UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Detection of man made targets using radar polarimetry Wasniewski, Flavio 2007

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

Item Metadata

Download

Media
831-ubc_2007-0635.pdf [ 12.32MB ]
Metadata
JSON: 831-1.0101077.json
JSON-LD: 831-1.0101077-ld.json
RDF/XML (Pretty): 831-1.0101077-rdf.xml
RDF/JSON: 831-1.0101077-rdf.json
Turtle: 831-1.0101077-turtle.txt
N-Triples: 831-1.0101077-rdf-ntriples.txt
Original Record: 831-1.0101077-source.json
Full Text
831-1.0101077-fulltext.txt
Citation
831-1.0101077.ris

Full Text

DETECTION OF MAN MADE TARGETS USING RADAR POLARIMETRY by FLAVIO WASNIEWSKI B.Sc, Universidade Estadual do Rio de Janeiro, Rio de Janeiro, Brazil, 1995 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF APPLIED SCIENCE in THE F A C U L T Y OF G R A D U A T E STUDIES (Electrical and Computer Engineering) THE UNIVERSITY OF BRITISH C O L U M B I A August 2007 © Flavio Wasniewski ABSTRACT Since the late 1970's , various synthetic aperture radar ( S A R ) satellite miss ions have p rov ided a valuable source o f informat ion about the earth's surface. P r o v i d i n g their o w n i l lumina t ion , these sensors generate imagery o f our planet 24 hours a day, regardless o f c loud cover. U n t i l recently, except for the short term space shuttle S I R - C system, these S A R miss ions carried s ingle-polar ized sensors, meaning that the informat ion extracted f rom the imagery, and the resul t ing qual i ty o f interpretation, was l imi ted . In 2006 the first o f a series o f planned satellites car ry ing fu l ly polar imetr ic S A R sensors was launched. In polar imetr ic S A R systems, images can be acquired w i t h hor izonta l (H) and vert ical ( V ) polar izat ions o f the electric f ie ld o n both t ransmission and reception. The result ing mul t ipo la r ized imagery makes avai lable a higher leve l o f informat ion content for a number o f applications, i nc lud ing m a n made target detection. T h e potential o f polar imetr ic S A R for m a n made target detection is investigated i n this thesis. In the m a n made target detection f ie ld , the purpose is to differentiate the signature o f the targets f rom those o f adjacent natural areas. The most operat ional ly successful appl icat ion i n this f ie ld is ship detection, as the sea surface usua l ly forms a quite homogeneous and easi ly dist inguishable clutter. Vegetated surfaces, however , usua l ly constitute a more chal lenging clutter, as the backscatter levels can be ve ry h igh and the signatures diverse, often be ing confused w i t h those o f m a n made targets. A few algori thms have been developed w i t h the purpose o f separating the targets from clutter. A methodology that employs the f o l l o w i n g three algori thms was recently tested i n order to discr iminate crashed aircraft f rom the surrounding terrain: Polar imet r ic W h i t e n i n g F i l t e r ( P W F ) , E v e n B o u n c e A n a l y s i s and C a m e r o n Decompos i t i on . In these tests, successful results were achieved when the terrain was composed o f a homogeneous f ie ld o f grass and the targets were crashed airplanes. i i In this w o r k the same methodology is tested w i t h different m a n made targets and w i t h different clutters. A l s o , methodologies i n v o l v i n g the use o f two other algori thms are tested i n order to reduce the false a larm rates. These algori thms are Coherence Test o f the Symmet r i c Scattering Character izat ion M e t h o d ( S S C M ) and Freeman-Durden Decompos i t ion . T h e methodologies are appl ied to three different data sets acquired over different Canadian locations - Ot tawa ( O N ) , Gage town ( N B ) and V a n c o u v e r ( B C ) - w i t h the C V - 5 8 0 polar imetr ic S A R system and the false a larm rates are assessed. Resul ts show that the Coherence Test can lower false a larm rates on h igh vegetation clutter when appl ied i n combina t ion w i t h other algori thms, w h i l e Freeman-Durden D e c o m p o s i t i o n does not perform effect ively i n the same experiments. i i i T A B L E O F C O N T E N T S Abstract » List of Tables vi List of Figures vii Acknowledgements x Dedication xii Chapter 1 - Introduction 1 1.1 Radar Polarimetry 1 1.2 Thesis Objectives and Outline 6 Chapter 2 - Foundations of Radar Polarimetry 9 2.1 EM Waves and Their Polarization 9 2.1.1 EM Waves 9 2.1.2 Linear Polarization 10 2.1.3 Elliptical Polarization 10 2.1.4 Partial Polarization 12 2.2 Interaction of EM Waves with Scattering Objects 14 2.2.1 Introduction 14 2.2.2 Scattering Mechanisms 14 2.2.3 The Scattering Matrix 16 2.2.4 The Covariance Matrix 17 2.2.5 The Coherency Matrix 17 2.3 Radar Systems and Polarimetry 18 2.4 Polarization Synthesis 19 2.5 Polarization Signatures 20 Chapter 3 - The Detection of Man Made Targets 22 3.1 Man Made Targets in Radar Polarimetry 22 3.2 Target Detection Methods 23 3.2.1 Introduction 23 3.2.2 Coherent and Non-Coherent Target Decompositions 23 3.2.3 Some Target Detection Methods 24 Chapter 4 - Methodology and Detection Algorithms 25 4.1 Methodology Used in This Work 25 4.2 The Polarimetric Whitening Filter (PWF) 31 4.2.1 Introduction : 31 4.2.2 Speckle and Polarimetric Speckle Filtering 32 4.2.3 Mathematical Description 32 4.3 Even Bounce Analysis 34 iv 4.4 Cameron Decomposition 36 4.5 The Coherence Test of the Symmetric Scattering Characterization Method (SSCM) 40 4.6 The Freeman-Durden Decomposition 42 Chapter 5 - Datasets 44 5.1 Datasets Used 45 5.1.1 The Ottawa Dataset 45 5.1.2 The Westham Island Dataset 46 5.1.3 The Gagetown Dataset 48 5.2 CV-580 SAR System Specifications and Data 50 5.2.1 The CV-580 SAR System 50 5.2.2 Data Format 52 Chapter 6 - Data Processing Results and Analysis 53 6.1 Selected Targets 53 6.2 Results 55 6.2.1 Results for Target 7 (House Among Trees) 58 6.2.2 Results for Target 14 (House in Forest) 63 6.2.3 Results for Target 21 (House in Forest) 68 6.2.4 Results for Target 1 (Two Vertical Cylinders) 73 6.2.5 Results for Target 2 (Plow in Grass) 75 6.2.6 Results for Target 4 (Large Farm Cart) 77 6.2.7 Results for Target 5 (Horizontal Cylinders) 79 6.2.8 Results for Target 12 (Artillery Pieces in Gagetown) 81 6.2.9 Results for Target 20 (Crashed Airplane) 83 6.3 Algorithm Threshold Values 85 6.4 General Analysis 87 Chapter 7 - Conclusions 90 7.1 Summary 90 7.2 Research Contributions 92 7.3 Future Work.... 93 BIBLIOGRAPHY 95 Appendix A - Supporting processing results 99 A.l - Target 1 100 A.2-Target 2 102 A.3-Target4 104 A.4-Target 5 106 A.5-Target 12 108 A.6-Target 20 110 v LIST OF TABLES Table 1: S ing le polar ized spacebome miss ions 2 Table 2: Current and planned spaceborne polar imetr ic radar systems 4 Table 3: z values for the elemental scatterers and their scattering matrices 39 Table 4: Target and clutter descriptions and locat ions 53 Table 7: Target 21 - False A l a r m count and False A l a r m Rate 69 Table 8: Target 1 - False A l a r m count and False A l a r m Rate 74 Table 9: Target 2 - False A l a r m count and False A l a r m Rate 76 Table 10: Target 4 - Fa lse A l a r m count and False A l a r m Rate 78 Table 11: Target 5 - False A l a r m count and False A l a r m Rate 80 Table 12: Target 12 - False A l a r m count and False A l a r m Rate 82 Table 13: Target 20 - False A l a r m count and False A l a r m Rate 84 Table 14: Optimal thresholds when Methodologies 1, 2 and 3 are applied 85 Table 15: O p t i m a l thresholds when M e t h o d o l o g y 4 is appl ied 86 Table 16: False a larm count for l o w , m e d i u m and h igh vegetation types, total false a larm count and total false a larm rate (false alarms/ k m 2 ) for each methodology 88 v i LIST OF FIGURES Figure 1: S w a m p area i n Gage town 5 Figure 2: Q u i c k b i r d and R A D A R S A T -1 images over the S w a m p area i n Gage town 6 Figure 3: Po la r iza t ion el l ipse i n the x - y plane 11 Figure 4: Representation o f three scattering mechanisms 15 Figure 5: Examples o f synthesized images 20 Figure 6: C o - p o l and cross-pol polar iza t ion signatures o f a forested area 21 Figure 7: F l o w chart o f detection algori thms (Methodologies 1 - 4) 30 Figure 8: E v e n B o u n c e image o f Target 1 35 Figure 9: Cameron ' s unit d isc 40 Figure 10: Poincare Sphere 41 Figure 11: F reeman-Durden decomposi t ion appl ied to s imulated images 43 Figure 12: T h e C o n v a i r 580 S A R airplane 44 Figure 13: O v e r v i e w o f the Ot tawa dataset and photograph o f the crashed airplane 45 Figure 14: O v e r v i e w o f Wes tham Island 47 Figure 15: O v e r v i e w o f Gage town data 48 Figure 16: Ikonos image acquired o n Oc t 4th, 2001 over Gage town 49 Figure 17: Ikonos, C V - 5 8 0 data and photograph o f Gage town data 50 Figure 18: T a i l section o f C o n v a i r - 5 8 0 showing S A R radomes. (Source: C C R S , 2002).51 Figure 19: C o n v a i r Sys tem configurat ion (Source: C C R S , 2002) 51 Figure 20: Photos o f the 9 targets selected for this w o r k 54 Figure 21: Target 7 - R G B composi te and photograph 58 v i i Figure 22: Target 7 - F i n a l detection maps for Methodo log ies 1 to 4 60 Figure 23: Target 7 - Plots f rom methodology 1 61 Figure 24: Target 7 - Plots f rom methodology 2 62 Figure 25: Target 7 - Plots f rom methodology 3 and 4 62 Figure 26: Target 14 - R G B composi te and photograph 63 Figure 27: Target 14 - F i n a l detection maps for Methodo log ies 1 to 4 65 Figure 28: Target 14 - Plots f rom methodology 1 66 Figure 29: Target 14 - Plots f rom methodology 2 67 Figure 30: Target 14 - Plots f rom methodology 3 and 4 67 Figure 31: Target 21 - R G B composi te and photograph 68 Figure 32: Target 21 - F i n a l detection maps for Methodo log ies 1 to 4 70 Figure 33: Target 21 - Plots f rom methodology 1 71 Figure 34: Target 21 - P lo ts f rom methodology 2 72 Figure 35: Target 21 - Plots f rom methodology 3 and 4 72 Figure 36: Target 1 - R G B composi te and photograph 73 Figure 37: Target 1 - F i n a l detection maps for Methodo log ies 1 to 4 74 Figure 38: Target 2 - R G B composi te and photograph 75 Figure 39: Target 2 - F i n a l detection maps for Methodo log ies 1 to 4 76 Figure 40: Target 4 - R G B composi te and photograph 77 Figure 41: Target 4 - F i n a l detection maps for Methodo log ies 1 to 4 78 Figure 42: Target 5 - R G B composi te and photograph. Left 79 Figure 43: Target 5 - F i n a l detection maps for Methodo log ies 1 to 4 80 Figure 44: Target 12 - R G B composi te and photograph 81 Figure 45: Target 12 - F i n a l detection maps for Methodo log ies 1 to 4 82 Figure 46: Target 20 - R G B composi te and photograph 83 v i i i Figure 47: Target 20 - F i n a l detection maps for Methodo log ies 1 to 4 84 Figure 48: Target 1 - Plots f rom methodology 1 100 Figure 49: Target 1 - Plots f rom methodology 2 101 Figure 50: Target 1 - Plots f rom methodology 3 and 4 101 Figure 51: Target 2 - Plots f rom methodology 1 102 Figure 52: Target 2 - Plots f rom methodology 2 103 Figure 53: Target 2 - Plots f rom methodology 3 and 4 103 Figure 54: Target 4 - Plots f rom methodology 1 104 Figure 55: Target 4 - Plots f rom methodology 2 105 Figure 56: Target 4 - Plots f rom Methodolog ies 3 and 4 105 Figure 57: Target 5 - Plots f rom methodology 1 106 Figure 58: Target 5 - Plots f rom methodology 2 107 Figure 59: Target 5 - Plots f rom methodology 3 and 4 107 Figure 60: Target 12 - Plots f rom methodology 1 108 Figure 61: Target 12 - Plots f rom methodology 2 109 Figure 62: Target 12 - Plots f rom methodology 3 and 4 109 Figure 63: Target 20 - Plots f rom methodology 1 110 Figure 64: Target 20 - Plots f rom methodology 2 I l l Figure 65: Target 20 - Plots f rom methodology 3 and 4 I l l i x ACKNOWLEDGEMENTS I owe m y gratitude to m y supervisor, D r . Ian C u m m i n g , w h o a lways had a w o r d o f good advice and academic guidance when needed and for his understanding attitude. H e was also generous both o n sharing his knowledge i n S A R and on m a k i n g h i m s e l f avai lable even after h is retirement. I a m also thankful to D r . Rabab W a r d for co-supervis ing m e i n the f inal stage o f m y thesis and for m a k i n g the Image Process ing L a b such a friendly environment for m y s e l f and for m y lab mates dur ing these years. M a n y thanks to D r . T o m L u k o w s k i , f rom the Defence Research and Deve lopment Canada ( D R D C ) , and D r . Francois Charbonneau, f rom the Canada Centre for Remote Sensing ( C C R S ) and the authors o f some o f the k e y papers ci ted i n this work , for sharing their knowledge o n radar polar imetry and for patiently answering the m a n y questions I asked. I also w i s h to thank the R A D A R S A T - 2 Group f rom Radarsat International (now M D A -G S I ) , l ed b y D r . John Hornsby , for their support and for p r o v i d i n g the Gage town C V - 5 8 0 data for m y research project. K a a n Er sah in and Colet te W a b n i t z : thank y o u for deny ing m e authorization to be unmotivated i n the most diff icul t days. Stephen, E r i n , J o A n n e , Jeff, G i l l i a n , A l i n e , G w e n d a l , Tan ia , V e r d a , K a r i n a , C la r i ce , K y l e , Pamela , Trav i s , E r i n J , W e l l i n g t o n , Nata l ia , M a r i a n and E v a are among the friends w h o deserve thanks for a l l their encouragement and for m a k i n g m e feel at home i n Vancouver . A n d so do fe l low S A R students K a a n , M i l l i e , Be rnd , S h u and Y e w L a m , w h o were very important i n this journey for a l l their help and good company. x V e r y special thanks to K e l l y H a m i l t o n for be ing both a compan ion and an inspira t ion dur ing the hard t imes o f m y thesis comple t ion per iod. A n d f ina l ly I w o u l d l i k e to thank m y fami ly for the m a n y years o