate polari-metric properties of the target. Polarization responses show marked differences for different types of scatterers, they are not unique, though [12]. D.5 Target Decomposition and Its Use for Classi-fication Radar polarimetry also provides the unique possibility for separating scattering terms of different nature which leads to a detailed characterization of the scattering. A num-155 ber of algorithms were developed, they range from analysis of a single look (scattering) matrix to the decomposition of distributed (i.e., multilooked matrices). Coherent Target Decompositions Coherent target decomposition refers to the analysis of single look scattering matri-ces. The various methods available are of particular interest for the detection and identification of small targets (in the order of the S A R resolution). Sea ice is a dis-tributed target, with vast areas to be observed. A detailed analysis of single scattering matrices is therefore of little interest for ice observation. Much more important for sea ice monitoring is the incoherent analysis of scat-tering, which can be applied to power representations like covariance and coherency matrices. Target decomposition can then be used to analyze multilooked imagery. The algorithms show varying degree of complexity. A simple approach for target classification was developed in 1989 [127]. The method is based on differences in the received backscatter depending whether the signal has bounced an odd, even or several number of times (diffuse) before reaching the receiver. A fourth class is assigned to non-classifiable pixels, which do not meet the following two requirements Van Zyl (D.18) (\SVV\)> (\Shv\) . (D.19) 156 If the initial requirements are met, the three remaining categories can then easily be found as Odd: (\Re(S*hhSvv\) > (\Shvf) Even: \(Re (S*hhSvv))\> (\Shv\2) Diffuse: \(Re (S*hhSvv))\ < (\Shv\2) (D.20) (D.21) (D.22) where Re(.) represents the real component of a complex number and * means conju-gate complex. If the result is ambiguous, the pixel is not classified. Even bounce is of-ten characteristic for manmade objects whereas odd bounce often indicates backscat-ter from a rough surface as for example ocean. Diffuse scattering is the third class available and indicates volume scattering. The simplicity of the algorithm makes it a good supplemental tool for target analysis. The method can be applied to multilooked and filtered imagery. One disad-vantages is that trihedral structures also result in odd bounce (i.e., triple) scattering and might be confused with surface scattering. Freeman - Durden While the Van Zyl method is a rather mathematical consideration of the problem, Freeman and Durden proposed a similar concept based on models for three scattering mechanisms volume scattering, even bounce (i.e., double bounce) and odd bounce (i.e., from a rough surface) [44]. The volume scattering model is based on long thin cylindrical scatterers with uniform orientation. The normalized volume scatterer can then be expressed as (D.23) 157 The double bounce scattering model is based on a dihedral corner reflector with different dielectric materials for the reflective surfaces. Rth and Rtv are the reflection coefficients for horizontal and vertical polarization on the vertical surface. Rgh and Rgv are the Fresnel reflection coefficients for the horizontal surface. Incorporation of propagation factors e^2lh'v (the jh,v a r e complex) for horizontal and vertical polariza-tion, respectively makes the model more general. The angle ay is the phase angle between the two co-polarized channels Shh a n d Svv (\Shh\2) = \af\2 (\SVV\2) = 1 (ShhS*v) = af (\shv\2) = 0 (ShhSly) = (ShvSyy) = 0 where (D.24) a / = ej2(lh-lv) f ^ h ^ ) . (D.25) For the surface scatter model a first order Bragg model is used which can be normalized to (\Su,\2) = I/3/I2 <|S™| 2) = 1 {ShhSyy) = • Bf (\shv\2) = o (ShhSly) = (ShVS*yy) = 0 (D.26) with the parameter 8j being real. 158 For all components of the backscatter reciprocity is assumed. Co- and cross-polarized channels are uncorrelated. Further assuming that the three scattering com-ponents are uncorrelated, the model for the total backscatter is then (\Shh\2) = fs\Pf\2 + fd\Vf\2 + fv " ( l ^ l 2 ) = fs + fd + fv (ShhS:v) = fspf + fdaf + ^ (\shv\2) = f (ShhSl) = <S,,.S;*,.) = 0 . (D.27) The parameters fs, fd and fv are the surface, double bounce and volume scatter contributions to the V V cross section. If these three parameters can be estimated, the contribution of the scattering mechanism to H H , H V and V V can be separated. Since neither the surface component nor the double bounce component con-tributes to the H V term, the volume contribution fv can directly be estimated using H V only. H V generally is an order of magnitude smaller than H H or V V , fv and can therefore be neglected in several equations. Freeman - Durden use Van Zyl's decision rule in the process to decide between even and odd bounce to be able to solve three equations in four unknowns (\shh\2) = / ^ / i 2 - . / ; , | n . / | - ' (\SVV\2) = fs + Id (ShhS*vv) ' = fs3s I fdOf . (D.28) If Re(Shh-S*v) is positive (i.e., surface scatter is dominant) a) is set to -1. If Re(Shh-S*v) is positive (i.e., double bounce scatter is dominant) cxf is set to 1. This approach ob-viously works best if either fs or fd are close to zero. A n estimation of the contribution of each scattering mechanism to the total 159 power TP (= span of the coherence or the covariance matrix) is given as TP Ps + Pd + Pl V where Ps Pd fs (1 + \Pf) fd ( i + M 2 ) 8/„ (D.29) V 3 The Freeman-Durden target analysis method provides more flexibility for clas-sification than the Van Zyl method. The relative contributions of surface scattering, volume scattering and double bounce can be used in a variety of ways: • Strongest backscatter contribution • Relative contributions (e.g., Surface scattering / Volume scattering) • Allow mixed classes The second advantage is the immediate availability of class interpretation based on the scattering mechanisms involved. This provides much information for scene inter-pretation. C l o u d e - P o t t i e r E igenvec to r B a s e d D e c o m p o s i t i o n The generally complex coherency matrix T is Hermitian (i.e., T T = T*) for any natural target thus leading to real eigenvalues, A; . Eigenvector based decomposition is therefore applicable without restrictions and the advantage of basis invariance makes this a powerful tool for the interpretation of polarimetric imagery. The resulting orthogonality of the eigenvectors allows the subtraction of any known noise power component [48]. The coherency matrix T can be expressed in diagonal form using its eigenvalues, A i , which can be interpreted as statistical independence of a set of target vectors, a 160 physical interpretation which justifies this approach for general decomposition. Us denotes the 3 x 3 matrix composed of the three orthogonal eigenvectors, Ai 0 0 I T = U3 where U3 = and Ai > A 2 > A 3 . 0 A 2 0 0 0 A 3 1 e i e 2 e 3 U+ (D.30) (D.31) (D.32) For the case of T being diagonal, which is for example the case for azimuthal symmetry, the eigenvalues are equal to the values in the diagonal elements of the matrix and can easily be interpreted as [21] = \HH + VV\2 Ai A 2 A 3 \HH - vvy \HV\ (D.33) (D.34) (D.35) Entropy H : The values of the A^ is somewhat data dependent, their ratios are therefore of much more interest. The dominant scattering mechanism then has a degree of randomness or entropy H (in the von Neumann sense) derived from the normalized eigenvalues A^ as 3 A,; i=l (D.36) H is dimensionless and represents the randomness of the scattering. H = 0 in-dicates a single scattering mechanism (isotropic scattering) while H = 1 indicates a random mixture of scattering mechanisms with equal probability and hence a depo-larizing target. P; can be interpreted as probability describing the partition of the scattering process described by the corresponding eigenvalue. a-Angle: While the eigenvalues are a measure for the partition of the scattering mechanisms, the eigenvectors ej describe the mechanism itself. A n interpretation of 161 the eigenvectors as scattering vectors (vectorized scattering matrices) leads to three orthogonal scattering mechanisms which can be separated even with the presence of noise and depolarization effects [15]. For an unrestricted model a parametrization of the eigenvectors into four angles is suggested [21] where the cti correspond to the variation of the scattering mechanism and ranges from 0° to 90° cos (ai) sin (ai) cos (3i)ei6i (D.37) sin (ctj) sin (/?i)e J 7 i for i = 1,2,3. (D.38) If the scattering medium does not have reflection symmetry there exists a pre-ferred orientation angle (or polarization angle), 9, for the scatterer which represents the orientation of the scatterer about the line of sight. The angle 0 is twice the polar-ization angle. The angles 7 and 8 are target phase angles and can in general have any value. 