f l o v e and investment i n m y education that brought me to this stage, and also for their continuous uncondi t ional support even f rom distant R i o de Janeiro, B r a z i l . F l a v i o W a s n i e w s k i The University of British Columbia August 2007 x i To my friends WoCymir, JoseCita, (Ricardb and MariCia CHAPTER 1 - INTRODUCTION 1.1 Radar Polarimetry T h e w o r d Radar is an acronym for radio detection and ranging. Radar systems transmit short pulses o f m ic rowave energy and record the strength o f the echoes received from the objects w i t h i n the radar f ie ld o f v i e w . Radar remote sensing systems use an antenna f ixed b e l o w an aircraft or spacecraft and pointed to the side. B y p r o v i d i n g their o w n i l l umina t ion i n the m i c r o w a v e por t ion o f the spectrum, these systems prov ide informat ion on ground features that neither photogrammetr ic systems nor opt ica l satellites systems - the two most c o m m o n i m a g i n g systems used i n remote sensing - can provide . D u e to this characteristic and to its capabi l i ty o f penetrating the atmosphere under v i r tua l ly a l l condit ions, i nc lud ing poor weather and night imaging , Synthet ic Aper ture Radar ( S A R ) has been, especia l ly i n the last three decades, a f ie ld that attracted both research and commerc ia l interest i n remote sensing. T h e development o f airborne and spaceborne earth observation S A R systems and the consequential increase i n data ava i lab i l i ty have led to a h igh demand for technical expertise i n the area. T h e R A D A R S A T - 1 , E R S - 1 , J E R S - 1 and ERS-2 satellites are some o f the most representative systems o f this era (see Tab le 1), and some o f these satellites are s t i l l operating. A m o n g these systems S I R - A , S I R - B and S I R - C were short duration miss ions (3 to 10 days) on board the N A S A Space Shuttle. These S A R systems transmit and receive waves i n the same polar iza t ion state. T h e measured data are t yp i ca l ly the ampli tude and differential phase o f the received wave as compared to the transmitted one. However , as a result o f its interaction w i t h the earth's surface, the received wave can have its polar iza t ion state changed into one the receiver m a y be par t ia l ly or comple te ly b l i n d to. 1 Sensor Country Launch Polarization Band S E A S A T U S A 1978 H H L S I R - A U S A 1981 H H L S I R - B U S A 1984 H H L A L M A Z Soviet U n i o n 1991 H H S E R S - 1 E U 1991 V V C JERS-1 Japan 1992 H H L S I R - C U S A 1994 Q u a d X , C , L E R S - 2 E U 1995 W C R A D A R S A T - 1 Canada 1995 H H C Table 1: S ing le polar ized spaceborne miss ions (wi th the except ion o f S I R - C ) . In order to be able to measure the received waves i n different polar iza t ion states and to increase the informat ion content o f S A R data, polar imetr ic S A R ( w h i c h i n this thesis w i l l be referred to as polarimetric radar) systems were developed. These systems can transmit and receive waves i n two orthogonal polar iza t ion states, and therefore provide more informat ion o n the phys ica l attributes o f the target. Polar imet r ic radar systems generate data i n four channels, corresponding to the combinat ions o f the transmitted and received hor izonta l (H) and ver t ical ( V ) waves. These channels are t yp i ca l ly named H H , H V , V H and V V . A n i l lus t ra t ion o f the dependence o f the radar 's response to polar iza t ion is shown i n F igure 1, where three C-band images ( H H , H V and V V ) acquired b y the C V - 5 8 0 airborne system over Gage town ( N B , Canada) are shown a long w i t h a co lour composi te o f them, for better v i sua l iza t ion . F igure 2 shows the same area imaged b y the opt ical satellite system Q u i c k b i r d and b y the C-band , H H pola r ized R A D A R S A T - 1 satellite for compar ison. T h e opt ical image is a R G B colour composi te o f the R e d , Green and B l u e bands ( Q u i c k b i r d channels 3, 2 and 1, respectively) and is therefore ve ry close to the terrain's natural colours (Figure 2(a)). A v i sua l compar ison w i t h the polar imetr ic colour composi te - F igure 1(d) - is a snapshot o f h o w different their informat ion contents are. Differences between the H H , H V and V V 2 channels are easi ly seen i n the R G B colour composi te ( H H - R , H V - G and W - B ) . T h e R A D A R S A T - 1 image shown i n F igure 2(b) and the C V - 5 8 0 image shown i n F igure 1(a) both have the same polar iza t ion and have a h i g h degree o f s imi lar i ty . T h e differences between them are due to their different resolutions, inc idence angles and look directions. Ter ra in differences such as moisture content cou ld also p l ay a role as the images were acquired o n different dates. T h e research o n radar polar imetry has received important contributions at least since the 1940's , w h e n George S inc la i r showed that a radar target can change the incident wave ' s polar iza t ion and have its properties expressed as a 2 x 2 scattering matr ix . In 1952 K e n n a u g h demonstrated that there are polar iza t ion states for w h i c h the radar receives m a x i m u m or m i n i m u m power, and h o w these states can be op t imized [3]. H u y n e n fo l l owed i n the 60 's showing h o w the targets phys ica l structure and geometry cou ld be determined b y the radar wave ' s backscattered polar iza t ion states, and i n the late 70 ' s Boerner ' s extension o f K e n a u g h ' s w o r k o n opt imal monostat ic polar iza t ion to the bistatic case was inf luent ial i n the addi t ion o f polar imetr ic capabi l i ty to the c i v i l and mi l i t a ry radars [3]. N A S A ' s Jet P ropu l s ion Labora tory ( J P L ) bui l t the first fu l ly polar imetr ic radar i n 1985. T h i s system was upgraded to the A I R S A R system i n the early 90 ' s , w h i c h was fo l lowed b y s imi la r systems f rom research institutions i n Canada ( C V - 5 8 0 ) , Denmark ( E M I - S A R ) and others. W i t h the data acquired b y these airborne sensors, recent research has been able not o n l y to advance the theory but to start b r idg ing the gap between theory and successful operational applications. T h e recent launch o f the first three satellites car ry ing polar imetr ic radar sensors and the approaching launch o f others (see Tab le 2) w i l l increase the data ava i lab i l i ty and provide extensive and per iodica l coverage o f the earth's surface, b r ing ing p romis ing improvements i n radar remote sensing applicat ions compared to the current s ingle po la r ized satellites. Some o f these fields are: sea ice c lassi f icat ion, ship detection, crop type and condi t ion classif icat ion, so i l moisture, search and rescue and forest management [6] [7]. W h i l e dual polar iza t ion systems can acquire simultaneous images i n two different polar izat ions, quad polar iza t ion systems can acquire four simultaneous images ( H H , H V , W , V H ) . 3 These four channels can provide complete radar scattering information of the earth's surface, for the given frequency, incidence angle and look direction. Sensor Country Launch Polarization Band E N V I S A T / A S A R E U 2002 Dual C A L O S / P A L S A R Japan 2006 Quad L Ter raSAR-X Germany 2007 Quad X R A D A R S A T - 2 Canada 2007 Quad C R I S A T India 2008 Quad C COSMOS-Skymet Italy 2008 Quad X S A O C O M Argentina 2008 Quad L Table 2: Current and planned spaceborne polarimetric radar systems. 4 5 (a) (b) Figure 2: Q u i c k b i r d and R A D A R S A T -1 images over the S w a m p area i n Gage town (same area as i n F igure 1) (a) Q u i c k b i r d image (b) R A D A R S A T - 1 F ine B e a m M o d e (C-band). 1.2 Thesis Objectives and Outline The noticeable increase o f research i n radar polar imetry applications i n the last few years led to the development o f a number o f algori thms and methodologies i n various fields. The detection o f man made targets, encompassing many possible applications such as ship detection, search and rescue, mapp ing and survei l lance, are some o f these fields, and are expected to benefit s ignif icant ly from the advent o f the spaceborne polar imetr ic radars. The m a i n purpose o f this w o r k is to experiment w i t h a few o f the algori thms that can potential ly be used for the detection o f man made targets and describe their performance 6 for different targets and clutters. The results that will be obtained in this thesis with airborne data are expected to be, to a large extent, applicable to datasets that will be generated by future radar satellites mentioned in Section 1.1, providing a contribution to the users of polarimetric radar data. Target detection involves discriminating a discrete target within a vegetation clutter and identifying it as a likely man made target. This is all that is required for some applications, while for others it is necessary to add one more step: target recognition. The development of Automated Target Recognition (ATR) systems is a very specialized field of research, and typically relies on high resolution data (0.3 - lm). Some of the algorithms used here might help with the identification when the targets are large compared to the sensor resolution. However, target recognition is not within the objectives of this work. In order to test the polarimetric radar detection capabilities, a specific methodology will be used as a starting point: the use of polarimetric imagery for the detection of crashed airplanes, as used by Tom Lukowski and others [8] [9] [10]. This methodology uses three main algorithms applied with a constant false alarm rate (CFAR) detector and image morphology processing in order to detect crashed aircraft in a homogeneous clutter of grass. These algorithms are Polarimetric Whitening Filter (PWF), Cameron Decomposition (CD) and Even Bounce Analysis. In this work this methodology will be referred to as D C A (detection of crashed airplanes). In this thesis, the D C A methodology is applied to more generic situations: instead of crashed airplanes only, any man made feature is accepted as a target, and instead of grass only, any kind of vegetation is accepted as clutter. In some of these situations, particularly when the clutter is composed of high vegetation, it is reasonable to expect an increase in the false alarm rates. In order to keep these rates at a low level, and in order to test other algorithms' effectiveness for man made target detection, two other methodologies are proposed and tested here: 7 First , P W F and E v e n B o u n c e A n a l y s i s are replaced i n the D C A methodology b y Coherence Test, w h i c h is part o f the Symmet r i c Scattering Character izat ion M e t h o d ( S S C M ) [16]. Second, the Freeman-Durden decomposi t ion is appl ied us ing the percentage o f dihedral response as a threshold and also fo l lowed b y morpho logy processing. T h e methodologies and algori thms used are described i n Chapter 4. Chapter 2 describes the most relevant pr inciples related to radar polar imetry and used i n this work , f rom the behavior o f E M wave propagation and its interaction w i t h surfaces to the various matrices used i n the descript ion and process ing o f polar imetry data. Chapter 3 provides a general descr ipt ion o f target decompos i t ion i n polar imetry and a few o f the techniques resul t ing f rom recent developments i n the target detection f ie ld . T h i s provides context to the methods described i n Chapter 4. Chapter 5 describes the datasets used i n this work , a b r i e f descr ipt ion o f the data acquis i t ion platform ( C V - 5 8 0 ) and the data format. F i n a l l y , details o n the targets, processing, results, analysis and conclus ions are p rov ided i n Chapters 6 and 7. 8 CHAPTER 2 - FOUNDATIONS OF RADAR POLARIMETRY 2.1 E M Waves and Their Polarization 2.1.1 E M Waves Electromagnetic (EM) Waves are composed of electric (E) and magnetic (H) fields. These fields are perpendicular to each other, and their relationship and description at any position and time satisfy Maxwell's equations [1]. Accordingly, an electromagnetic wave can be fully characterized by its electric field vector, described by the following equation [2]: E(r,t) = Exx + Eyy = [ax.exp(j8x)x + ay.exp(j8y)y].exv[j(wt - kz)] (1) where: w is the angular frequency of the radar wave; 8X and 8 are the x and y phases of the is components. The real vector Real( E ) has the following Cartesian components: is, = ax COS(T + 8x) E2=aycos(T + 8y) (2) where r = wt - kz . If we eliminate the parameter r in (2), we get the following equation of a conic: (3) where 5 = 3-5,,. 9 A s the radar wave (1) can be considered monochromat ic and fu l ly po lar ized , ax, ay and 8 are constant, m a k i n g (3) the equation o f an el l ipse {fully polarized is defined i n Sect ion 2.1.4). T h i s shows that, as the wave propagates, its electric f ie ld vector tip traces an el l ipse. 2.1.2 Linear Polarization One o f the important properties o f plane E M waves is polarization, w h i c h is the w a y an electromagnetic wave oscil lates as it travels. T h e polar iza t ion o f a wave can be linear, c i rcular or, more generical ly, e l l ip t ica l . I f w e take a plane perpendicular to the d i rec t ion o f propagation, a l inear ly po la r ized wave w i l l trace a straight l ine on this plane. T w o part icular cases, a ve r t i ca l ly po la r ized wave and a hor izonta l ly polar ized one, w i l l respect ively trace a ver t ical l ine and a hor izonta l l ine . T h e directions H (horizontal) and V (vertical) are c o m m o n l y used i n polar imetry instead o f the x and y directions. In l inear ly po la r ized waves the two components w i l l present a phase difference 8 = 0 or a mul t ip le number o f n. 2.1.3 Elliptical Polarization Instead o f t racing a ver t ical or hor izontal l ine on a plane perpendicular to the direct ion o f propagation, the wave cou ld trace a c i rc le or an el l ipse - w h i c h correspond, respectively, to c i rcular or e l l ip t ica l polarizat ions. It can also have a d i rec t ion o f rotation, w h i c h can be right-handed or left-handed - respect ively c lockwise or counterc lockwise , for an observer l o o k i n g i n the d i rec t ion o f propagation. T h e e l l ip t ica l shape that w o u l d be traced b y the wave ' s E vector, whether l inear or e l l ip t i ca l , is ca l led the wave ' s polar iza t ion el l ipse (Figure 3). T h e length o f the semi-major 10 axis of the ellipse characterizes the wave's amplitude, and its polarization is defined by two angles: the ellipticity x a n d the orientation y/. ELLIPSE Figure 3: Polarization ellipse in the x-y plane. The ellipticity describes the "fatness" of the polarization ellipse. It is measured from the major axis of the ellipse to the line connecting the intercept of the major and minor axes with the outline of the ellipse, and it can range from -45° to +45°. When the ellipticity angle is 0 degrees, the polarization is linear, and when it is 45°, the polarization is circular (+45° for left circular and -45° for right circular). The ellipticity % c a n be defined as: Z = avctan(an/a^) (4) The orientation angle describes the "tilt" of the ellipse. It is measured from the horizontal axis (x, in Figure 3) to the major axis of the ellipse, and can range from 0 to 180 degrees. 11 T h e polar iza t ion angles can be expressed i n terms o f the wave parameters ax, ay and 8 by : tan 2y/ = (tan 2a) cos 8 s in 2% = (s in 2a) s in 8 where a is defined by : tana-ay/ax (6) 2.1.4 P a r t i a l P o l a r i z a t i o n A wave is said to be comple te ly polar ized when its orientation and e l l ip t ic i ty are constant. W h e n its polar iza t ion state is t ime dependent, the wave is either comple te ly unpolar ized or par t ia l ly polar ized . The par t ia l ly po lar ized wave can be understood as a sum o f a comple te ly po la r ized and a comple te ly unpolar ized wave, and has therefore a degree o f polar iza t ion. Par t ia l ly po la r ized waves are w e l l characterized b y mathematical relations among the Stokes vector elements. The Stokes vector describes the polar iza t ion state o f an E M wave and is : " S o " ~\Ev\2+\Eh\2 S0 Q \Ev\2-\Eh? S0 cos 2y/ cos 2% u 2Re{EX) S0 s in 2y/ cos 2x V _2lm{EX) . 5 0 s i n 2 ^ (7) The comple te ly po la r ized case fo l lows the relat ion: S02 = Q2+ U2+ V2 (8) 12 where 5 0 2 is the total power, and the terms Q', if and V2 o f the Stokes vector g ive the po la r ized part o f the power. That explains the equal i ty w i t h a comple te ly polar ized wave , and the f o l l o w i n g inequal i ty for a par t ia l ly po la r i zed wave: S02 > Q2+ U2+ V2 (9) T h e degree o f polar iza t ion (d) is shown be low. It has an important phys i ca l meaning: it is related to the pur i ty o f the scattering mechan i sm ( w h i c h is defined i n i tem 2.2.2). Surface scattering results i n values o f d close to 1 w h i l e diffuse (volume) scattering results i n values o f d close to 0. T h i s loss o f polar iza t ion caused mos t ly b y terrain interaction and represented b y the l o w e r i n g o f d is ca l led depolarization. The elements Q , U and V represent respect ively the l inear polar iza t ion , orientation and ci rcular po lar iza t ion contents o f the wave [5]. Therefore, more specif ic relations can be defined, such as the degree o f l inear polar iza t ion (equation 11) and the degree o f c i rcular polar iza t ion (equation 12): d = (10) (11) V (12) o 13 2.2 Interaction of EM Waves with Scattering Objects 2.2.1 Introduction Afte r interacting w i t h the incident wave , each scatterer o n the ground w i l l produce a scattered wave w h i c h cou ld have a different polar iza t ion from the incident wave. F o r any g iven p i x e l , the resul t ing scattered wave is a vector addi t ion o f a l l the i n d i v i d u a l scatterers waves i n that p i x e l and w i l l have contributions i n both the ver t ical and hor izonta l axes. T h e orthogonal contributions o f the signal received b y the radar can be described b y different types o f matrices, w h i c h contain informat ion about the ampli tude and phase o f the scattered wave i n both directions. Polar imetr ic radar scattering informat ion can be stored i n a scattering matr ix , a covariance matr ix , a coherency mat r ix or a Stokes matr ix . These matrices are expla ined b e l o w i n sections 2.2.3 to 2.2.6. T h e coherency and covariance matrices contain informat ion o n correlat ion properties o f the scatterers and are therefore more convenient for manipu la t ion o f data conta ining distributed targets such as clutter. 2.2.2 Scattering Mechanisms T h e different ways the ground features scatter the incident waves can be expressed as a compos i t ion o f s imple scattering behaviors. Those behaviors, represented b y elemental scatterers, are ca l led scattering mechanisms. Some examples o f scattering mechanisms are: Sphere (single bounce), dihedral (double bounce), he l ix , tr ihedral (odd bounce), cy l inder and d ipole (a random dis t r ibut ion o f dipoles represents diffuse scattering). F igure 4 shows sketches o f three c o m m o n scattering mechanisms. These scattering mechanisms are expressed as specif ic scattering matrices, and can therefore be retrieved when fu l ly polar imetr ic informat ion is avai lable. Identifying the 14 dominant scattering mechanisms in each resolution cell gives information on the target structure [14]. This identification can be a complex task, which is made easier by knowing the result of the wave's interaction. For example: a double bounce typically causes a 180 degrees shift in the phase difference between the vertical and horizontal components of the wave. In general, but not necessarily, the higher the number of bounces - as in diffuse scattering - the more l ikely the wave wi l l have its polarization state transformed, a phenomenon called repolarization [3]. A lso , the larger wi l l be its unpolarized component. Figure 4: Representation of three scattering mechanisms: diffuse canopy scatter (top), double-bounce scatter (middle) and surface scatter (bottom). Adapted from Freeman and Durden [18]. 15 2.2.3 The Scattering Matrix For a radar system that coherently transmits and receives both horizontal and vertical polarizations, a complete scattering matrix can be constructed for each pixel. Both the amplitude and phase for each channel are recorded. The elements of the scattering matrix are functions of the radar wavelength and the illuminating and scattering geometries. There w i l l be a separate complex scattering matrix for each pixel of the image, but other data representations derived from this matrix can be more directly useful in data analysis. The scattering matrix also contains noise, which is typically introduced at the radar receiver. The scattering matrix can be represented as the following relation between the electric fields of the incident and scattered waves: X 5 " (13) where the superscripts " z " stand for "incident" and "s" for "scattered", and subscripts h and v stand for horizontal and vertical. In the monostatic case (when the radar transmitter and receiver are located in the same spot), the reciprocity principle dictates that 5/, v = Svf,, thus reducing the number of independent parameters. From now on, only the monostatic case wi l l be considered in this thesis. 16 2.2.4 The Covariance Matrix Many forms of analysis are performed more efficiently when power representations of the scattering matrix are used, and the Covariance Matrix is one of the most common ones. This matrix is the inner product of the scattering vector with itself. The scattering vector has the same components of the scattering matrix, but in vector form: °hh V25 hv (14) and therefore the covariance matrix has the following form: Jhh -J2S„S. ^ S h v S h h s s* vv hh hh^hv 2 hv\ 25, <j2SmS, ^hh^vv •J2S, S* v hv vv In |2 (15) vv hv where + denotes conjugate transpose and * the conjugate. As this matrix is formed by the products of the backscatter measured in the antenna (voltage units) it relates the power of the incidence and the scattered waves. 2.2.5 The Coherency Matrix Another commonly used matrix is the "coherency matrix". Sometimes the interpretation of the physical scattering mechanisms is easier using this matrix, which is: T T 2 U+2W(ShXv) + K |2 'hh^vv) 2ShXh+2ShvSl \Shh\ ~2fi(ShhSvv)-\Svv\ 2ShhShv+2SvvShv 2 I r~i |2 / r% 1-.* x . I r-* |2 « r i ri* O O O * hh hv vv hv Shh +2}3(ShhS*vv)- Svv Shh -2^(ShhS*vv)+ S1 hh'-'vvJ (16) 4S, 17 This matrix is obtained by using the following vector: V2" 2S hv (17) Both this matrix and the covariance matrix are by definition hermitian matrices and have the same real non-negative eigenvalues, but different eigenvectors. The covariance matrix and the coherency matrix are linearly related. 2.3 Rada r Systems and Polar imetry Radar polarimetry tries to determine the polarization content of every resolution cell in the image as a way to derive information on the scatterers on the ground. In order to fully characterize a scatterer, the full scattering matrix must be measured. This requires the radar system to transmit and receive at two orthogonal polarizations, and that the system retains the coherent phase information for the transmitted and received waves. The transmitting antenna of the radar system determines the polarization of the emitted wave, and the receiving antenna determines which polarization of the returned signal will be recorded. In fully polarized data there are four polarization combinations based on the systems transmit/receive polarization. These linear polarization combinations are named using the following convention: H H = horizontally polarized transmit and received signal, H V = horizontally polarized transmit signal and vertically polarized received signal, V H = vertically polarized transmit signal and horizontally polarized received signal, V V = vertically polarized transmit and received signal. 18 Different transmit polarizations can cause radar waves to interact differently with the surface and produce different returns. Therefore, when interpreting surface characteristics from a radar image it is important to know the polarization combination used to collect the image. In radar polarimetry the amplitude and phase of the four possible combinations are used to extract terrain information. Like-polarized images ( W and HH) will in most situations have stronger returned signals and higher signal to noise ratios than cross polarized images (VH and HV). This is because natural scatterers, on average, only repolarize a small amount of the transmitted wave. Many differences in contrast between differently polarized images can be noticed for various types of surfaces. When the plane of polarization of the transmitted wave is parallel to the dominant plane of linear features on the surface, the like-polarized radar return will be stronger than if the radar wave had the orthogonal polarization. As an example, some crops like wheat, which have a vertical structure, will result in a stronger signal in the VV channel, and appear more clearly in the associated W image. 2.4 Polarization Synthesis Once the scattering matrix is known, one can synthesize the radar response for any possible transmit and receive polarizations. This is called polarization synthesis, and is used to construct images in different polarization combinations. The synthesized image can be constructed by computing the Stokes matrix from the scattering matrix and then pre-multiplying and post-multiplying it by the unit Stokes vector (see Section 2.1.4) that contains the chosen receive and transmit polarizations. 19 (a) (b) F i g u r e 5: Examples o f synthesized images, (a) V H pola r ized image, (b) R igh t -Lef t c i r cu la r ly po lar ized synthetic image. The data was acquired b y the C V - 5 8 0 S A R system. In radar polar imetry applicat ions, creating a synthetic set o f images that enhance the features to be detected can be ve ry useful. F o r discrete targets, as an example , the target-to-clutter ratio can be m a x i m i z e d b y selecting the appropriate polar iza t ion combinat ion , therefore increasing the detectabili ty o f the target. 2.5 P o l a r i z a t i o n S igna tu re s A polar iza t ion signature is a w a y to v i sua l ize the response o f a target for different possible transmit and receive polar izat ions. T h e y are a plot o f the backscatter received as a function o f four independent variables: the e l l ip t ic i ty and orientations o f the incident wave and the e l l ip t ic i ty and orientation o f the rece iv ing wave . T h e plot is a surface plot, as i n F igure 6, mapped over a two d imens iona l g r id (%= -45deg. to 45 deg. and y/ = 0 deg. to 180 deg.). A s a complete m a p p i n g o f a l l the variables is too compl ica ted for interpretation, o n l y two variables are used at a t ime to portray the polar iza t ion signatures: the e l l ip t ic i ty and orientation o f the incident wave . F o r s impl i c i ty , 20 the plot shows two particular cases: The co-polarization signature, where the backscatter is computed for when the transmitted and received waves are the same polarization; and the cross-polarization signature, where the received polarizations are orthogonal to the transmitted ones. The two main features that help with the interpretation of the plots are the shape of the 3D surface and the pedestal height. The surface symmetry, and the existence and location of peaks are indicators of the scattering mechanism that originated it. The pedestal height is the minimal value of intensity of the plotted surface, and its presence indicates the existence of an unpolarized scattering component in the received signal. The larger this component is, the larger the presence of random scattering from the target. Polarization signatures are generally not very useful for target analyses as they are not clear in the absence of pixel averaging. CO-POL R E S P O N S E CROSS-POL RESPONSE Figure 6: Co-pol and cross-pol polarization signatures of a forested area in Gagetown ( N B , Canada). 