5 denotes the phase difference between the decomposed Shh + Svv and Shh — Svv whereas 7 represents the phase difference between the decomposed Shh + Svv and Shv terms. The best estimate in terms of maximum likelihood of the average angles can be given using the probabilities based on the eigenvectors a = P i « i + P2a2 + P3«3 where P = — 3 ^ . (D.39) a represents the average dominant scattering mechanism. Given random media problems only the average a can be estimated unambiguously and is therefore of use for an automated classification scheme [28]. The lower limit is given by surface scattering in the Geometrical Optics (GO) limit (a. = 0) followed by surface scattering under Physical Optics (PO) and the Bragg surface model. Dipole scattering (a — 45) is followed by double bounce scattering from metallic surfaces and finally dihedral scatter from metallic surfaces (a = 90°), the upper bound. Anisotropy A: Entropy and a describe a scattering process based on eigenvalues and eigenvectors, however, the entropy is not a unique function of the eigenvalue 162 ratios. One and the same entropy could be the result of different combinations of eigenvalues. A second eigenvector parameter therefore proves useful [85] and can be defined as Anisotropy A A - (D.40) A represents the relative powers of the second and third scattering mechanisms, A = 0 indicates azimuth symmetry while values of A > 0 indicate increasing amounts of anisotropic scattering. When A = 0 the second and third eigenvalues are equal. This is the case for azimuthally symmetric surfaces as well as for a random scatter-ing type where all three eigenvalues are equal. A > 0 generally suggests multiple scattering [20]. For one-dimensional surfaces A becomes one. In this case the H V component of the backscatter is zero. Classification based on eigen-analysis Using H and a for classification pur-poses leads to a transformation of each pixel into the H / a plane, which is given in Figure D . l a . The observable a values given a certain entropy are bounded between curves I and II. These bounds are generated using a diagonal coherency matrix with degenerate minor eigenvalues [21]. The bounds (Curve I and Curve II) as shown in Figure D . l a indicate that high entropy limits the ability to classify different scattering mechanisms. A n initial parti-tion into nine (eight usable) classes as shown in Figure D . l a is suggested [21]. Classes are chosen based on general properties of the scattering mechanism and do not depend on a particular data set. This allows unsupervised classification based on physical properties of the signal. • Class Z 1: Double bounce scattering in a high entropy environment • Class Z 2: High entropy multiple scattering • Class Z 3: High entropy surface scattering (this is N O T a feasible region) • Class Z 4: Medium entropy multiple scattering • Class Z 5: Medium entropy dipole scattering 163 90 -80 • 70 • 60 • B 50 -_ 40 • e < 30 • 20 -10 • 0 -. 0 (a 1 r 0.9 • 0.8 • 0.7 • 0.6 -< a. o 0.5 • L_ O CO 'c 0.4 -< 0.3 • 0.2 • 0.1 • 0 -one-dimensional rough surface smooth azimuthally symmetric surfaces _l I L 0.8 1 Entropy H [1] 0 0.2 (b) H/A-plane 0.4 0.6 0.8 1 Entropy H [1] Figure D . l : (a) H / a plane with bounds and partitioning. A description of the classes . ( Z l - Z9) as well as the observable area is given in the text, (b) H / A plane of surface scattering 164 • Class Z 6: Medium entropy surface scattering • Class Z 7: Low entropy multiple scattering (double or even bounced scattering) • Class Z 8: Low entropy dipole scattering (strongly correlated mechanisms with a large imbalance between H H and V V ) • Class Z 9: Low entropy surface scattering (GO, P O and Bragg surfaces) It is important to note that the boundaries are somewhat arbitrary and do cer-tainly depend on radar calibration, measurement noise floor, and variance of parame-ter estimates. Nevertheless, this classification method is linked to physical scattering properties making it independent of training data sets. The number of classes needed as well as the usability of the method highly depends on the application. The H / A -plane representation for surface scattering is given in Figure D . l (b). The bound can be calculated using a diagonal coherency matrix with small A 2 and A3 varying from 0 to A 2 . Introduction of the anisotropy leads to a third dimension for the classification. The H / A / a classification space, given in Figure D.2, now provides further capabilities to distinguish between different scattering mechanisms. For example, high entropy and low anisotropy correspond to random scattering whereas high entropy and high anisotropy indicate the existence of two scattering mechanisms with equal probability. One approach is to simply divide the space.into two H / a planes shown in green in Figure D.2, one associated to A < 0.5 the other one associated to A > 0.5 thus introducing 16 classes if the H / a planes are divided according to Figure D . l a . The three parameters H , A , and a are based on eigenvectors and normalized eigenvalues of a local estimate for the 3 x 3 Hermitian coherency matrix. The basis invariance of the target decomposition makes these three parameters roll invariant, i.e., the parameters can be computed independent of the polarization basis. Estima-tion of these three parameters allows a classification of the pixel according to the type of scattering process within the pixel (H, A) and the corresponding physical scatter-ing mechanism (a). Classification is therefore performed on a pixel level, however, 165 Figure D.2: H / A / a classification space. a high number of looks is required to estimate the parameters accurately [68]. The classification method requires the imagery to be multi looked or filtered [64]. D.6 Classification Based on the Wishart Statistics M a x i m u m L i k e l i h o o d Class i f i ca t ion B a s e d on the W i s h a r t S ta t i s t ics of the Cova r i ance M a t r i x Supe rv i s ed Class i f i ca t ion : Image classification in the maximum likelihood sense is not based on the physical event responsible for image generation (i.e., scatter-ing) but on data statistics. Probability considerations become important in pattern 166 recognition because of the randomness under which the pattern classes normally are generated. Knowledge of the data statistics (i.e., the theoretical statistical distribu-tion) allows the derivation of a classification approach that is optimal in the sense that, on average, its use yields the lowest probability for misclassification [45]. Initial classes are derived using a training data set in known terrain. The image is then classified according to the distance to the class means. Each pixel is assigned to the class to which it has the minimum distance. The distance itself is calculated to provide that the Bayes maximum likelihood classification condition is satisfied [61]. Bayes classification for polarimetric S A R imagery was first used in 1988 [58]. The authors showed that the use of the polarimetric imagery gives optimum classifi-cation results. The algorithm was only developed for single-look polarimetric imagery, though. For most applications in radar remote sensing, multi-looking is applied to the imagery to reduce the effects of speckle noise. The number of looks is therefore an important parameter for the development of a probabilistic model. The full polarimetric information content is given in the scattering matrix S, the covariance matrix C as well as the coherency matrix T. It has been shown that T and C both have a complex Wishart distribution. A l l equations below are applicable to both C and T, which is why the symbol Z is used instead. The probability density function (pdf) for a given number of looks, n, is given as n < ? n | ^ | « - 9 e-nTrace(y 1 <Z)) "Z « Z » = — K(n,q)\V\ p ( g ( ( » = " ^ , T „ w (D.41) where V = E{(Z)} . (D.42) E{ } is the expected value, q represents the dimensionality (3 for reciprocal case, else 4) and Trace is the sum of the elements in the main diagonal of a matrix. V is the expected value for the covariance or coherency matrix, in other words its mean. <> represents an ensemble average. K(n, q) is a normalization factor given as