21 f CHAPTER 3 - THE DETECTION OF M A N MADE TARGETS 3.1 Man Made Targets in Radar Polarimetry The concept of a feature as being a target has military origins and brings the notion of being both "significant" and surrounded by a background clutter. Typical man made targets are buildings, vehicles, power lines or train tracks, whereas typical natural targets are forests, crop fields, water and bare soils. In remote sensing, man made targets (as well as natural ones) in a scene are often perceived differently by different sensors. This wi l l depend not only on the physical properties o f the target but also on factors such as the frequency in the E M spectrum and the resolution of each sensor. In the case of polarimetric radars, these targets wi l l typically be perceived by their: characteristic changes in the phase and amplitude of the wave's orthogonal components due to the large presence of scatterers of right angles, planes or cylinders; or - higher levels of backscattering compared to the clutter. In the polarimetric analysis, man made targets can be considered as pure targets - whose signature can be completely determined by the measured scattering matrix. Man made targets are also characterized by local structure rather than texture [4], and by tridimensional geometrical shapes. Natural targets, which can be treated as homogeneous, are considered distributed targets. 22 3.2 Target Detection Methods 3.2.1 Introduction The purpose of this section is to provide an overview of the current target detection (TD) techniques used in radar polarimetry, including some of the recent research and developments. This wi l l provide the context for understanding the methodology used in this work, as described in Chapter 4. For S A R image users in general, man made targets are often expected to have a higher backscatter level than the mean of the surrounding natural clutter. Therefore, one could think that we could detect them by thresholding the magnitude o f the data. However, this alone may not produce acceptable results. This is because natural targets can sometimes have a high backscatter response and speckle noise, which can cause high statistical variability in the image. A s these factors can cause a large number of false targets, different algorithms have been developed to take advantage of the relationships that exist between phase and amplitude at different polarizations. Many of these algorithms employ target decompositions and different polarimetric discriminators in order to detect targets. Being considered pure targets, man made targets wi l l typically have their physical properties better characterized when coherent decompositions are employed, while non-coherent decompositions are commonly employed for distributed targets. 3.2.2 Coherent and Non-Coherent Target Decompositions Target decomposition techniques aim at expressing the average scattering mechanism as a sum of independent components [2]. These components wi l l usually, but not necessarily, belong to orthogonal vector spaces, and for operational purposes a decomposition should successfully associate a meaningful physical property to each component. The physical properties identified for each data pixel w i l l contain information about the imaged target. 23 Methods o f target decompos i t ion can be dis t inguished as: coherent target decompos i t ion ( C T D ) and non-coherent target decomposi t ion. W h e n the target can be characterized b y a h i g h l y po la r ized wave and can be w e l l represented b y a scattering matr ix , a coherent target decompos i t ion can be used. Some o f the m a i n coherent decompos i t ion methods are K e n n u a g h - H u y n e n C T D [24, 25] , P a u l i C T D , C a m e r o n C T D [22] and the Symmet r i c Scattering Character izat ion M e t h o d ( S S C M ) [16]. W h e n the scattered wave is par t ia l ly po lar ized and there is a h i g h var iab i l i ty i n the scattering properties among the different p ixe l s i n an area, it is more meaningful to extract informat ion f rom a pixel-averaged matr ix , such as the Stokes, covariance, or coherency matrices. These matrices do not retain the absolute phase informat ion, but the average phase difference between the polar izat ions is retained. S o m e o f the m a i n non-coherent decomposi t ion methods are: V a n Z y l [26], F r eeman-Durden [18] and C loude -Po t t i e r [27]. Chapter 4 w i l l discuss the S S C M , C a m e r o n and Freeman-Durden decompos i t ion methods i n more detail , since these methods have been appl ied i n this study. 3.2.3 Some Target Detection Methods V a r i o u s developments i n the detection o f m a n made targets have been publ i shed since the 1960's , and have received noticeable h igh act iv i ty i n recent years. E v e n though a complete survey o f methods is not i n the scope o f this work , some o f the most recent ones are ment ioned b e l o w i n an attempt to draw an approximate picture o f the state o f the art: • A d a p t i v e Polar imet r ic Target Detector us ing a polar imetr ic Genera l i zed L i k e l i h o o d Ra t io Test ( G L R T ) [28]; • Sub-aperture coherence detection us ing the 2 L - I H P ( two- look internal He rmi t i an product) a lgor i thm [29]; • Polar imet r ic Texture Discr imina tor : a c lassif icat ion scheme based on the polar imetr ic texture signature [30]. A d d i t i o n a l methods that w e study i n more detail are discussed i n Chapter 4. 24 CHAPTER 4 - METHODOLOGY AND DETECTION ALGORITHMS 4.1 Methodology Used in This Work There are different algori thms that cou ld help i n detecting m a n made targets. T h e methodology appl ied b y L u k o w s k i [10] for the detection o f crashed airplanes ( D C A ) uses a combina t ion o f three o f these algori thms. It has an important appl icat ion: to assist i n search and rescue operations i n remote areas b y p r o v i d i n g coordinates o f poss ible crashed airplanes, w i t h a l o w false a larm rate. T h e D C A methodology was tested i n fields w i t h l o w vegetation (characteristic o f m a n y remote areas, as the Canad ian prairies and most o f the arctic regions) o n both serviceable and crashed airplanes. It uses the different responses o f targets i n polar imetr ic data to discr iminate them from their surroundings. D C A uses the Pola r imet r ic W h i t e n i n g F i l t e r ( P W F ) [11], [12] to reduce the effect o f speckle w h i l e main ta in ing the target resolut ion (see Sect ion 4.2). Its output is an image that no longer holds polar imetr ic informat ion thus it is not used as input to other algori thms, but shows more c lear ly the sharp point discontinuit ies i n the clutter. These points are a first set o f potential targets, as the airplanes are supposed to have a higher backscatter than the clutter. In order to detect the targets, a constant false a larm rate ( C F A R ) is imposed , us ing a threshold dependent o n the loca l variance. T h e C F A R magnitude is calculated as: y >u + K<j (18) Where y is the intensity value o f a p i x e l , u is the mean o f the intensity image, cr is the variance o f the intensity image and K is a parameter used to set the detection threshold for each scene [23]. 25 Anothe r characteristic o f crashed airplanes is that they n o r m a l l y retain their tails intact after the fa l l , and dihedrals are thus the most dominant elemental feature present. A s a result, a strong double bounce effect can be expected, and to detect points w i t h a predominant double bounce, the E v e n Bounce a lgor i thm [13] is used (see Sec t ion 4.3). A C F A R threshold is also appl ied to the a lgor i thm and another set o f potential targets is generated. Bes ides the E v e n B o u n c e a lgor i thm, a double bounce scattering mechan i sm is also detected b y us ing the C a m e r o n decomposi t ion a lgor i thm (see Sect ion 4.4). T h i s a lgor i thm is appl ied i n order to classify the scatterer responses accord ing to their dominant scatterer type. T h e scatterer is compared w i t h the responses o f s ix elemental scatterers and mapped i n Cameron ' s unit disc shown i n F igure 9. T h e p ixe l s c lassif ied as dihedrals and narrow dihedrals are chosen as potential targets. M a n y natural features i n the images can have the same signatures as crashed airplanes -tree trunks, for example, can sometimes act as perfect dihedrals. T h e y w i l l typ ica l ly , however , have fewer p ixe l s than an airplane. Therefore, after app ly ing a detecting a lgor i thm, morpho log ica l processes can be employed i n order to dis t inguish false targets f rom real ones us ing s ize and cont inui ty as criteria. In this thesis the morpho logy processing consists of: 1 - A p p l y i n g the morpho log ica l operator " c l o s i n g " , w h i c h encompasses a d i la t ion operation fo l lowed b y an erosion operation. Here w e choose a 2x2 structuring element i n order to close a 1 -p ixe l gap between detected p ixels . 2 - E l i m i n a t i n g the clusters that are smaller than 3 p ixe l s . T h e resul t ing image is used to calculate the false a larm rate ( F A R ) . F A R is calculated as: F A R = False A l a r m s / (False A l a r m s + Correct Reject ions) 26 The number of false alarms is the number of clusters that were left after the morphology operation minus the target clusters. The F A R unit used in this thesis is the number of targets per km 2 . D C A showed promising results and its potential usefulness in operationally practical situations stimulate questions on what are the limits of this methodology and what possible extensions are there. The motivation of our present work is to address a few of these questions. These are: • Can we use the D C A methodology to detect other kinds of man made targets and in areas with different vegetation clutters? • Can the more recently developed Symmetric Scattering Characterization Method (SSCM) algorithm also be used successfully in this task? • Can a non-coherent decomposition be successfully applied for the detection of man made targets? In order to find the answers to these questions, we used different datasets (see Chapter 5), in which different types of man made targets are selected. These targets are either surrounded by or lie on the edge of different clutters: sparse, low or high vegetation. Then, the D C A methodology is applied for these different situations and its false alarm rate is assessed. In addition to the D C A , three other methodologies are proposed, tested and compared. These four methodologies are called here Methodologies 1, 2, 3 and 4 and are explained below. A flow chart for each of these methodologies is shown in Figure 7. 27 Methodology 1: This is the D C A methodology, which is applied as explained above. Methodology 2: This is a variation of the Symmetric Scattering Characterization Method (SSCM) . The S S C M algorithm was introduced in 2002 by Touzi and Charbonneau [16], who proposed the use of a coherence test followed by the application of Cameron Decomposition and the mapping of the results into the Poincare Sphere (shown in Figure 9, in Section 4.5), instead of mapping them into the Cameron's unit disc. In our work, the Coherence Test algorithm is applied as proposed in the S S C M , but Cameron Decomposition is applied as originally designed and as used in Methodology 1. Therefore, the results are mapped in Cameron's unit disc, which is a classification scheme successfully applied in previous target detection work [19], [20] and [21] . In Methodology 2 the Coherence Test generates a set of potential targets and Cameron Decomposition is applied on this set of pixels. Then, only dihedrals and narrow dihedrals are selected and the morphology processing is applied as in Methodology 1 to generate the final map of detections. False Alarm Rate is calculated from this map. Methodology 3: The aim here is to apply a non-coherent target decomposition for detection of man made targets. We apply the Freeman-Durden decomposition that classifies every pixel into one of the three following classes of scattering types - volume scattering, surface scattering and double bounce scattering. This decomposition was chosen for two reasons: first, since it was designed to detect the dihedral scattering generated by ground-trunk interaction, and it is thus reasonable to suppose that it w i l l also detect the dihedrals present in man made targets. Second, this decomposition method calculates the percentage of each scattering mechanism present in each pixel. / In this methodology, the percentage of dihedral scattering yielded by this decomposition is used as a threshold. Thus, even i f the dihedral response is not the dominant one for a 28 particular target, a lower value of the dihedral percentage can still generate a target map where the target is present. The morphology processing is also applied in this methodology and the false alarm rate is assessed. Methodology 4: This is the intersection of Methodologies 1 and 2, which means that a detection is declared if it appears in Methodology 1 and in Methodology 2. This methodology is expected to be beneficial on high vegetation clutters, when the false alarm rates of Methodologies 1 and 2 can be increased due to the high backscatter response from the tree canopies. Methodology 4 is equivalent to adding the Coherence Test algorithm to the DCA methodology. The four methodologies above are applied with various thresholds. The chosen thresholds are the ones that keep the false alarm rate to a minimum without erasing the target. For each target the optimal thresholds used in this work are reported in Chapter 6. 29 METHODOLOGY 1 Ft!t»r PVW Target Map Matrix Even Bounce Analysis Even 8ow©» Target Map ,/ Dtedrais * Target Map 1 * * Rftfrfl METHODOLOGY 2 Scattering Matrix ' SSCM Coherence Test GCteWfit Target Map Image Deletion Morphatogy * RassiJ Cameron QtwJrote* Tsrgottep METHODOLOGY 3 Seatturing. ^Covatlanc© Proeman-Ourtfen OSwxSrot* Imago Dotection Matrix 11 MnMx Decomposition Target Map Morphology Ressitt METHODOLOGY 4 METHODOLOGY 1 tmr.g« Morphology . Dotection Result METHODOLOGY a Figure 7: Flow chart of detection algorithms (Methodologies 1 - 4). 30 4.2 The Polarimetric Whitening Filter (PWF) 4.2.1 Introduction P W F (Polar imetr ic W h i t e n i n g Fi l ter) was developed to reduce the effect o f speckle o f the natural targets i n a Synthet ic Aper ture Radar ( S A R ) polar imetr ic image wi thout affecting its resolut ion [11]. In the context o f m a n made target detection, the general ro le o f P W F is to enhance the detectabili ty o f an ind iv idua l target i n clutter. It can be used i n two m a i n ways : the first is to reduce the speckle l eve l thus prepar ing the image for further process ing or image interpretation. Secondly , it can also be used direct ly as a target discr iminator . A s every speckle filter, P W F can lead to loss o f informat ion, and this is a negative point to be considered w h e n us ing it for speckle noise reduction. T h e D C A methodology [10] for target detection uses the filter for the second a i m ment ioned above. P W F is used as a d iscr iminator and the results form a set o f potential targets, instead o f a filtered image for further processing. These targets w i l l be later cross-checked against other methods used i n the detection process. In the D C A methodology a Constant False A l a r m Rate ( C F A R ) detector is appl ied to an image, us ing a threshold exper imental ly determined to generate an in i t i a l set o f targets. These targets w i l l not o n l y be m a n made targets, but it is expected that m a n made targets w i l l be among the ones detected due to the h igh target-to-clutter ratio they t yp i ca l l y present. T h i s part o f L u k o w s k i ' s methodology uses P W F as N o v a k proposed i n 1993 [10], [12]. N o v a k also proposed the M u l t i l o o k Polar imet r ic W h i t e n i n g F i l t e r ( M P W F ) and its adaptive vers ion ( A M P W F ) . These filters are not adopted here, because w e o n l y use single l ook complex ( S L C ) data. 31 4.2.2 Speckle and Polarimetric Speckle Filtering Radar imaging systems are of the coherent type, where the phases of the transmitted and received signals are carefully recorded. The return signal from one resolution cell on the ground is the result of the vector addition of reflections from all the scatterers within that resolution cell. Even i f adjacent cells on the ground visually appear to be very similar, the signal from their components may combine to result in completely different backscatter. This variation in backscatter for otherwise similar scatterers is known as speckle. Polarimetric speckle filtering is a technique that could be applied to the original radar image for reducing image speckle while preserving spatial resolution, after the image is formed. It uses the three complex channels ( H H , H V , W ) to reduce the speckle. 4.2.3 Mathematical Description The purpose of the P W F filter is to process the three channels H H , H V and W in order to obtain an intensity image with the minimum amount of speckle. A s reported above, the area chosen is homogeneous, and thus can be characterized by a complex Gaussian clutter model. This means that the expected returns H H , H V and V V can be expressed by a covariance matrix of the form: HH 0 1 0 £ p 4r o (19) r where HH E\HH\2\ is the H H power, 32 E' \HV\2 \HH\2 is the cross-pol ratio, E) W 7 = —v S is the co-pol ratio and (20) E{HH.W*} ^E\HH^yE^vf is the complex correlation. We want to find the optimal single channel (or b/w) image from H H , H V and V V . The parameter used to measure the noise (speckle) reduction performance w i l l be o/u, which is the ratio of the standard deviation of the pixel to mean of intensities: . = ^ w ( 2 1 ) where y is pixel intensity in the filtered image. Now, considering X = [ H H , H V , W ] (complex elements), we want to find a weighting matrix A that results in an image y = X ^AX whose o7p ratio is minimal. Let E{y} = tr{2ZcA)^i (22) ™{y} = tr(2ZcA)2=^2 7=1 where and A.3 are the eigenvalues of the matrix E c A. Thus, we want A such that O7LI w i l l be minimum. From the above we can derive o7u as 33 A is called the whitening filter, or P W F . It should have equal values for its eigenvalues, and thus the minimum-speckle image is: y = X*2Z~JX (24) This takes us back to the covariance matrix of the clutter given in (19). The solution y, considering the covariance matrix (19), is: \HH\2 \W\2 \HV\2 0-/W(HP| 2 ) ^uni1 \p\2)r ° m E (25) / \HH\ \W\ cos (<j>HH - <t>vv - <j>p ) 4.3 Even Bounce Analysis Even bounce analysis can contribute as a discriminator to the detection of man made targets. This is because the dihedrals normally present in those targets provide strong double bounces (see Figure 4). Whi le few dihedral structures exist in natural clutter, natural clutter tends to exhibit more odd-bounce reflected energy than even-bounce reflected energy. 34 The double bounce scattering mechan i sm creates a 180° phase shift between the H H and V V polar izat ions [13], and the f o l l o w i n g relat ion gives the strength o f the even bounce scattering: Eeven = ^ h h ~ ^ + 2\Shv\2 (26) In this w o r k an even bounce image (example shown i n F igure 8) is created us ing equation (26), and a C F A R (explained i n Sect ion 4.1) can be appl ied to this image i n order to generate an even bounce bright target map. A m a n made target stands out and can be detected b y app ly ing a C F A R to the E v e n B o u n c e image. Even Bounce Image 5 1 10I 2 0 1 2 5 1 301 3 5 1 4 0 1 100 200 300 400 500 600 Figure 8: E v e n B o u n c e image o f Target 1 (see section 6.2). 35 4.4 Cameron Decomposition The Cameron decomposition detects an elemental scatterer based on the physical scattering mechanism associated with the image backscatter, thus providing information about the structure of the scatterers. This method is based on two assumed properties of radar targets: symmetry and reciprocity. Symmetric scatterers have an axis of symmetry in the plane orthogonal to the line of sight (LOS) of the radar. In monostatic radar the transmitting and receiving antennas can be considered as being in the same position, and therefore every target can be considered reciprocal - having thus the orthogonal elements of its scattering matrix equal to each other: S = S hv vh (27) Given the scattering matrix S, its vector form can be expressed as S = 'hh 'hv 'vh (28) S can also be represented in terms of Pauli spin matrices as S = aSa+ fiSb+ 5SC (29) where a, /? and y are the Pauli components given as follows: a = $hh + $vv 4i (30) 36 y = 42Shv ( reciprocal case) (32) A c c o r d i n g to Cameron , a reciprocal target can be expressed as the sum o f two components: 5 = 4 c o s ^ ™ x + s i n r 5 ™ n ] (33) where: S?ym l s m e m a x i m u m symmetr ic component ; is the m i n i m u m symmetr ic component; cos t is the degree o f symmetry, measur ing h o w far S is f rom 5 ™ " ; and A = S (34) S £ = a S . + sSt (35) s =6cos0 + 5s\n9 (36) tan20 = S/ + P'S/(\/3\2 -\sf) (37) A n arbitrary symmetr ic scatterer can be decomposed accord ing to: S=aejp[R(¥)]k(z) (38) 37 Where a is the ampli tude o f the scattering matr ix , p is the nuissance phase and y/ is the scatterer orientation angle. T h e matr ix [R(i//)] denotes the rotation operator and A ( z ) is g iven b y The purpose o f the decompos i t ion is to find the value o f the scattering type parameter (z) for a g iven p i x e l ; this is because each elemental scatterer has its z value, and the classif icat ion proposed b y C a m e r o n compares the scattering mat r ix w i t h those o f the elemental scatterers (Table 1). In order to f ind scattering type parameter the value o f z is plotted i n the Symmet r i c Scatterer U n i t D i s c (Figure 9). T h e classif icat ion o f the p ixe ls is performed b y ca lcula t ing the distance between z for each p i x e l and the zref o f each o f the p r imi t ive scatterers: Tr ihedra l , dihedral , d ipole , cyl inder , narrow diplane, quarter wave device. Table 1 shows the value o f zreftor each o f these scatterers. 1 0 (39) 38 Target SM z Trihedral ' l 0^ 1 Dihedral / \ ,o - h -1 Dipole 0 Cylinder / V \ o s 0 0.5, 1 0.5 Narrow Diplane f l 0 1^ 0 -O . f \ ; J -0.5 Quarter Wave : ' l 0^ ^° , h i Table 3: z values for the elemental scatterers and their scattering matrices. Finally, the following relation gives the distance from a given pixel's z value to the elemental scatterer's zref. This distance w i l l tell how close they are from each other J ( z , z r e / ) = cos" 1 + z z ref 're/ (40) and which one is the closest elemental scatterer. 39 4.5 The Coherence Test of the Symmetric Scattering Characterization Method (SSCM) T h e S S C M is an extension o f C a m e r o n decompos i t ion for symmetr ic coherent scatterers. It employs the m a x i m u m symmetr ic component S™* o f the scattering matr ix as used b y Cameron . T h e S S C M scheme includes the f o l l o w i n g steps: - Ca l cu l a t i on o f the parameters a and s (as i n C a m e r o n decomposi t ion , Sect ion 4.4); - Class i f ica t ion o f target scattering as coherent or n o n coherent, us ing the degree o f coherence (equation 41) and the R i c i a n threshold [16]. R i c i a n Thresho ld is a threshold based on the R i c i a n dis t r ibut ion o f clutter; - C lass i f ica t ion o f coherent scatterers us ing C a m e r o n decompos i t ion and mapp ing the results i n the Poincare Sphere (Figure 10). 40 The Poincare Sphere is illustrated in this section in order to provide a complete overview of the S S C M method. Nevertheless, in this thesis, only the Coherence Test of the S S C M w i l l be used. The degree of coherence psym o f a target is calculated as follows: The degree of coherence p separates coherent from partially coherent scattering targets. This classification is important, as the S S C M wi l l try to characterize only the coherent targets. This characterization is done by mapping the scatterers into one half of the Poincare sphere. This provides a higher resolution mapping, according to the authors, than the Cameron's unit disc. sym (41) Dihedral * I M W a v c Figure 10: Poincare Sphere. 41 4.6 The Freeman-Durden Decomposition T h e Freeman-Durden decomposi t ion models the covariance mat r ix as the contr ibut ion o f three different scattering mechanisms: 1. Surface scattering - or s ingle bounce scattering; 2. D o u b l e bounce scattering, mode l l ed b y a dihedral mechanism; 3. V o l u m e scattering from a c loud o f r andomly oriented dipoles , w h i c h is typ ica l f rom a canopy scatterer. A n i l lustrat ion o f these scattering mechanisms can be seen i n F igure 4 (Sect ion 2.2.2). In detection o f m a n made targets, the purpose o f us ing the Freeman-Durden decomposi t ion is to assess h o w effective it is i n ident i fy ing double bounce scattering originated from corners i n a phys ica l structure. F o r each p i x e l , Freeman-Durden decomposi t ion calculates the percentage o f each o f the three scattering mechanisms described above. T h e p i x e l is assigned to the image that corresponds to the dominant scattering mechanism. Before app ly ing this technique on real data w e have performed the f o l l o w i n g test: typ ica l dihedral and surface scattering matrices were s imulated and added i n a rectangle (no v o l u m e contr ibut ion was added). Af t e r that, Gauss ian noise was added to the w h o l e image, i nc lud ing the rectangle. F i n a l l y , F reeman-Durden decompos i t ion was applied. F igure 11 shows the results for four different percentage values o f dihedral and surface scattering. T h e percentage o f dihedral varies f rom 100% to 4 0 % , w h i l e the percentage o f surface varies f rom 0% to 60%. T h e results show that the a lgor i thm tends to classify the p ixe l s correctly, but the presence o f noise causes many p ixe l s to be classif ied as vo lume-dominated scatterers. 42 S = 0 D=1 V=0 S = 0 D=1 V = 0 S = 0 D= 1 V = 0 FD Target Map - Surface (S) FD Target Map - Dihedral (D) FD Target Map - Volume (V) S = 0.2 D = 0.8 V = 0 FD Target Map-Surface (S) • • • • : • • " > . • I • S = 0.2 D = 0.8 V = 0 FD Target Map-Dihedral (D) • v " .• •. ' » ; • > > • • V*.£>*.s S = 0.2 D = 0.8 V = 0 FD Target Map -Volume (V) S = 0.4 D = 0.6 V = 0 S = 0.4 D=0.6 V = 0 S = 0.4 D= 0.6 V = 0 FD Target Map-Surface (S) FD Target Map-Dihedral (D) FDTarget Map-Volume (V) S = 0.6 D = 0.4 V = 0 S = 0.6 D = 0.4 V = 0 S = 0.6 D = 0.4 V = 0 FDTarget Map-Surface (S) FD Target Map - Dihedral (D) FD Target Map - Volume (V) F i g u r e 11: F reeman-Durden decomposi t ion appl ied to s imulated images. 