K{n,q) = TT^)q{-q-l)T{n)...T(n-q + l) . (D.43) 167 The class mean Vm is defined as Vm = E{{Z)\{Z) eum} (D.44) where iom is the set of pixels belonging to class m. The actual value for the class mean can then be estimated as (D.45) According to Bayes maximum likelihood classification a distance measure, d, can be derived as [61] d((Z),Vm) = n (In\Vm\ + Trace ( V ^ Z ) ) ) -In (P (u,m))-(n-q) In\(Z)\+In (K(n,q)) . (D.46) The last two terms are not related to the class mean, they can be eliminated without changing the classification result. A n increasing number of looks, n, decreases the contribution of the a priori probability P(u>m). Also, the a priori probability can be assumed to be equal for all classes since no information on the classes is available. The elimination of this term is therefore justified [61]. A n appropriate distance measure can then be written as [60] d((Z),Vm) = In | V m | + Trace {Vm-\Z)) (D.47) thus leading to a look independent minimum distance classification d({Z),Vm) < d((Z),Vj). for all uj ^ um . (D.48) A pixel is assigned to a certain class if the distance between pixel and class mean is minimum. The look independence of this scheme allows its application to multilooked as well as speckle filtered imagery. This classification scheme can also be generalized for multi frequency polarimetric imagery provided that the frequencies are sufficiently separated to ensure statistical 168 independence between frequency bands [61]. Also, the imagery from different bands need to be properly co-registered in this case. The new distance measure, d c L is then simply calculated as the sum of the distance measures (dc, £?L) of the various single-frequency imagery [37] dCL = dc + dL. (D.49) The classification depends on a training set and must therefore be applied under supervision. Also, it is not based on the physics of the scattering mechanism, which might well be considered a disadvantage of the scheme. It does however utilize the polarimetric information and allows a look independent image classification. Unsupervised Classification: The Wishart distribution based M L classifi-cation relies on a training set or initial clustering of the imagery and is therefore supervised. A combination with an unsupervised method is suggested in [60]. While any unsupervised method can be used, scattering based classifiers were recommended more recently. The unsupervised method is then used to set up the initial classes for the Wishart distribution based classifier. The basic concept is shown in Figure D.3. To further improve the result, a K-means approach is suggested, where for each iteration, the classifier is initialized with the result from the previous run. Possible options for the use of the K-means clustering algorithm which uses a Wishart classifier are shown, in Figure D.4, including the fully supervised direct classification based on operator selected training areas. For operational use a more automated approach is desirable. Class interpretation (i.e., assignment of classes to ice types) is required for all results with one exception. Figure D.4 is intended to illustrate the versatility of the Classifier. Some of the approaches shown were developed as part of this work and are described in more detail in Chapter 3. Using the H / A / a space to initialize the K-means clustering algorithm which uses a Wishart classifier provides improved results over the H / A / a classification. For the H / A / a classification, the thresholds are somewhat arbitrary and not the entire 169 Filtered data Polarimetric Data Polarimetric Filter Seed classification Unsupervised Classifier Classification update (n iterations) Wishart Classifier Result Interpretation Classified Image Figure D.3: Wishart-based classification concept. Result interpretation is a required manual task. 170 Fully supervised approach (One iteration only) Operator interaction required before application No operator interaction before classification, class interpretation required. No operator interaction before classification; main scattering mechanism known C l a s s n u m b e r d e t e r m i n a t i o n Manual through test areas Manual (visual interpretation of the scene) Fixed number of classes (need not be all populated) Automatic ' V - - „ " - «. In i t i a l i za t ion Manual through test areas Modified H/a or H/A/a (selective use of thresholds) Image with "n" classes randomly distributed H/a or H/A/a or Hc/H L/H P Separation of scattering mechanisms C l a s s n u m b e r r e d u c t i o n Not required Not required Manual through class interpretation Automatic 1 C l a s s i n t e r p r e t a t i o n Not required Interpretation of classes is generally required Main scattering mechanism known Figure D.4: Possible options for the Wishart based method with degree of automation indicated. 171 polarimetric information can be used due to the inability to determine all four angles that parameterize the eigenvalues. Due to the iterations, however, the class means wil l not remain constant and the final classes therefore deserve further discussion. In particular, the initial clustering is made in the H / A / a domain which is normalized and therefore backscatter strength invariant, the minimum distance classification is performed using the coherency matrix and wil l be sensitive to backscatter strength variations. After the K-means classification, the clusters may therefore well overlap in the H / A / a plane. They have to be assigned to target types manually. The number of classes may be fixed [21] or changed based on the application [109]. The highest degree of automation is achieved by incorporating the interpretation of scattering mechanisms in the classification procedure [63]. The Freeman-Durden method itself is very much related to backscatter strength, there is a closer relationship between initial classes and subsequent iterative classification result. The creation of three super classes based on dominant scattering mechanisms is suggested. The iterative classifier is only allowed to shift classes within one super class (i.e., the interpretation of the super class remains valid throughout the process). The number of classes is a required input, though. This approach has shown useful in urban and rural areas. Manual class interpretation is required for a more detailed analysis [108]. For all approaches, if some knowledge of the scene is available, the number of classes can be restricted and the class interpretation effort can be reduced. Restricting the number of classes involves user intervention and reduces the degree of which the classifier is automated [108]. M A P Classifier A maximum a posteriori ( M A P ) classification method is suggested in [90]. Each sample point is characterized by two attributes: • A polarimetric measurement vector • A region label which identifies the region to which the sample belongs 172 Given the polarimetric measurement vectors, a model for the posterior distribu-tion labels is proposed. Optimal region labelling is defined as maximizing the posterior distribution of the region labels. The entire polarimetric information is contained in the covariance/coherency matrix, each class is represented by a single matrix V/, the ensemble average. The conditional distribution of the Covariance/coherency matrix given the entire region label array, L is p(Z/L) = l[p(Zs/Ls) . (D.50) s The conditional distribution of Zs in each region can be written in a Gibbs represen-tation e-NUf(Zs/La=l) p(Zs/.Ls = l) = 3 N . (D.51) TV U( is the energy function built from the elements of the Covariance matrix. N is the number of looks, s denotes the pixel location and 1 denotes the class. Region labelling is achieved by using Markov random fields ( M R F ) . M R F are a computationally efficient way of representing local interactions between neighboring pixel attributes. They can also be described in terms of Gibbs energy functions. The conditional distribution of the region label array Ls, given region labels elsewhere, only depends on the immediate neighborhood C/| is the energy function built from the immediate neighborhood. Using Bayes' theorem, the posterior distribution of the region label Ls given the polarimetric imagery and the region labels Lr of a neighborhood N® can be written as . p(Ls = l/Lr,reN°,Z3). = e [ - w f ( ^ . = 0 - J v t / | ( ^ = ^ ; r g w ? ) ] _ : .. ( D . 