43 CHAPTER 5 - DATASETS This chapter describes the three datasets used i n the experiments i n this work . These are the Ot tawa dataset (2002), the Wes tham Island dataset (2004) and the Gage town dataset (2001). The descriptions w i l l also include the ground reference data col lected for each o f the datasets. The ground reference data are used to validate the experimental results. The photographs o f the specific targets selected for the experiments i n this w o r k w i l l be shown i n Chapter 6, a long w i t h the processing results. A l l the datasets are ful ly polar imetr ic and were acquired by the Conva i r -580 ( C V - 5 8 0 ) S A R airplane (Figure 12), w h i c h was o r ig ina l ly developed, owned and operated by the Canada Centre for Remote Sensing ( C C R S ) , and later transferred to Envi ronment Canada [15]. The data are processed by M D A ' s Geospat ia l Solut ions at the Canad ian Data Process ing Fac i l i t y ( C D P F ) at Gat ineau (Quebec). The C V - 5 8 0 specifications and data format are described at the end o f this chapter. Figure 12: The C o n v a i r 580 S A R airplane (Source: C C R S , 2002) 44 5.1 Datasets Used 5.1.1 The Ottawa Dataset The first dataset is acquired from an area to the northwest o f Ot tawa ( O N , Canada) , where a control led experiment was carried out by a group from the Canada Centre for Remote Sensing ( C C R S ) , to test their methodology to detect crashed airplanes [10]. In this thesis part o f the data is used. It covers a grass f ie ld where a single target was p laced a long wi th corner reflectors for data cal ibrat ion. In this work , these data are used both as a check for the methodology implementat ion and as a first test for the compar i son o f algori thms performances proposed here. The data was acquired by the C V - 5 8 0 radar system on June 25th, 2002. Figure 13: O v e r v i e w o f the Ottawa dataset and photograph o f the crashed airplane. Lef t : ( H H , red; V V , green; H V , blue). Orange arrow shows the orientation o f the airplane, and blue arrows indicate azimuth ( A ) , range (R) , and true north ( T N ) . R ight : Photograph o f the crashed airplane. (Source: L u k o w s k i et al. [10]). 45 5.1.1.1 G round Reference Data F r o m Ottawa F o r the experiments i n Ot tawa a prev ious ly crashed airplane was p laced on a grass field i n order to simulate a real crash o f a smal l aircraft on a l o w , un i form vegetation. The corner reflectors and A c t i v e Radar Cal ibrators were also p laced nearby and were used for data cal ibrat ion. 5.1.2 The Westham Island Dataset The second dataset covers Wes tham Island, a very flat i s land south o f V a n c o u v e r ( B C , Canada). The i s land 's l and cover is most ly composed o f agricul tural fields, w i t h a few houses and machinery spread along the roads. The is land 's borders used to be f looded by the tide. Th i s effect was avoided i n the crop fields by the construction, years ago, o f a d i tch surrounding the is land. O n the north and northwest sides there are forested areas and swamps. M a n y migratory birds populate this area for part o f the year. The area is located between 123° 0 7 ' 4 7 " W and 123° 11 ' 3 6 " W i n the E - W direct ion and between 4 9 ° 0 4 ' 1 3 " N and 4 9 ° 0 6 ' 4 0 " N i n the N - S direct ion. 46 Figure 14: O v e r v i e w o f Wes tham Island. H H - R e d , H V - Green and V V - B l u e . T h i s data was col lected and processed for the U B C Radar Remote Sens ing G r o u p ( R R S G ) as part o f a larger acquis i t ion campaign on September 30, 2004. T h e data swath extends up to the c i ty o f N o r t h Vancouve r . 5.1.2.1 Ground Reference Data from Westham Island T h e ground data for the Wes tham Island dataset comprises ground photographs taken b y members o f the R R S G ( U B C ) dur ing two f ie ldwork trips to the is land. The first trip took place on September 2 8 t h 2004 - two days before the acquis i t ion date - w i t h the purpose o f m a k i n g an overa l l examinat ion o f the terrain and its potential targets. A c c e s s i b i l i t y was checked and some photographs o f f ixed targets were taken. T h e second f ie ld trip took place on the acquis i t ion date (September 30 t h ) , and other targets were photographed. T h e photos were taken over approximate ly 2 hours. The C V - 5 8 0 airplane was f l y i n g over the area (between 11:00am and 1:00pm, loca l t ime). Parked machinery , wreckage, bu i ld ings , boats and other relevant m a n made targets were inc luded i n the photographs and G P S readings were taken. D u e to the almost total absence o f forests i n the is land, few targets are located w i t h i n o f h i g h vegetation clutters. H i g h resolut ion opt ical data avai lable onl ine were also used i n this w o r k as complementary reference informat ion o n the ground features. These data are a mosa ic o f images acquired between M a r c h 10 t h , 2004 and September 19 t h , 2004, and were helpful i n ident i fy ing non-movab le targets seen i n the radar data. 47 5.1.3 T h e G a g e t o w n Datase t The third dataset was acquired on September 11 , 2001 , over Gage town ( N B , Canada). It was acquired for Radarsat International (now M D A ' s Geospat ia l Services) , w h o p rov ided the U B C R R S G w i t h a copy for research purposes. M o s t o f the area includes the Canad ian Forces Base ( C F B ) Gage town, w h i c h has l o w bui ld ings , mi l i t a ry vehicles , t raining ranges and a residential area. The Freder ic ton airport is on the N W corner o f the imaged area, the Saint John river is on the north edge and forests, ponds and l o w vegetation fields cover the rest o f the image. The area is approximately 16 k m ( W / E direction) X 8 k m ( N / S direction) and is located between 6 6 ° 2 0 ' 1 7 " W and 6 6 ° 34 ' 0 0 " W i n the E - W direct ion and between 4 5 ° 2 7 ' 2 0 " N and 4 5 ° 5 3 ' 0 0 " N i n the N - S direct ion. F i g u r e 15: O v e r v i e w o f Gage town data. 48 5.1.3.1 Ground Reference Data from Gagetown T w o field campaigns were performed to col lect ground truth data at selected sites i n the Gage town area by A e r d e Envi ronmenta l Research Inc. The first one took place on September 10, 2001 , w h i c h was one day before the acquis i t ion date. Th i s fieldwork generated 16 ground photos o f different targets l y i n g both inside and outside C F B Gage town. The S A R data were processed i n 2002, and v i sua l analysis showed the need for more ground truth. Based on the radar responses seen i n the image and auxi l ia ry informat ion (both from a 1:50,000 map) and a h igh resolut ion Ikonos image, specific targets were selected for a second fieldwork. M o s t o f these targets were man made targets such as pub l ic monuments , fixed art i l lery pieces for d isplay and bui ld ings . Th i s fieldwork took place on N o v e m b e r 27, 2002, and around 70 photographs were taken o f the targets - except those that presented access problems. Figure 16: Ikonos image acquired on Oc t 4 , 2001 over Gage town. A n Ikonos image was acquired over part o f the area on Oc t 4 t h , 2001 (23 days after the C V -580 acquis i t ion date) and was used i n this w o r k to identify larger targets o f interest i n the 49 area. A R A D A R S A T - 1 F ine beam mode (10 meters resolution) image was acquired on the same day as the C V - 5 8 0 S A R data were acquired and is part o f the reference data. Figure 17: Ikonos, C V - 5 8 0 data and photograph o f Gage town data. Top left: De ta i l o f Ikonos image (Oct. 4 t h , 2001) used as auxi l ia ry ground reference data showing helicopters i n the C F B Gage town. Top right: En la rged detail o f V V channel o f C V - 5 8 0 data (Sept. 1 1 t h , 2001) o f the same area. Bottom: Photograph showing the helicopters. N o t i c e that, due to the difference i n acquis i t ion dates, the helicopters are not i n the same posit ions i n the images. 5.2 CV-580 SAR System Specifications and Data 5.2.1 The CV-580 SAR System A l l the data used i n this w o r k were acquired by the C V - 5 8 0 S A R system. Th i s airborne system was developed by the Canada Centre for Remote Sensing ( C C R S ) o f Na tura l Resources Canada, and i n 1996 transferred to Envi ronment Canada ( E C ) . The system can 50 produce both interferometric and polar imetr ic data, i n both X - b a n d (not polar imetr ic) and C-band , and the data used i n this w o r k are fu l ly polar imetr ic C-band data. - . JWU . ' l l l -Ui 'M / S I • F i g u r e 18: T a i l section o f C o n v a i r - 5 8 0 showing S A R radomes. (Source: C C R S , 2002). ESP Convair 580 \ InSAR Main Antenna Real-time Display Antenna Radome Radome Station F i g u r e 19: C o n v a i r System configurat ion (Source: C C R S , 2002). 51 5.2.2 Data Format T h e raw data received dur ing acquis i t ion were transcribed to a computer compat ible format. Af te r a large set o f processing stages, the four channels o f complex polar imetr ic data are generated. F o l l o w i n g the focusing and cal ibrat ion o f the data, two m a i n classes o f files are generated: (1) T h e P o l G A S P (output f rom the Pola r imet r ic Genera l i zed A i r b o r n e S A R Processor); and (2) The S I R - C data format [15]. T h e S I R - C format is p rov ided w i t h equal range and az imuth resolutions, and is used to generate geocoded imagery i n a compressed format. T h e P o l G A S P format, w h i c h is used i n this work , carries S ingle L o o k C o m p l e x ( S L C ) , f loat ing point data, and its extension is . img . E a c h file carries one o f the four polar izat ions ( H H , H V , V H and W ) and is accompanied b y a header A S C I I file (.hdr extension). T h e addi t ion o f a master header f i le that contains the general informat ion about the four po lar iza t ion headers makes a total o f nine files per data. T h e b inary data were wri t ten on a S G I machine, i n b ig-endian format (i.e., byte swapped compared to a P C ) . T h i s should be considered b y the routine used to read the data into M a t l a b . P O L S A R single look complex ( S L C ) data have different resolutions i n az imuth and range. T h i s is because i n radar i m a g i n g these resolutions depend on different parameters. T h e range resolut ion depends o n the chirp bandwidth , and the az imuth resolut ion depends o n system parameters such as the antenna length and the fraction o f the D o p p l e r bandwid th processed. 52 CHAPTER 6 - DATA PROCESSING RESULTS AND ANALYSIS 6.1 Selected Targets T h e targets for this w o r k were selected from both the Wes tham Island and Gage town datasets, and one target i n the Ot tawa dataset where a crashed airplane was intent ional ly p laced i n the scene (see Chapter 5). The f o l l o w i n g two cri teria were kept i n m i n d dur ing the select ion o f these data: targets should be surrounded b y different backgrounds (bare so i l , l o w , m e d i u m and h igh vegetation) and should be t yp i ca l ly m a n made ones, as defined i n Sec t ion 3.1. T h i s task was completed b y ana lyz ing both the photographs taken i n the f i e ldwork trips and the Ikonos image. Tab le 4 describes the targets, the clutter type and their locat ion. Coordinates are g iven i n W G S - 8 4 datum. In F igure 20, photographs or opt ical satellite images o f the targets are presented. Target Target description Clutter type Coordinates (WGS 84 Datum) and Dataset 1 T w o cy l ind r i ca l steel containers Forest + wheat f ie ld 49° 05'49"N/123° 09' 51"W Westham Island 2 A g r i c u l t u r a l p l o w i n g machine L o w grass + bare so i l 49° 05'02"N/123° 08'20"W Westham Island 4 Cart beside a 10 metre h igh tree L o w grass 49° 05' 14"N/123° 08' 25"W Westham Island 5 T w o steel cyl inders (water tanks) on stands C o r n f ie ld 49° 05'42"N/ 123° 08'45"W Westham Island 7 Large house surrounded b y trees B r o c c o l i f ie ld 49° 05'33"N/123° 09'59"W Westham Island 12 A r t i l l e r y pieces L o w grass 45° 50' 25"N / 66° 27' 59"W Gagetown 14 House Forest 45°52'37"N/66°32' 17"W Gagetown 21 House Forest 45°51'59"N/66°28' 13"W Gagetown 20 Crashed airplane M e d i u m grass Ottawa Table 4: Target and clutter descriptions and locations. 53 Target 5 Target 7 Target 12 Target 14 Target 21 Target 20 F i g u r e 20: Photos and images o f the 9 targets selected for this work . 