5 3 ) The posterior distribution of the entire region label array L given the polarimetric information Zs p(L/Z) = e-»Ze[u^z3=i/L3)+uz(La=i/u, re jy? )} . ( D 5 4 ) 173 The estimate of L that minimizes Equation D.54 is the Bayes' most likely estimate ( M A P estimate) of the region labels given the polarimetric estimate £ M A P = iUi (Z* = l/L^} - N U 2 (L° = l/Lr, r E 7V°) . (D.55) s The energy functions U\ and XJi depend on local polarimetric information and region labels. The concept can be extended to multi-frequency imagery as well as to single-frequency multi-polarization imagery. The classifier was found to be sensitive to inci-dence angle dependent variations of the backscatter, a split of the region in different incidence angle ranges is therefore recommended. The M A P classifier is a supervised method as it requires initial cluster centers to classify a scene. Similar to the Wishart classification, an unsupervised approach is possible if class means can somehow be determined automatically. The use of an I S O D A T A algorithm is suggested which involves a small random subset of the scene. Due to significant intensity differences in co- and cross-pol channels in the linear domain, the use of the log domain is suggested. Information on the number of classes is required input, this information is application specific [91]. The M A P classifier is closely related to the Wishart classifier, both operate on a Gaussian data distribution channel. The M A P classifier takes neighborhood into consideration during classification, whereas the use of an advanced directional filter is suggested for the Wishart method. A n example result is shown in Chapter 3. Classification Based on Expectation Methods A n approach involving the estimation of class mean with expectation maximization is presented in [25]. The authors argue that the estimate of the class mean is skewed for poorly separated classes. In an iterative approach, some a priori knowledge on the number of classes as well as class statistics is available. Assumed is the following 174 probabilistic model M p(x\&) - ^UiPiixfii) (D.56) i=l where x is (multi-dimensional data from M component densities of parameter Oi present in spatial variations uJi. For the polarimetric case, the class weights, u>m, and class means, Vm, can then iteratively be estimated using the previous parameter guess over N segments based on expectation maximization N N Urn = 1 E "rssPz ( m | Z 4 , V^">) (D.57) i = l Vm = ^ ^ " V m v • (D-58) The method assumes that the number of classes, M, is known and that each class is completely defined by a fixed covariance matrix. M is application specific and cannot easily be determined. The final implementation involves the use of a K-means iteration for the parameters to converge to 0.01%. While the method is fast for smaller test scenes, the approach is memory intensive and it is unclear how it would work on large scenes used in an operational environment. The concept is closely related to other classification approaches such as "fuzzy classification" [31]. The method can be modified to work on reduced versions of the covariance (or coherency) matrix, i.e., intensity only. While not documented, an application to multi-frequency imagery seems possible, also. The method is also related to the Wishart classifier as it is based on the data distribution. The main difference is the introduction of weights in an attempt to better estimate the class means. A n example result for polarimetric imagery is shown in Chapter 3. 175 D.7 Multi-Polarization SAR Imagery The second generation of spaceborne SARs provides a number of modes that con-tain only part of the scattering matrix. R A D A R S A T - 2 and A L O S - P A L S A R provide dual-polarization modes, whereas E N V I S A T - A S A R provides a so-called alternating-polarization mode which differs slightly in the way the imagery is acquired [73], [34]. The three methods described above can be modified to work with a partial co-variance matrix. This is not the case for classification algorithms based on target decompositions, as vital information for the decomposition is missing. In this section the modification of the Wishart method is described in more detail. The main impact of this is the unavailability of decomposition methods to initialize the classifier. The classification methods for E N V I S A T A S A R imagery in its various formats are summarized in [35]. This summary is based on the Wishart classifier. The authors of [30] suggest the use the use of amplitude and texture characteristics, their method is a variation of the M A P classifier. The authors of [62] quantitatively compare the classification capability of polari-metric vs. dual-polarization imagery based on the same approach. They conclude that the importance of having polarimetric information available over dual-polarization im-agery varies with the application. The radar frequency plays a significant role based on the application. In the dual-polarization case, the scattering matrix only has two available el-ements, the resulting covariance matrix is also reduced to a 2 x 2 matrix. Possible options for this matrix are '2hhhv — \Shh\ ShhS, 2 (D.59) '2vvvh — SvhS*. Svh^*v \SVV\ vv 2 (D.60) 176 '2hhvv \Shh\2 ShhS*v (D.61) <? <?* I <> I The distance measure is then derived similar to the polarimetric case and can be written as d2((Z2),V2rn) = In | F 2 m | +Trace ( F 2 m - 1 ( Z 2 ) ) (D.62) where Z 2 is the reduced sample covariance matrix and V 2 m represents the class mean. The classification concept holds for single intensity S A R imagery as well with d i ( ( ^ i ) , ^ i m ) = In | V l T O | + T r a c e (Vx^-^Zx)) (D.63) where Z\ is the sample channel intensity (e.g., ISTI / J 2 ) and V i m represents the class mean. The multi-polarization modes of spaceborne SARs are of particular significance as they are available with wider antenna swaths, thus providing more coverage. In particular, R A D A R S A T - 2 wil l provide dual-polarization ScanSAR imagery with 500 km swath width. Currently, single-polarization ScanSAR imagery from R A D A R S A T -1 are the most commonly used imagery at the CIS due to their coverage and revisit requirements. 177 Appendix E Confusion Matrices This Appendix contains the confusion matrices discussed in Chapter 3. Table E . l summarizes the references used. A l l methods for generating reference data are de-scribed in Chapter 3. Table E . l : Validation data used in Chapter 3 to calculate the confusion matrices (References). Reference Description Reference A ful l dual-frequency result Reference B top 50%, based on conf and a requirement that the C-band only classification in the last iteration results in the same class as the dual-frequency classification Reference C top 10%, based on conf. Reference D Use of user selected validation data 178 E . l Dual-Frequency Reference - Reference A Table E.2: Confusion Matr ix of C-band T P initialization compared with dual fre-quency reference given in % of the reference class (Reference A) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 60 33 0 0 0 0 M Y I 38 66 11 0 0 0 F Y I R R 2 1 75 23 1 0 R F Y I 0 0 14 68 43 0 S F Y I 0 0 0 9 55 33 T h i 0 0 0 0 1 67 Overall classification accuracy: 65.9% Table E.3: Confusion Matr ix of C-band Freeman-Durden initialization compared with dual frequency reference given in % of the reference class (Reference A) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 25 6 0 0 0 0 M Y I 1 69 90 . 9 . 0 0 0 F Y I R R 2 1 67 !5 1 0 R F Y I 0 0 22 80 58 0 S F Y I 0 0 0 4.-'. 40 35 T h i 1 4 3 '2 1 1 65 Overall classification accuracy: 66.1% (1): formed from two different classes 179 Table E.4: Confusion Matr ix of L-band T P initialization compared with dual fre-quency reference given in % of the reference class (Reference A) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 83 0 23 0 0 0 M Y I 1 17 54 72 43 0 0 F Y I R R 0 44 5 41 0 0 R F Y I 2 — — — — — — S F Y I 0 2 0 16 59 0 T h i 0 0 0 0 41 100 Overall classification accuracy: 42.1% (1): formed from two different classes (2): not classified Table E.5: Confusion Matr ix of L-band Freeman-Durden initialization compared with dual frequency reference given in % of the reference class (Reference A) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 76 0 31 0 0 0 M Y I 1 11 77 6 90 36 0 F Y I R R 12 23 61 4 0 0 R F Y I 2 — — —. — — — S F Y I 0 0 0 0 60 38 T h i 3 1 0 2 ••2 3 62 Overall classification accuracy: 58.4% (1): formed from three different classes (2): not classified (3): formed from two different classes 180 Table E.6: Confusion Matr ix of dual channel (HH+HV) classification compared with dual frequency reference given in % of the reference class (Reference A) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 59 32 0 0 0 0 M Y I 39 67 10 0 0 0 F Y I R R 2 1 . 72 20 1 0 R F Y I 0 0 17 66 35 0 S F Y I 0 0 1 14 63 27 T h i 0 0 0 0 1 73 Overall classification accuracy: 66.3% Table E.7: Confusion Matr ix of dual channel ( V V + V H ) classification compared with dual frequency reference given in % of the reference class (Reference A) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 58 32 0 0 0 0 M Y I 39 67 9 0 0 0 F Y I R R 3 1 71 17 0 0 R F Y I 0 0 19 68 31 0 S F Y I 0 0 1 14 63 23 T h i 0 0 0 0 6 77 Overall classification accuracy: 66.2% 181 Table E.8: Confusion Matr ix of single channel (HH) compared with dual frequency-reference given in % of the reference class (Reference A) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 59 43, 0 0 0 0 M Y I 37 54 19 2 0 0 F Y I R R 4 3 63 30 6 0 R F Y I 0 0 17 57 49 0 S F Y I 0 0 1 11 44 29 T h i 0 0 0 0 1 71 Overall classification accuracy: 56.3% Table E.9: Confusion Matr ix of single channel (HV) compared with dual frequency reference given in % of the reference class (Reference A) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 56 25 o 0 0 0 M Y I 41 74 12 0 0 0 F Y I R R 3 1 73 25 2 0 R F Y I o 0 15 66 49 17 S F Y I 0 0 0 9 48 73 T h i 0 0 0 0 1 10 Overall classification accuracy: 64.4% 182 Table E.10: Confusion Matr ix of single channel (VV) compared with dual frequency reference given in % of the reference class (Reference A) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 57 45 0 0 0 0 M Y I 39 54 16 1 0 0 F Y I R R 4 1 63 26 3 0 R F Y I 0 0 20 59 37 0 S F Y I 0 0 1 14 55 25 T h i 0 0 0 . o 5 75 Overall classification accuracy: 57.7% Table E . 11: Confusion Matr ix of expectation based classification compared with dual frequency reference given in % of the reference class (Reference A) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 14 2 0 0 0 0 M Y I 1 70 88 8 1 0 0 F Y I R R 3 1 43 37 10 1 R F Y I 13 9 33 7 1 0 S F Y I 0 0 15 49 67 9 T h i 0 0 1 6 . 22 90 Overall classification accuracy: 50.8% (1): formed from two different classes 183 E.2 Classification Confidence and Band Restricted Dual-Frequency Reference - Reference B Table E.12: Confusion Matr ix of C-band T P initialization compared with dual fre-quency reference given in % of the reference class (Reference B) C F Y I M Y I F Y I R R R F Y I S F Y I . T h i CFYI 100 14 0 0 0 0 MYI 0 86 1 0 0 0 FYIRR 0 o 99 10 0 0 RFYI 0 0 0 90 33 0 SFYI 0 o 0 0 67 2 T h i 0 0 0 0 0 98 Overall classification accuracy: 89.2% Table E.13: Confusion Matr ix of C-band Freeman-Durden initialization compared with dual frequency reference given in % of the reference class (Reference B) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 47 4 0 0. 0 0 M Y I 1 53 93 0 0 0 0 F Y I R R 0 0 97 0 0 0 R F Y I 0 0 1 99 56 0 S F Y I ' 0 o •• 0 0 44 6 T h i 1 0 4 2 1 0 94 Overall classification accuracy: 81.0% (1): formed from two different classes 184 Table E.14: Confusion Matr ix of dual channel (HH+HV) classification compared with dual frequency reference given in % of the reference class (Reference B) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 98 15 0 0 0 0 M Y I 2 85 1 0 0 0 F Y I R R 0 0 97 7 0 0 R F Y I 0 0 2 92 25 0 S F Y I 0 0 0 1 75 1 T h i 0 0 0 0 0 99 Overall classification accuracy: 89.6% Table E.15: Confusion Matr ix of dual channel ( V V + V H ) classification compared with dual frequency reference given in % of the reference class (Reference B) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 98 16 0 0 0 0 M Y I 2 84 1 0 0 0 F Y I R R 0 0 96 5 0 0 R F Y I 0 0 3 94 17 0 S F Y I 0 0 0 1 80 5 T h i : 0 0 0 0 3 95 Overall classification accuracy: 89.4% 185 Table E.16: Confusion Matr ix of single channel (HH) compared with dual frequency reference given in % of the reference class (Reference B) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 89 31 0 0 0 0 M Y I 11 68 14 0 0 0 F Y I R R 0 1 79 28 1 0 R F Y I 0 0 7 68 56 0 S F Y I 0 0 0 4 43 1 T h i 0 0 0 0 0 99 Overall classification accuracy: 71.1% Table E . 17: Confusion Matr ix of single channel (HV) compared with dual frequency reference given in % of the reference class (Reference B) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 96 12 0 0 0 0 M Y I 4 88 5 0 0 0 F Y I R R 0 0 92 20 0 0 R F Y I . 0 0 3 77 38 10 S F Y I 0 0 0 3 60 76 T h i 0 0 0 0 2 14 Overall classification accuracy: 82.4% 186 Table E.18: Confusion Matr ix of single channel (VV) compared with dual frequency reference given in % of the reference class (Reference B) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 88 36 0 0 0 0 M Y I 12 64 10 0 0 0 F Y I R R 0 0 82 22 0 0 R F Y I 0 0 8 75 32 0 S F Y I 0 0 0 3 66 7 T h i 0 0 0 0 2 93 Overall classification accuracy: 73.8% 187 E.3 Classification Confidence Restricted Dual-Frequency Reference - Reference C Table E.19: Confusion Matr ix of C-band T P initialization compared with dual fre-quency reference given in % of the reference class (Reference C) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 99 14 0 0 0 0 M Y I 1 86 0 0 0 0 F Y I R R 0 0 74 . 0 0 0 R F Y I 0 0 25 79 44 0 S F Y I 0 0 1 21 56 0 T h i 0 0 0 0 0 100 Overall classification accuracy: 83.9% Table E.20: Confusion Matr ix of C-band Freeman-Durden initialization compared with dual frequency reference given in % of the reference class (Reference C) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 55 8 0 0 0 0 M Y I 1 45 88 0 o 0 0 F Y I R R 2 0 58 0 0 0 R F Y I 0 0 40 88 74 0 S F Y I 0 0 0 11 26 0 T h i 1 0 4 2 ' 1 0 100 Overall classification accuracy: 67.3% (1): formed from two different classes 188 Table E.21: Confusion Matr ix of dual channel (HH+HV) classification compared with dual frequency reference given in % of the reference class (Reference C) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 99 12 0 0 0 0 M Y I 1 88 0 0 0 0 F Y I R R 0 0 68 0 0 0 R F Y I 0 0 31 67 47 0 S F Y I 0 0 1 33 53 0 T h i 0 0 0 0 0 100 Overall classification accuracy: 81.2% Table E.22: Confusion Matr ix of dual channel ( V V + V H ) classification compared with dual frequency reference given in % of the reference class (Reference C) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 98 16 0 0 0 0 M Y I 2 84 0 0 0 0 F Y I R R 0 0 64 0 0 0 R F Y I 0 0 35 65 21 0 S F Y I 0 0 1 34 78 0 T h i 0 0 0 1 1 100 Overall classification accuracy: 81.8% 189 Table E.23: Confusion Matr ix of single channel (HH) compared with dual frequency reference given in % of the reference class (Reference C) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 93 27 0 0 0 0 M Y I 7 73 0 0 0 0 F Y I R R 0 0 65 4 6 0 R F Y I 0 0 33 70 82 0 S F Y I 0 0 2 26 12 0 T h i 0 0 0 0 0 100 Overall classification accuracy: 71.4% Table E.24: Confusion Matr ix of single channel (HV) compared with dual frequency reference given in % of the reference class (Reference C) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 98 16 0 0 0 0 M Y I 2 83 0 0 0 0 F Y I R R 0 0 76 2 0 0 R F Y I 0 0 23 80 23 6 S F Y I 0 0 1 18 74 77 T h i 0 1 0 0 3 17 Overall classification accuracy: 83.0% 190 Table E.25: Confusion Matr ix of single channel (VV) compared with dual frequency reference given in % of the reference class (Reference C) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 93 39 0 0 0 0 M Y I 7 61 0 0 0 0 F Y I R R 0 0 60 2 1 0 R F Y I 0 0 38 67 42 0 S F Y I 0 0 2 31 56 0 T h i 0 0 0 0 .1 100 Overall classification accuracy: 71.2% 191 E.4 Analyst Selected Validation Data -Reference D Table E.26: Confusion Matr ix of C-band T P initialization compared with dual fre-quency reference given in % of the reference class (Reference D) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 59 18 0 0 0 0 M Y I 39 82 6 0 0 0 F Y I R R 2 0 77 10 0 0 R F Y I 0 0 15 85 3 0 S F Y I 0 0 2 5 96 13 T h i 0 0 0 0 1 87 Overall classification accuracy: 82.0% Table E.27: Confusion Matr ix of C-band Freeman-Durden initialization compared with dual frequency reference given in % of the reference class (Reference D) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 18 2 0 0 0 0 M Y I 1 75 95 5 0 0 0 F Y I R R 2 0 69 1 0 0 R F Y I 0 0 23 97 12 0 S F Y I 0 0 2 2 84 21 T h i 1 5 3 1 . 0 . 4 79 Overall classification accuracy: 84.6% (1): formed from two different classes 192 Table E.28: Confusion Matr ix of dual channel (HH+HV) classification compared with dual frequency reference given in % of the reference class (Reference D) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I ' 58 19 0 0 0 0 M Y I 40 80 6 0 0 0 F Y I R R 2 1 74 5 0 0 R F Y I 0 0 17 77 2 0 S F Y I 0 0 3 18 96 5 T h i 0 0 0 0 2 95 Overall classification accuracy: 80.4% Table E.29: Confusion Matr ix of dual channel ( V V + V H ) classification compared with dual frequency reference given in % of the reference class (Reference D) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 53 23 0 0 0 0 M Y I 44 77 6 0 0 0 F Y I R R 3 0 73 4 0 0 R F Y I 0 0 17 74 1 0 S F Y I . 