54 6.2 Results The methodologies described i n Chapter 4 have been appl ied to images containing the targets and an area o f clutter. The images are a l l sub-images f rom the or ig ina l datasets and were cropped i n the largest possible area o f homogeneous clutter a long w i t h the target. Roads , fences, poles and other m a n made features were kept out o f the scenes i n order to avo id scene complex i ty and keep o n l y one target i n each image. In our study w e are m a i n l y concerned w i t h the detection o f the targets and decrease o f the number o f false alarms. In order to achieve that, var ious thresholds were appl ied for each a lgor i thm. T h e detection results are obtained b y app ly ing the thresholds that y i e l d the lowest false a larm rates w h i l e s t i l l keep ing at least three connected p ixe l s o f the target. These thresholds are considered the op t imal thresholds for that part icular target and clutter. Detect ions conta ining one or two p ixe ls cannot be considered as targets or as false alarms. T h e v a l i d targets are h ighl ighted w i t h i n ell ipses i n the var ious detection results shown i n the figures i n this chapter. F o r each target, a descr ipt ion o f both target and clutter is p rov ided , fo l lowed b y graphical and numer ica l results. F o r clari ty, fu l l results for Targets 7, 14 and 21 are shown i n this chapter. F i n a l detection results for Targets 1, 2, 4, 5, 12 and 20 are also shown i n this chapter, and plots showing intermediate processing results for these targets are shown i n A p p e n d i x A . T h e methodologies are appl ied as fo l lows : Methodology 1: T h e Polar imet r ic W h i t e n i n g Fi l te r ( P W F ) and the E v e n B o u n c e ( E B ) a lgor i thm were each appl ied to the or ig ina l image, generating a P W F processed image and an E v e n B o u n c e processed image. Different Constant False A l a r m Rate ( C F A R ) values (represented b y the constant K , as expla ined i n Sect ion 4.1) were appl ied to the P W F and to the E v e n B o u n c e 55 images. The optimal thresholds are determined and used to generate both the P W F target map and the Even Bounce target map. A mask was created for each of these two target maps and Cameron decomposition was applied to both masks. After that, only dihedrals and narrow dihedrals were selected. The intersection between the results above resulted in a target map to which the morphological processing was applied, as explained in Section 4.1. The intersection is an " A N D " operator, and selects only the pixels that are detected in both images. The resulting target map was the detection map, from which the false alarm rate was calculated. Methodology 2: The Coherence Test procedure from the S S C M algorithm was applied to the original image. Various degree of coherency and Rician threshold values were applied. The Coherence Test target map was generated by applying the optimal degree of coherence and Rician thresholds. Cameron decomposition was applied to this target map and only the pixels classified as dihedrals and narrow dihedrals were kept. Morphological processing was applied to the resulting map in order to generate the detection map, from which the false alarm rate was calculated Methodology 3: The Freeman-Durden decomposition was applied to the original image. Various dihedral percentage threshold values were applied to the dihedral scattering image, generating a target map. Morphological processing was applied to the resulting map in order to generate the detection map, from which the false alarm rate was calculated. Methodology 4: The detection maps resulting from Methodologies 1 and 2 were cross-checked (only those targets appearing in both detection maps were retained). If the number of connected pixels from the target in the resulting image was larger than three, the target was considered to be 56 retained. Otherwise, Methodologies 1 and 2 were applied again with different thresholds until the target was retained. When the target was finally retained, morphological processing was applied and the final detection map was generated. After that, the false alarm rate was calculated. 57 6.2.1 Results for Target 7 (House Among Trees) 6.2.1.1 Target Descript ion Target 7 is a house surrounded b y trees i n its immediate v i c in i ty . The clutter is a b r o c c o l i field. Figure 21: Target 7 - R G B composi te and photograph. Lef t : C o l o u r composi te image ( H H - R e d , H V - Green , V V - B l u e ) showing Target 7 ( in the ell ipse) and the surrounding ground. R ight : Target 7 photograph. Analys is : M e t h o d o l o g y 1 generated one false a larm close to the house (see F igure 22). Th i s can be due to the h igh scattering from the large tree or, more l i k e l y , to a vehic le parked close to this tree. In M e t h o d o l o g y 2 the Coherence Test detected many p ixe l s o n the house (see Figure 24) as w e l l as i n the b rocco l i field. Results f rom both Methodolog ies 1 and 2 show that, al though the P W F , E v e n B o u n c e and Coherence Test c o u l d detect most p ixe ls o f the house, C a m e r o n Decompos i t i on ident i f ied very few o f these as dihedrals. Th is m a y be due to the r o o f causing a strong surface scattering, w h i c h C a m e r o n w o u l d classify as tr ihedral. A l s o , these three algori thms detected part o f the trees that surround the house as targets, w h i l e C a m e r o n decompos i t ion d i d not. Th is kept the trees f rom being confused w i t h the house, al though they are very close and present a h igh backscatter. 58 Methodology 3 detected the house with a considerably lower false alarm rate than it did in the other targets. Freeman-Durden decomposition calculated a high percentage of dihedral in the house scattering. This allowed us to use for this image the highest threshold value for a target in this work (dihedral percentage > 94%). This caused the drop in the false alarm rate, although it had still the highest false alarm rate among the four methodologies. Methodology 4 detected the target without incorrect detections. Methodology 1 Methodology 2 Methodology 3 Methodology 4 False Alarm count 1 7 10 0 False Alarm Rate 20 146 209 0 Table 5: Target 7 - False Alarm count and False Alarm Rate (false alarms/km 2) 59 6.2.1.2 Processing Results for Target 7 Detection Map - Methodology 1 Detection Map - Methodology 2 50 100 150 200 250 300 350 50 100 150 200 250 Detection Map - Methodology 3 Detection Map - Methodology 4 50 100 150 200 250 300 350 50 100 150 200 250 Figure 22: Target 7 - Final detection maps for Methodologies 1 to 4. 60 PWF Target Map (K = 2) i i 50 100 150 200 250 300 350 50 100 150 200 250 300 350 Cameron, PWF and Even Bounce 50 100 150 200 250 300 350 Figure 23: Target 7 - Plots f rom M e t h o d o l o g y 1. U p p e r left: P W F target map. U p p e r right: E v e n B o u n c e target map. B o t t o m : combined result o f C a m e r o n (dihedrals plus narrow dihedrals), P W F and E v e n Bounce . K is the C F A R constant. 61 Coherence Test Target Map Cameron and Coherence Test Map J .1 'i 50 100 150 200 250 300 350 100 150 200 250 300 350 (a) (b) Figure 24: Target 7 - Plots f rom M e t h o d o l o g y 2. (a) Coherence Test target map. (b) C o m b i n e d result o f C a m e r o n (dihedrals plus narrow dihedrals) and Coherence Test. Target Map Cameron, PWF, EB and Coherency Test Target Map 50 100 150 200 250 300 350 (a) (b) Figure 25: Target 7 - Plots f rom Methodolog ies 3 and 4. (a) F reeman-Durden target map showing p ixe l s that present a percentage o f dihedral scattering above the chosen threshold, (b) C o m b i n e d results o f C a m e r o n (dihedrals plus narrow dihedrals) , P W F , E v e n B o u n c e and Coherence Test. 62 6.2.2 Results for Target 14 (House in Forest) 6.2.2.1 Target Description Target 14 is a house located i n a remote area and is surrounded by forest. Figure 26: Target 14 - R G B composi te and photograph. Left : C o l o u r composite image ( H H - R e d , H V - Green , V V - B l u e ) showing Target 2 ( in the ell ipse) and the surrounding ground. The white diagonal feature i n the image is a d i tch separating two crop fields. R igh t : Target 2. Analysis: Methodo log ies 1 and 2 presented very few false alarms (one each - see F igure 27). O n l y two out o f the three detected target p ixels were the same i n these two detection maps. M e t h o d o l o g y 4 d i d not detect the target w i t h these thresholds, as the m i n i m u m size a l l owed is three p ixels . Therefore, M e t h o d o l o g y 4 was appl ied after re-applying Methodo log ies 1 and 2 w i t h lower thresholds. The false a larm rate i n M e t h o d o l o g y 4 is lower than i n Methodolog ies 1 and 2. The false a larm detected was the on ly one present i n a l l results f rom this methodology. T h i s 63 detection happened in area of high forest, and no man made feature could be seen in the Ikonos image. The percentage of dihedrals calculated by Freeman-Durden for this image was comparatively low all over the scene, but the target was not detected. For illustration purposes the detection result of Methodology 3 with a 1% threshold is shown in the processing results. Methodology 1 Methodology 2 Methodology 3 Methodology 4 False Alarm count 2 2 56 1 False Alarm Rate 144 144 4,042 72 Table 6: Target 14 - False Alarm count and False Alarm Rate (false alarms/km 2). 64 6.2.2.2 Processing Results for Target 14 Detection Map - Methodology 1 Detection Map - Methodology 2 Detection Map - Methodology 3 Detection Map - Methodology 4 50 100 150 200 Figure 27: Target 14 - F i n a l detection maps for Methodo log ies 1 to 4. 65 PWF Target Map (K = 1) Even Bounce Target Map (K = 6 ) 10 I * Jl I " ' H r \f... . . . ii 1 • \ \ 1 ll 1 1 \ I -1 1 100 150 200 6 0 Cameron, PWF and Even Bounce 5 1 , 1 » 1 1 10 1 15 1 1 1 l 20 1 1 1 \ \ 25 • -30 • ' 1 -35 150 200 50 100 150 200 Figure 28: Target 14 - Plots from Methodology 1. Upper left: P W F target map. Upper right: Even Bounce target map. Bottom: combined result of Cameron (dihedrals plus narrow dihedrals), P W F and Even Bounce. K is the C F A R constant. 66 Coherence Test Target Map Cameron and Coherence Test Map 150 200 50 100 150 200 (a) (b) Figure 29: Target 14 - Plots f rom M e t h o d o l o g y 2. (a) Coherence Test target map. (b) C o m b i n e d result o f Cameron (dihedrals plus narrow dihedrals) and Coherence Test. Target Map 35 r I. UH II i I i , i I I i i 'i: „ i i I. f , . lf[ 1 ' \ : " 1i . ' , 1 1 'I . . ' . 1 " i i ,ii I . \ L 1 Cameron, P W F , E B and Coherency Test Target Map 10 20 1 , I 1 \ 1 .1 1 1 1 \ 1 50 100 150 200 50 100 150 (a) (b) Figure 30: Target 14 - Plots f rom Methodolog ies 3 and 4. (a) Freeman-Durden target map showing p ixe l s that present a percentage o f dihedral scattering above the chosen threshold, (b) C o m b i n e d results o f C a m e r o n (dihedrals plus narrow dihedrals) , P W F , E v e n B o u n c e and Coherence Test. 67 6.2.3 Results for Target 21 (House in Forest) 6.2.3.1 Target Descript ion Target 21 is a house by a forest near Gage town. B o t h the house and the forest surrounding it are bounded by a road, w h i c h is not inc luded i n the image. Figure 31: Target 21 - R G B composi te and photograph. Left : C o l o u r composi te image ( H H - R e d , H V - Green , V V - B lue ) showing Target 7 ( in the ell ipse) and the surrounding ground. R ight : Target 7. Analys is : B o t h Methodolog ies 1 and 2 showed 2 false alarms each i n the forest area (see F igure 32). In M e t h o d o l o g y 4 it was possible to detect the target wh i l e r emov ing these false alarms without hav ing to change the thresholds. The forest's h igh backscatter might be what degraded the P W F performance (see F igure 33). Coherence Test and E v e n Bounce A n a l y s i s had considerably less false detections than the P W F . 68 Methodology 3 generated a high number of false alarms, as the threshold had to be set in a low value (13%) in order to detect the house. If compared to the results of Target 7, the house in Target 21 was much less detectable by this methodology than the house in Target 7. This is probably due to the house's orientation with respect to the radar look direction: the house in Target 7 has walls perpendicular to the radar line of sight, while in Target 21 the house walls make a 45 degrees angle with the line of sight. Methodology 1 Methodology 2 Methodology 3 Methodology 4 False Alarm count 2 2 74 0 False Alarm Rate 44 44 1,652 0 9 Table 7: Target 21 - False Alarm count and False Alarm Rate (false alarms/km ) 69 6.2.3.2 Processing Results for Target 21 Detection Map - Methodology 1 Detection Map - Methodology 2 50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400 Detection Map - Methodology 4 Detection Map - Methodology 2 50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400 Figure 32: Target 21 - Final detection maps for Methodologies 1 to 4. 70 P W F Target Map (K = 2) Even Bounce Target Map (K = 7) 2 0 I l l I ' 1 i II 10 1 1 k ' ' ' . 1 • ( 1 1 1 1 1 1 I t 1 1 1 i i '-i 20 30 • 1 1 I , 1 • 1" • i I . -40 50 1 -• ' ' ' • , • 1 1 . . . i. ' ' ' h * ' " I ' 'i 1 60 70 'i Ii 1 -50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400 Cameron, P W F and Even Bounce 50 100 150 200 250 300 350 400 F i g u r e 33: Target 21 - Plots from Methodology 1. Upper left: PWF target map. Upper right: Even Bounce target map. Bottom: combined result of Cameron (dihedrals plus narrow dihedrals), PWF and Even Bounce. K is the CFAR constant. 71 Coherence Test Target Map Cameron and Coherence Test Map 50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400 (a) (b) Figure 34: Target 21 - Plots f rom M e t h o d o l o g y 2. (a) Coherence Test target map. (b) C o m b i n e d result o f C a m e r o n (dihedrals plus narrow dihedrals) and Coherence Test. 30 I II II Target Map T7 i III • i i i 1 • • I . I I I i i i I I I u, , 11' 50 100 150 200 250 300 350 400 Cameron, P W F , E B and Coherency Test Target Map 50 100 150 200 250 300 350 400 (a) (b) Figure 35: Target 21 - Plots f rom Methodolog ies 3 and 4. (a) Freeman-Durden target map showing p ixe l s that present a percentage o f dihedral scattering above the chosen threshold, (b) C o m b i n e d results o f C a m e r o n (dihedrals plus narrow dihedrals) , P W F , E v e n B o u n c e and Coherence Test. 72 6.2.4 Results for Target 1 (Two Ver t ica l Cyl inders) 6.2.4.1 Target Descript ion T w o ver t ical steel cyl inders about 3 meters h igh and w i t h a 4 meters diameter located on the boundary between a forested area and a wheat f ie ld i n Wes tham Island. Figure 36: Target 1 - R G B composi te and photograph. Lef t : C o l o u r composi te image ( H H - R e d , H V - Green, W - B l u e ) showing Target 1 ( in the ell ipse) and the approximate boundaries o f the processed sub-image ( in the rectangle). R ight : Photograph o f Target 1. Analys is : A l l the algori thms present a few more detections i n the forested area (left o f the target), than i n the wheat field (Figures 48, 49 and 50 i n Sect ion A . l ) . Methodolog ies 1, 2 and 4 detect the target w i t h no false alarms, and the detection algori thms show few detections other than the target. Th i s may be due to the fact that the target is large and has a strong response, a l l o w i n g for h igh threshold values and therefore l o w false a larm rate. Th i s is not on ly for its size but also because ver t ical cyl inders behave l ike a ver t ica l pole or tree trunk, presenting a strong dihedral response. M e t h o d o l o g y 3 has a higher false a larm rate than other methodologies. 