0 0 4 22 73 15 T h i 0 0 0 0 26 85 Overall classification accuracy: 74.6% 193 Table E.30: Confusion Matr ix of single channel (HH) compared with dual frequency reference given in % of the reference class (Reference D) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 65 33 0 0 0 0 M Y I 27 63 15 0 0 0 F Y I R R 8 4 64 26 0 0 R F Y I 0 0 18 65 18 0 S F Y I 0 0 3 9 81 8 T h i 0 0 0 0 1 92 Overall classification accuracy: 67.1% Table E.31: Confusion Matr ix of single channel (HV) compared with dual frequency reference given in % of the reference class (Reference D) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 53 14 0 0 0 0 M Y I 45 : 86 8 0 0 0 F Y I R R 2 . 0 77 . 6 0 0 R F Y I 0 0 14 88 15 15 S F Y I 0 0 1 6 82 66 T h i 0 0 0 0 3 19 Overall classification accuracy: 78.4% 194 Table E.32: Confusion Matr ix of single channel (VV) compared with dual frequency reference given in % of the reference class (Reference D) C F Y I M Y I F Y I R R R F Y I S F Y I T h i C F Y I 55 43 0 0 0 0 M Y I 37 56 14 0 0 0 F Y I R R 8 1 64 14 0 0 R F Y I 0 0 19 70 4 0 S F Y I 0 0 3 .16 74 19 T h i 0 0 0 0 22 81 Overall classification accuracy: 62.0% 195 Bibliography [1] R A D A R S A T - 2 Special Issue, June 2004. Canadian Journal of Remote Sensing (Vol.30, No.3). [2] R. De Abreu, 2001. Personal Communication. [3] R. De Abreu, 2005. Personal Communication. [4] R. De Abreu, D. Flett, B . Scheuchl, and B . Ramsay. Operational Sea Ice Mon-itoring with R A D A R S A T - 2 - A Glimpse Into the Future. In Proceedings of the IEEE International Geoscience and Remote Sensing Symposium, IGARSS'03. I E E E , 2003. [5] European Space Agency. ENVISAT ASAR Product Handbook, 2002. [6] D . G . Barber, 2002. Personal Communication (with D. Flett). [7] D. G . Barber, D. Flett, R. De Abreu, and E . F . LeDrew. Spatial and Temporal Variations in Sea Ice Geophysical Properties and Microwave Remote Sensing Observations: The SIMS'90 Experiment. Arctic, 45(3):233-251, 1992. [8] D. G . Barber and L . - E . LeDrew. S A R Sea Ice Discrimination Using Texture Statistics: A Multivariate Approach. Photogrammetric Engineering and Remote Sammy, 57(4):385-395 ; 1991. [9] D. G . Barber, M . J . Manore, T. A . Agnew, H . Welch, E . D. Soulis, and E . F . LeDrew. Science Issues Relating To Marine Aspects of the Cryosphere: Impli-196 cations for Remote Sensing. Canadian Journal of Remote Sensing, 18(l):46-55, 1992. [10] D . G . Barber and S. V . Nghiem. The Role of Snow on the Thermal Depen-dence of Microwave Backscatter over Sea Ice. Journal of Geophysical Research, 104(Cll):25789-25803, November 1999. [11] C. Bertoia, J . Falklngham, and F . Fetterer. Analysis of SAR Data of the Polar Oceans, chapter 10, Polar S A R Data for Operational Sea Ice Mapping, pages 201-234. Springer, 1998. [12] W . - M . Boerner, H . Mott , E . Lunenburg, C. Livingstone, B . Brisco, R. J . Brown, and J . S. Patterson. Polarimetry in Remote Sensing: Basic and Applied Con-cepts, chapter 5: in Manual of Remote Sensing, Volume 2: Principles and Applications of Imaging Radar. Wiley, 3rd edition, 1998. [13] C - C O R E . Validation of Iceberg Detection Capabilities of R A D A R S A T Syn-thetic Aperture Radar - Final Report. Technical Report Contract Report 00-C5 prepared for CIS, Centre for Cold Oceans Resources Engineering ( C - C O R E ) , St. John's, Nfld., 2000. [14] C - C O R E . The Capabilities for Iceberg Detection Using R A D A R S A T - 1 . Tech-nical Report Contract Report R 03 078 263 prepared for CIS, Centre for Cold Oceans Resources Engineering ( C - C O R E ) , St. John's Nfld., 2004. [15] R , Caves. R A D A R S A T - 2 Polarimetric Applications. Technical Report R X - R P -51-1572, M D A , 2000. [16] International Development Research Centre. Sharing Canada's Eye in the Sky: R A D A R S A T - Online Article, 2005. http://reseau.crdi.ca/en/ev-2684-201-l-DO-TOPIC.h tml . •17] M . S. Chowdhury and D. A . Clausi. Sea Ice Boundary Detection in S A R Satel-lite Images Using Conflicting Strength. In Proceedings of the First Canadian Conference on Computer and Robot Vision, pages 440-448, 2004. 197 [18] D. A . Clausi and H . Deng. Operational Segmentation and Classification of S A R Sea Ice Imagery. In Proceedings of the IEEE Workshop on Advances in Techniques for Analysis of Remotly Sensed Data, pages 268-275, 2003. [19] D. A . Clausi and B . Yue. Comparing Cooccurrence Probabilities and Markov Random Fields for Texture Analysis of S A R Sea Ice Imagery. IEEE Transac-tions on Geoscience and Remote Sensing, 42(l):215-228, January 2004. [20] S. R. Cloude, I. Hajnsek, and K . Papathanassiou. A n Eigenvector Method for the Extraction of Surface Parameters in Polarimetric S A R . In Proceedings CEOS'99, Toulouse, France, 1999. [21] S. R. Cloude and E . Pottier. A n Entropy Based Classification Scheme for Land Applications of Polarimetric S A R . IEEE Transactions on Geoscience and Remote Sensing, 35(1), January 1997. [22] S. R. Cloude, E . Pottier, and W . M . Boerner. Unsupervised Image Classifi-cation using the Entropy/Alpha/Anisotropy Method in Radar Polarimetry. In Proceedings of the AIRSAR Earth Science and Application Workshop, 2002. [23] I. G . Cumming and Noetix Research. C C R S Radar Polarimetry Tutorial (Online Version), 2004. http://www.ccrs.nrcan.gc.ca/ccrs/. [24] I. G . Cumming and F. H . Wong. Digital Processing of Synthetic Aperture Radar Data - Algorithms and Implementation. Artech House, Norwood, M A , 2005. [25] G . Davidson, K . Ouchi, N . Ishitsuka, G . Saito, K . Mohri , and S. Uratsuka. Polarimetric Classification Using Expectation Methods. In Proceedings of the Polarimetric and Interferometric SAR Workshop PISAR02, Tokyo, 2002. [26] W . Dierking, H . Skriver, and P. Gudmandsen. S A R Polarimetry for Sea Ice Classification. In Workshop on Applications of SAR Polarimetry and Polari-metric Interferometry, E S A / E S R I N , Frascati, Italy, January 14-16 2003. 198 [27] W . Dierking, H . Skriver, and P. Gudmandsen. What is Gained by Using an L-Band S A R for Sea Ice Monitoring. In Proc. ENVISAT/ERS Symposium, September 6-10 2004. [28] M . R. Drinkwater, R. Kwok, E . Rignot, H . Israelsson, R. G . Onstott, and D. P. Winebrenner. Microwave Remote Sensing of Sea Ice, chapter 24 - Potential Applications of Polarimetry to the Classification of Sea Ice, pages 419-430. A G U Geophysical Monograph 68, 1992. [29] M . R. Drinkwater, R. Kwok, D. P. Winebrenner, and E . Rignot. Multifrequency Polarimetric Synthetic Aperture Radar Observations of Sea Ice. Journal of Geophysical Research, 96(Cll):20679-20698, 1991. [30] L . Du, M . R. Grunes, and J.-S. Lee. Unsupervised Segmentation of Dual-Polarization S A R Images on Amplitude and Texture Characteristics. Interna-tional Journal of Remote Sensing, 23(20):4383-4402, October 2002. [31] L . D u and J.-S. Lee. Fuzzy Classification of Earth Terrain Covers Using Complex Polarimetric S A R Data. International Journal of Remote Sensing, 17(4):809-826, 1996. [32] C. Eckes and B . Fritzke. Classification of Sea-Ice with Neural Networks - Results of the E U Research Project I C E R O U T E S . Internal Report 2001-02, Ruhr-Universitaet Bochum, Institut fuer Neuroinformatik, March 2001. [33] L . Eriksson, M . R. Drinkwater, B . Holt, E . Valjavek, and O. Nortier. SIR-C Po-larimetric Radar Results from the Weddell Sea, Antarctica. In Proceedings of the IEEE International Geoscience and Remote Sensing Symposium, IGARSS'98. I E E E , 1998. [34] E S A . E N V I S A T web site, 2005. http://envisat.esa.int/. [35] E S A . Polsarpro - Polarimetry Tutorial. E S A , 2005. http://earth.esa.int / polsarpro / tutorial.html. 199 [36] D. L . Evans, T . G . Farr, J . J . Van Zyl , and K\ A . Zebker. Radar Polarimetry: Analysis Tools and Applications. IEEE Transactions on Geoscience and Remote Sensing, 26, 1988. [37] L . Ferro-Famil, E . Pottier, and J.-S. Lee. Unsupervised Classification of Mul t i -Frequency and Fully Polarimetric S A R Images Based on the H / A / a - W i s h a r t Classifier. IEEE Transactions on Geoscience and Remote Sensing, 39(11):2332-2342, November 2001. [38] D. Flett. C-Band Polarimetric Synthetic Aperture Radar Signatures of Winter Sea Ice Conditions. Master's thesis, University of Waterloo, 1997. [39] D. Flett, 2003. Personal communication. [40] D. Flett. Concepts of Operations for Operational Sea Ice Monitoring. Technical report, Canadian Ice Service, May 2003. [41] D. Flett. Operational Use of S A R at the Canadian Ice Service: Present Opera-tions and a Look to the Future. In Proceedings of the 2nd Workshop on Coast and Marine Applications of SAR, Svalbard, Norway, 2003. [42] D. Flett, 2004. Personal communication. [43] D. Flett, 2005. Personal communication. [44] T. Freeman and S. L . Durden. A Three-Component Scattering Model for Po-larimetric S A R Data. IEEE Transactions on Geoscience and Remote Sensing, 36(3):963-973, May 1998. [45] R. C. Gonzalez and R. E . Woods. Digital Image Processing. Addison Wesley, 1993. [46] International Ice Charting Working Group. Ice Information Services: Socio-Economic Benefits and Earth Observation Requirements. Technical report, The Group on Earth Observation (GEO) and Global Monitoring for Environment and Security ( G M E S ) , September 2004. 200 [47] International Ice Charting Working Group, 2005. http://nsidc.org/noaa/iicwg/. [48] I. Hajnsek. Inversion of Surface Parameters Using Polarimetric SAR. P h D thesis, Friedrich-Schiller-Universitaet Jena, 2001. [49] M . Hallikainen and D. P. Winebrenner. Microwave Remote Sensing of Sea Ice, chapter 3 - The Physical Basis for Sea Ice Remote Sensing, pages 29-46. A G U Geophysical Monograph 68, 1992. [50] Y . Hara, R. G . Atkins, R. T. Shin, J . A . Kong, S. H . Yueh, and R. Kwok. Application of Neural Networks for Sea Ice Classification in Polarimetric S A R Images. IEEE Transactions on Geoscience and Remote Sensing, 33(3): 740-748, May 1995. [51] S. Haykin, E . O. Lews, R. K . Raney, and J . R. Rossiter. Remote Sensing of Sea Ice and Icebergs. Wiley,. 1994. [52] T. Hirose. Evaluation of R A D A R S A T Applications and State-of-Preparedness of Users in Oceans and Sea Ice Disciplines. Technical report, Noetix Research (prepared for the R A D A R S A T Project Office), 1991. [53] I C E M O N . Project web site, 2005. www.icemon.org. [54] W . B . Tucker III, D: K . Perovich, A . J . Gow, W . F . Weers, and M . R. Drinkwa. ter. Microwave Remote Sensing of Sea Ice, chapter 2 - Physical Properties of Sea Ice Relevant to Remote Sensing, pages 9-28. A G U Geophysical Monograph 68, 1992. [55] R. Johansson. Detection and Characterization of Ice Ridges in the Baltic Sea Using CV-580 S A R Imagery. In Proc. 12th Canadian Symp. Remote Sensing, pages 386-389, Vancouver, Canada, 1989. [56] M . Johnston and D. Flett. First Year Ridges in R A D A R S A T ScanSAR Imagery: Influence of Incidence Angle and Feature Orientation. In Proceedings Interna-tional Symposium on Remote Sensing in Glaciology, College Park, Maryland, U S A , 2001. 201 [57] J . A . Karvonen. Baltic Sea Ice S A R Segmentation and Classification Using Modified Pulse-Coupled Neural Networks. IEEE Transactions on Geoscience and Remote Sensing, 42(7):1566-1574, July 2004. [58] J . A . Kong, A . A . Swartz, H . A . Yueh, L . M . Novak, and R. T. Shin. Identifi-cation of Terrain Cover Using the Optimal Polarimetric Classifier. Journal of Electromagnetic Waves and Applications, 2:171-194, 1988. [59] R. Kwok, S. V . Nghiem, S. H . Yueh, and D. D. Huynh. Retrieval of Th in Ice Thickness from Multifrequency Polarimetric S A R Data. Remote Sensing of the Environment, 51:361-374, 1995. [60] J.-S. Lee, M . Grunes, T . L . Ainsworth, L . - J . Du, D. L . Schuler, and S. R. Cloude. Unsupervised Classification Using Polarimetric Decomposition and the Complex Wishart Classifier. IEEE Transactions on Geoscience and Remote Sensing, 37(5):2249-2258, September 1999. [61] J.-S. Lee, M . Grunes, and R. Kwok. Classification of Mult i -Look Polarimetric S A R Imagery Based on Complex Wishart Distribution. International Journal of Remote Sensing, 15(11):2299-2311, 1994. [62] J.-S. Lee, M . R. Grunes, and E . Pottier. Quantitative Comparison of Clas-sification Capability: Fully Polarimetric versus Dual and Single-Polarization S A R . IEEE Transactions on Geoscience and Remote Sensing, 39(11):2343-2351, November 2001. [63] J.-S Lee, M . R. Grunes, E . Pottier, and L . Ferro-Famil. Unsupervised Terrain Classification Preserving Polarimetric Scattering Characteristics. IEEE Trans-actions on Geoscience and Remote Sensing, 42(4):722-731, 2004. [64] J.-S. Lee, R. Grunes, and G . de Grandi. Polarimetric S A R Speckle Filtering and Its Implication for Classification. IEEE Transactions on Geoscience and Remote Sensing, 37(5), September 1999. 202 [65] C. E . Livingstone, D. G . Barber, F . Spiring, and W . L i u . The SIMMS'93 S A R Polarimetry Experiment: Combined Surface and Airborne Radar Measurements of Winter Sea Ice. In Proceedings of the IEEE International Geoscience and Remote Sensing Symposium, IGARSS'96. I E E E , 1996. [66] C. E . Livingstone and M . R. Drinkwater. Springtime C-Band S A R Backscatter Signatures of Labrador Sea Marginal Ice: Measurements versus Modelling Pre-dictions. IEEE Transactions of Geoscience and Remote Sensing, 29(1):29-41, 1991. [67] C E . Livingstone, R. G . Onstott, L . D. Arsenault, A . L . Gray, and K . P. Singh. Microwave Sea Ice Signatures Near the Onset of Melt . IEEE Transactions of Geoscience and Remote Sensing, 25(2): 174 - 186, 1987. [68] C. Lopez-Martinez and E . Pottier. Statistical Assessment of Eigenvector-Based Target Decomposition Theorems in Radar Polarimetry. In Proceedings of the IEEE International Geoscience and Remote Sensing Symposium. IGARSS '04-I E E E , 2004. [69] T. Luscombe, 2003. Personal Communication. [70] T. Luscombe. R A D A R S A T - 2 Product Specifications. Technical Report R N -SP-50-9786, M D A Report, January 2003. [71] C. Maetzler and U . Wegmueller. Dielectric Properties of Freshwater Ice at Microwave Frequencies. J. Phys. D: Appl. Phys., 20:1623-1630, 1987. [72] M . Makynen and M . Hallikainen. C-Band Backscattering Signatures of Baltic Sea Ice. In Proc. Of the IEEE International Geoscience and Remote Sensing Sumposium IGARSS'98. I E E E , 1998. [73] M D A . R A D A R S A T - 2 web site, 2005. http://www.mdacorporation.com/radarsat-2/-[74] H . Melling. Detection of Features in First-Year Pack Ice by Synthetic Aperture R A D A R (SAR). Int. J. Remote Sensing, 19(6):1223-1249, 1998. 203 [75] K . Nakamura, H . Wakabayashi, K . Naoki, T . Moriyama, and S. Uratsuka. Ob-servation of Sea-Ice Thickness in the Sea of Okhotsk by Using Dual-Frequency and Fully Polarimetric Airborne S A R (Pi -SAR) Data. IEEE Transactions of Geoscience and Remote Sensing, 43(ll):2460-2469, November 2005. [76] S. V . Nghiem and C. Bertoia. Study of Multi-Polarization C-Band Backscatter Signatures for Arctic Sea Ice Mapping with Future Satellite S A R . Canadian Journal of Remoe Sensing, 27(5), October 2001. [77] S. V . Nghiem, R. Kwok, S. H . Yueh, and M . R. Drinkwater. Polarimetric Signatures of Sea Ice 1. Theoretical Model. Journal of Geophysical Research, 100(C7):13665-13679, July 1995. [78] S. V . Nghiem, R. Kwok, S. H . Yueh, and M . R. Drinkwater. Polarimetric Signatures of Sea Ice 2. Experimental Observations. Journal of Geophysical Research, 100(C7):13681-13698, July 1995. [79] S. V . Nghiem, R. Kwok, S. H . Yueh, A . J . Grow, D. K . Perovich, J . A . Kong, and C. C. Hsu. Evolution in Polarimetric Signatures of Thin Saline Ice Under Constant Growth. Radio Science, 32(1):127-151, January-February 1997. [80] C. J . Oliver and S. Quegan. Understanding Synthetic Aperture Radar Images. Artech House, 1998. [81] R. G . Onstott. Microwave Remote Sensing of Sea Ice, chapter 5 - S A R and Scat-, terometer Signatures of Sea Ice, pages 73-104. A G U Geophysical Monograph 68, 1992. [82] World Meteorological Organization. Sea-Ice Information Services in the World. Technical Report W M O No. 574, Secretariat of the W M O , Geneva, Switzerland, 2000. [83] World Meteorological Organization, 2005. http://www.wmo.int/index-en.html. 204 [84] D. Pearson, C. Livingstone, R. Hawkins, L . Gray, L . Arsenault, T . Wilkinson, and K . Okamoto. R A D A R Detection of Sea Ice Ridges and Icebergs in Frozen Oceans at Incidence Angles from 0 to 90. In Proc. 6th Canadian Symp. Remote Sensing, pages 231-237, Halifax, Canada, 1980. [85] E . Pottier and J.-S. Lee. Application of the H / A / a Polarimetric Decomposition Theorem for Unsupervised Classification of Fully Polarimetric S A R Data Based on the Wishart Distribution. In Proceedings CEOS'99, Toulouse, France, 1999. [86] D. Power, J . Youden, K . Lane, C. Randell, and D. Flett. Iceberg Detection Capabilities of R A D A R S A T Synthetic Aperture Radar. Canadian Journal of Remote Sensing, 27(5):476-486, October 2001. [87] B . Ramsay, D. Flett, H . S. Anderson, R. G i l l , S. Nghiem, and C. Bertoia. Prepa-ration for the Operational Use of R A D A R S A T - 2 for Ice Monitoring. Canadian Journal of Remote Sensing, 30(3):415-423, June 2004. [88] B . Ramsay, T. Hirose, M . Manore, J . Falkingham, R. Gale, D. Barber, M . Shokr, B . Danielowicz, B . Gorman, and C. Livingstone. Potential of R A D A R S A T for Sea Ice Applications. Canadian Journal of Remote Sensing, 19(4), November-December 1993. [89] J . A . Richards. Remote Sensing Digital Image Analysis. Springer-Verlag, Berlin, Germany, 1994. [90] E . Rignot and R. Chellappa. Segmentation of Polarimetric Synthetic Aperture Radar Data. IEEE Transactions on Image Processing, 1(3), July 1992. [91] E . Rignot, R. Chellappa, and P. Dubois. Unsupervised Segmentation of Po-larimetric S A R Data Using the Covariance Matrix. IEEE Transactions on Geoscience and Remote Sensing, 30(4):697-705, July 1992. [92] E . Rignot and M . R. Drinkwater. Winter Sea Ice Mapping From Mul t i -Parameter Synthetic Aperture .Radar. J . Glaciology, 40(134) :31-45, 1994. 205 [93] A . Rodrigues, D. Corr, K . Partington, E . Pottier, arid L . Ferro-Famil. Unsu-pervised Wishart Clasification of Sea Ice Using Entropy, Alpha and Anisotropy Decomposition. In Workshop on Applications of SAR Polarimetry and Polari-metric Interferometry, January 14 - 16 2003. [94] S. Sandven, M . Lundhaug, 0 . Dalen, K . Kloster, H . Gronvall, A . Seina, and J . Vainio. Sea Ice Monitoring by Satellites - Lectures for the C E O Training Course. Technical Report Special Report No.58, N E R S C , Bergen, 1999. [95] B . Scheuchl. Analysis of CV-580 Fully Polarimetric Data of Sea Ice - Data Cal-ibration Test. Technical Report RX-TN-51-4118, MacDonald Dettwiler, Rich-mond, B C , September 2002. [96] B . Scheuchl. Utilization of Multi-Polarization and Polarimetric Data for Sea Ice Monitoring in an Operational Environment. Technical Report RX-RP-51-4919, MacDonald Dettwiler, Richmond, B C , November 2003. [97] B . Scheuchl, R. Caves, I. G . Cumming, and G . Staples. Automated Sea Ice Classification Using Spaceborne Polarimetric S A R Data. In Proceedings of the IEEE International Geoscience and Remote Sensing Symposium, IGARSS'Ol, pages 3117-3119, Sydney, Australia, July 9-13, 2001. [98] B . Scheuchl, R. Caves, I. G . Cumming, and G . Staples. H / A / a - b a s e d Classifi-cation of Sea Ice Using S A R Polarimetry. In Proceedings of the 23rd Canadian Symposium on Remote Sensing, Quebec City, August 21-24, 2001. [99] B . Scheuchl, R. Caves, D. Flett, R. De Abreu, M . Arkett, and I. G . Cumming. E N V I S A T A S A R A P Data for Operational Sea Ice Monitoring. In Proc. IEEE International Geoscience and Remote Sensing Symposium, IGARSS'04, vol-ume 3, pages 2142-2145, Anchorage, September 20-24, 2004. [100] B . Scheuchl, R. Caves, D. Flett, R. De Abreu, M . Arkett, and I. G . Cumming. Potential of Cross-Polarization Information for Operational Sea Ice Monitoring. In Proc. ENVISAT/ERS Symposium, Salzburg, Austria, September 6-10 2004. 206 [101] B . Scheuchl and I. G . Cumming. Analysis of the Influence of N E S Z Variations on Cross-Polarized Signatures of Sea Ice. In Proc. IEEE International Geoscience and Remote Sensing Symposium, IGARSS'05, Seoul, Korea, July 25-29 2005. [102] B . Scheuchl, I. G . Cumming, and I. Hajnsek. Consolidation of a Pixel-Based Classification Using Neighbourhood Information. In Proceedings of the IEEE In-ternational Geoscience and Remote Sensing Symposium, IGARSS'03, Toulouse, July 21-25, 2003. [103] B . Scheuchl, I. G . Cumming, and I. Hajnsek. Classification of Fully Polarimetric Single- and Dual-Frequency S A R Data of Sea Ice Using the Wishart Statistics. Canadian Journal of Remote Sensing, 31(l):61-72, February 2005. [104] B . Scheuchl, D. Flett, R. Caves, and I. G . Cumming. Potential of R A D A R S A T -2 Multiple Polarization and Polarimetric S A R Data for Operational Sea Ice Monitoring. Canadian Journal of Remote Sensing, 30(3):448-461, June 2004. [105] B . Scheuchl, D. Flett, G . Staples, G . Davidson, and I. G . Cumming. Prelim-inary Classification Results of Simulated R A D A R S A T - 2 Polarimetric Sea Ice Data. In Workshop on Applications of SAR Polarimetry and Polarimetric In-terferometry, E S A / E S R I N , Frascati, Italy, January 14-16 2003. [106] B . Scheuchl, I. Hajnsek, and I. G . Cumming. Model-Based Classification of Polarimetric S A R Sea Ice Data. In Proceedings of the IEEE International Geoscience and Remote Sensing Symposium, IGARSS'02, pages 1521-1523, Toronto, June 24-28, 2002. [107] B . Scheuchl, I. Hajnsek, and I. G . Cumming. Sea Ice Classification Using Multi-Frequency Polarimetric S A R Data. In Proceedings of the IEEE Interna-tional Geoscience and Remote Sensing Symposium, IGARSS'02, pages 1914-1916, Toronto, June 24-28, 2002. [108] B . Scheuchl, I. Hajnsek, and I. G . Cumming. Classification Strategies for Po-larimetric S A R Sea Ice Data. In Workshop on Applications of SAR Polarimetry 207 and Polarimetric Interferometry, E S A / E S R I N , Frascati, Italy, January 14-16 2003. [109] B . Scheuchl, I. Hajnsek, and I. G . Cumming. Potential of S A R Polarimetry for Sea Ice Discrimination. In Proceedings of the Advanced SAR Workshop, ASAR'03, Canadian Space Agency, Montreal, June 25-27 2003. [110] S. Schuetz. Payload to ground segment interface control document. Technical Report PO-ID-64-9461, E N V I S A T Report, Apr i l 1999. [ I l l ] A . J . Sephton and K . C. Partington. Analysis of SAR Data of the Polar Oceans - Recent Advances, chapter 12 - Towards Operational Monitoring of Arctic Sea Ice by S A R , pages 259-280. Springer, 1998. [112] Canadian Ice Service. MANICE - Manual of Standard Procedures for Observing and Reporting Ice Conditions. Environment Canada, 9th edition, A p r i l 2002. [113] S.-E. Shih, K . - H . Ding, S. V . Nghiem, C . - C . Hsu, J . Kong, and A . Jordan. Th in Saline Ice Thickness Retrieval Using Time-Series C-Band Polarimetric Radar Measurements. IEEE Transactions on Geoscience and Remote Sensing, 36(5):1589-1598, 1998. [114] M . Simila, M . Lepparanta, H . Granberg, and J . Lewis. The Relation Between S A R Imagery and Regional Sea Ice Ridging Characteristics for B E P E R S - 8 8 . Int. J. of Remote Sensing, 13(13):2415-2432, 1992. [115] H . Skriver and L . T. Pederson. Polarimetric Signatures of Sea Ice in the Green-land Sea. In Proceedings of the IEEE International Geoscience and Remote Sensing Symposium, IGARSS'95. I E E E , 1995. [116] L . - K . Soh, C. Tsatsoulis, D . Gineris, and C. Bertoia. A R K T O S : A n Intelligent System for S A R Sea Ice Image Classificaiton. IEEE Transactions on Geoscience and Remote Sensing, 42(l):229-248, January 2004. 208 [117] B . B . Thomsen, S. V . Nghiem, and R. Kwok. Polarimetric C-Band S A R Ob-servations of Sea Ice in the Greenland Sea. In Proceedings of the IEEE Interna-tional Geoscience and Remote Sensing Symposium, IGARSS'98. I E E E , 1998. [118] M . Topan. R A D A R S A T - 2 Polarimetric Product Simulation Tool (RSAT2SIMU) . Technical Report R X - R P 52-1296, M D A , March 2004. [119] R. Touzi. Calibrated Polarimetric S A R Data for Ship Detection. In Proceedings of the 22nd Canadian Symposium on Remote Sensing, pages 93-96, Victoria, Canada, 2000. [120] R. Touzi and F. Charbonneau. Characterization of Target Symmetric Scatter-ing Using Polarimetric SARs. IEEE Transactions on Geoscience and Remote Sensing, 40(11):2507-2516, November 2002. [121] C. Tsatsoulis and R. Kwok. Analysis of SAR Data of the Polar Oceans, chapter 1, Recent Advances in the Analysis of S A R Data of the Polar Oceans, pages 3-8. Springer, 1998. [122] F . T. Ulaby and C. Elachi. Radar Polarimetry for Geoscience Applications. Artech House, 1990. [123] F . T. Ulaby, R. K . Moore, and A . K . Fung. Microwave Remote Sensing, Active and Passive, volume 1 - Fundamentals and Radiometry. Addison Wesley, 1981. [124] H . Wakabayashi, T. Matsuoka, K . Nakamura, and F . Nishio. Polarimetric Characteristics of Sea Ice in the Sea of Okhotsk Observed by Airborne L-Band S A R . IEEE Transactions of Geoscience and Remote Sensing, 42(ll):2412-2425, November 2004. [125] D. P. Winebrenner, L . D. Farmer, and I. R. Joughin. On the Response of Polarimetric S A R Signatures at 24 cm Wavelength to Sea Ice Thickness in Arctic Leads. Radio Science, 30(2):373-402, 1995. [126] H . Zebker and J . J . Van Zyl . Imaging Radar Polarimetry: A Review. Proceedings of the IEEE, 79(11), November 1991. 209 [127] J . J . Van Zyl . Unsupervised Classification of Scattering Behavior Using Radar Polarimetry Data. IEEE Transactions on Geoscience and Remote Sensing, 27(1), January 1989. 210 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.831.1-0065869/manifest

Comment

Related Items