73 6.2.4.2 Processing Results for Target 1 Detection Map - Methodology 1 Detection Map - Methodology 2 Figure 37: Target 1 - Final detection maps for Methodologies 1 to 4. Methodology 1 Methodology 2 Methodology 3 Methodology 4 False Alarm count 0 0 203 0 False Alarm 0 0 5,138 0 Rate 2 Table 8: Target 1 - False Alarm count and False Alarm Rate (false alarms/km ) 74 6.2.5 Results for Target 2 (Plow in Grass) 6.2.5.1 Target Descript ion Target 2 is an agricultural plowing machine. It is located close to a ditch that separates two crop fields. The ground in the lower left field is mostly bare soil with sparse, low grass. The upper right field is covered in higher grass. Figure 38: Target 2 - R G B composite and photograph. Left: Colour composite image (HH - Red, H V - Green, W - Blue) showing Target 2 (in the ellipse) and the surrounding ground. Right: Target 2. Analys is : A l l the algorithms present more detections on the upper right field, where the grass causes higher, more diffuse backscatter than the bare soil. The Coherence Test detects a feature that is larger than the target in the lower field, but this feature is erased when the Coherence Test target map is combined with Cameron Decomposition. Even Bounce analysis shows less false alarms than both P W F and Coherence Test. A l l algorithms have detections in the ditch, where the slope facing the radar can present occasional single and double bounce behaviour. But these detections, as the others left after the combined results of Methodologies 1 and 2 are generated, are random pixels erased by morphological processing. Methodology 3 has a very high false alarm rate. Methodology 4 cannot be assessed as both Methodologies 1 and 2 presented zero false alarms. 75 6.2.5.2 Processing Results for Target 2 Detection Map - Methodology 1 Detection Map - Methodology 2 Figure 39: Target 2 - F i n a l detection maps for Methodo log ies 1 to 4. M e t h o d o l o g y 1 M e t h o d o l o g y 2 M e t h o d o l o g y 3 M e t h o d o l o g y 4 False Alarm 0 0 152 0 count False Alarm 0 0 5,875 0 Rate Table 9: Target 2 - False A l a r m count and False A l a r m Rate (false a la rms/km ) 76 6.2.6 Resu l t s fo r T a r g e t 4 ( L a r g e F a r m C a r t ) 6.2.6.1 T a r g e t D e s c r i p t i o n Target 4 is a large farm cart (blue cart i n the photo) i n Wes tham Island. Due to target locat ion i n the scene other m a n made targets had to be inc luded, m a k i n g this a complex scene. A single tree is located to the right hand side o f the target. There are two clutter fields i n the scene: l o w grass on the lower left part and m e d i u m grass on the upper right. F i g u r e 4 0 : Target 4 - R G B composi te and photograph. Left : C o l o u r composi te image ( H H - R e d , H V - Green, V V - B l u e ) showing Target 4 ( in the ell ipse) and the surrounding ground. Other m a n made features are present i n this image. Right : Target 4 (blue cart), a single tree and other features: barn and machinery. A n a l y s i s : The taller grass f ie ld presented a considerable higher number o f detections than the l o w grass field i n a l l algori thms (Figures 54, 55 and 56), showing that the higher the vegetation the more l i k e l y it is to produce double bounce behaviour. M o s t o f these detections were erased for not be ing classif ied as dihedrals by Cameron , and the remain ing ones were o f smal l size and were erased by the morpho log ica l processing. N o false alarms were present i n Methodolog ies 1, 2 and 4. Other man made features were detected by a l l algori thms, but were not considered i n the false a larm rate count. It is noteworthy that the single tree had p ixe ls detected by E v e n Bounce , Coherence Test and C a m e r o n (see Sect ion A . 4 ) . It d i d not appear i n the final detection map o f M e t h o d o l o g y 1 because it was not detected by P W F and because on ly two p ixe ls were detected. M e t h o d o l o g y 3 generated a m u c h higher F A R . 77 6.2.6.2 Processing Results for Target 4 Detection Map - Methodology 1 Detection Map - Methodology 2 25 - 5 -10 1 h I" ? 1 i 1 • 15 0° 20 25 20 40 60 80 100 120 140 160 180 200 220 20 40 60 80 100 120 140 160 180 200 220 Detection Map - Methodology 3 Detection Map - Methodology 4 20 40 60 80 100 120 140 160 180 200 220 20 40 60 80 100 120 140 160 180 200 220 Figure 41: Target 4 - F i n a l detection maps for Methodo log ies 1 to 4. P i x e l s i n the rectangle are other m a n made targets i n this complex scene. T h e c i rc le indicates the loca t ion o f a tree. M e t h o d o l o g y 1 M e t h o d o l o g y 2 M e t h o d o l o g y 3 M e t h o d o l o g y 4 False Alarm 0 0 16 0 count False Alarm 0 0 1,749 0 Rate Table 10: Target 4 - False A l a r m count and False A l a r m Rate (false a l a rms /km 2 ) 78 6.2.7 Results for Target 5 (Hor izontal Cyl inders) 6.2.7.1 Target Descript ion Target 5 is a set o f two metal water tanks on stands w i t h a hor izonta l cy l i nd r i ca l shape located i n Wes tham Island. The tanks are 2-3 meters i n diameter. The target is a few meters above the ground. There ' s some bare so i l to the left o f the target, but most o f the clutter i n the scene is composed o f corn fields. Figure 42: Target 5 - R G B composi te and photograph. Left : C o l o u r composi te image ( H H - R e d , H V - Green , W - B l u e ) showing Target 5 ( in the ell ipse) and the surrounding ground. The clutter is a corn f ie ld. Right : Photograph o f Target 5. Analys is : M o s t o f the p ixe ls cover ing this target were detected by P W F and E v e n Bounce . Thresholds were lowered to cover more target p ixe ls , but on ly one o f the target p ixe ls occupied was classif ied by C a m e r o n Decompos i t i on as a dihedral . The c y l i n d r i c a l shape o f the target is the l i ke ly cause. In M e t h o d o l o g y 2, the Coherence Test detected 4 p ixels o f the target, but also here combined result w i t h C a m e r o n Decompos i t i on d idn ' t a l l o w target detection. The target was detected i n M e t h o d o l o g y 3 at a very h igh false a larm rate. 79 6.2.7.2 Processing Results for Target 5 23 a «D 80 «U 120 1(0 160 180 Detection Map - Methodology 3 20 40 60 80 100 120 140 160 180 Detection Map - Methodology 2 2D 25' 20 40 60 80 100 120 140 160 180 Ostcooi nap- macdottg/ 4 i r r-—-i——i r -—r-0 I I - I t_ 20 to m m too 120 ic m m Figure 4 3 : Target 5 - F i n a l detection maps for Methodo log ies 1 to 4. M e t h o d o l o g y 1 M e t h o d o l o g y 2 M e t h o d o l o g y 3 M e t h o d o l o g y 4 False Alarm count - - 54 -False Alarm Rate - - 6,724 -9 Table 11: Target 5 - False A l a r m count and False A l a r m Rate (false a la rms /km ) 80 6.2.8 Results for Target 12 (Art i l lery Pieces in Gagetown) 6.2.8.1 Target Descript ion Target 12 is a set o f four art i l lery pieces on permanent display at the Gage town mi l i t a ry base. There are a few trees close to one o f the pieces, and the remain ing clutter is short grass. Figure 44: Target 12 - R G B composi te and photograph. Left: C o l o u r composi te image ( H H - R e d , H V - Green, W - B l u e ) h ighl igh t ing the pieces that Target 12 ( in the ell ipse) and the surrounding ground. Righ t : Target 7. Analys is : M e t h o d o l o g y 1 performs w e l l , detecting the targets w i t h no false alarms. One o f the false alarms come from the trees, w h i c h was also detected by the P W F . M e t h o d o l o g y 2 results in three false alarms, w i t h the Coherence Test detecting the trees and part o f the grass. It was not possible to detect the targets w i th M e t h o d o l o g y 3, because there were less than 3 connected target samples for each o f the targets. The target map f rom M e t h o d o l o g y 3 shows that one target was detected wi th a threshold value as l o w as 1% (Figure 62). 81 6.2.8.2 Processing Results for Target 12 Detect ion M a p - Methodology 1 Detect ion Map - Methodology 2 80 100 120 140 160 40 60 80 100 120 140 160 Detect ion Map - Methodology 3 Detect ion M a p - Methodology 4 20 40 60 80 100 120 140 160 40 60 80 100 120 140 160 Figure 45: Target 12 - F i n a l detection maps for Methodo log ies 1 to 4. M e t h o d o l o g y 1 M e t h o d o l o g y 2 M e t h o d o l o g y 3 M e t h o d o l o g y 4 False Alarm count 0 3 13 0 False Alarm Rate 0 751 3,225 0 2 Table 12: Target 12 - False A l a r m count and False A l a r m Rate (false a la rms/km ) 82 6.2.9 Results for Target 20 (Crashed Ai rp lane) 6.2.9.1 Target Descript ion Target 20 is a crashed plane (same scene used by L u k o w s k i et al. [8,9,10]). 200 4O0 600 800 1000 1200 Figure 46: Target 20 - R G B composi te and photograph. Left : P W F image showing Target 20, corner reflectors and the surrounding ground. Right : Target 20 . Analys is : The Coherence Test was more sensitive to the brightness o f the grass then the P W F and the E v e n Bounce . C a m e r o n Decompos i t i on el iminated the false detections on a l l three algori thms (see Figures 63 , 64 and 65). The corner detectors are detected w e l l by the P W F due to their h igh target-to-clutter ratio, but have very few samples in E v e n Bounce A n a l y s i s and Coherence Test. A s these devices are trihedrals, C a m e r o n Decompos i t i on d i d not detect them and they were not inc luded i n the detection maps. Methodolog ies 1, 2 and 4 generated no false alarms, w h i l e M e t h o d o l o g y 3 presented a very h igh false a larm rate. T h i s may be due to the fact that the target-to-clutter ratio does not affect Freeman-Durden classif icat ion. 83 6.2.9.2 Processing Results for Target 20 M e t h o d o l o g y 1 M e t h o d o l o g y 2 M e t h o d o l o g y 3 M e t h o d o l o g y 4 False Alarm count 0 0 1,122 0 False Alarm Rate 0 0 5,.933 0 Table 13: Target 20 - False A l a r m count and False A l a r m Rate (false a la rms /km 2 ) 84 6.3 Algorithm Threshold Values Tables 14 and 15 present the threshold values used i n the algori thms. T h e thresholds are: K , degree o f coherence, R i c i a n threshold and percentage o f dihedral scattering contr ibut ion i n F r ee man -D u rd en Decompos i t ion . CFAR PWF (K) CFAR Even Bounce (K) Degree of Coherence Rician threshold (dB) Percentage of Dihedral scattering contribution (Freeman -Durden) Target 1 3 8 0.6 10 4 0 % Target 2 2 6 0.55 4 5 0 % Target 4 2 2 0.5 8 7 6 % Target 5 1 4 0.5 8 3 8 % Target 7 2 4 0.5 7 9 4 % Target 12 2 6 0.38 11 6 9 % Target 14 1 6 0.65 8 1% Target 21 2 7 0.6 11 1 3 % Target 20* 3 5 0.9 8 6 3 % * Data used by Lukowski et al. [10] Table 14: O p t i m a l thresholds when Methodolog ies 1, 2 and 3 are appl ied. 85 CFAR PWF (K) CFAR Even Bounce (K) Degree of Coherence Rician threshold (dB) Target 1 3 8 0.6 10 Target 2 2 6 0.55 4 Target 4 2 2 0.5 8 Target 5 1 4 0.5 8 Target 7 2 4 0.5 7 Target 12 2 6 0.38 11 Target 14 1 6 0.6 8 Target 21 2 7 0.6 11 Target 20* 3 5 0.9 8 * Da ta used by Lukowski etal.[10] Table 15: O p t i m a l thresholds when M e t h o d o l o g y 4 is appl ied. 86 6.4 Genera l Analysis T h e results obtained b y the appl icat ion o f the methodologies were compared and l ed to the conclusions that fo l l ow. M e t h o d o l o g y 3 is not suitable for this appl icat ion. M a n y threshold values were tested for each target and, i n order to retain the target, a large number o f false alarms had to be kept as w e l l . T h e m a n made targets do not seem to be perceived b y Freeman-Durden D e c o m p o s i t i o n as stronger double bounce scatterers than the surrounding clutter. A l s o , the thresholds used i n M e t h o d o l o g y 3 are not consistent f rom scene to scene. T h e values range f rom 13 to 94 (the 1% threshold was not counted here, as the target was not detected). T h i s was the m i n i m u m percentage necessary to have three connected samples o f the target detected. G i v e n the poor results o f M e t h o d o l o g y 3, w e proceeded to compare the other three methodologies . Tab le 16 shows the total number o f false alarms and the total false a larm rates per methodology. In order to keep the results f rom be ing affected b y complex scenes, w e tried, as m u c h as possible, to select homogeneous fields close to w h i c h there was ground-truthing information. T h i s led to a l imi ta t ion on image size, and the average image size was 150 x 200 m . F o r this reason, i f any false alarms were present the false a larm rates w o u l d result i n a h igh number. 87 M e t h o d o l o g y 1 M e t h o d o l o g y 2 M e t h o d o l o g y 4 False Alarm count (Low Vegetation) 0 3 0 False Alarm count (Medium Vegetation) 1 7 0 False Alarm count (High Vegetation) 4 4 1 Total False Alarm count 5 14 1 Total False Alarm Rate 210 1087 72 Table 16: False a larm count for l o w , m e d i u m and h i g h vegetation types, total false a larm count and total false a larm rate (false alarms/ k m 2 ) for each methodology. Some targets do not have enough samples classif ied b y C a m e r o n Decompos i t i on as dihedrals and cannot be detected b y Methodo log ies 1 and 2. In this work , this is the case o f Target 5. In order to apply M e t h o d o l o g y 4 the or ig ina l thresholds m a y have to be changed, so a l l the algori thms can detect the m i n i m u m number o f p ixe ls . A compar i son o f Tables 14 and 15 show that o n l y i n one case the thresholds were changed, and b y a smal l amount. T h i s shows that the p ixe l s detected b y both methodologies i n the target are ve ry s imi lar . T h e threshold values used i n Methodo log ies 1, 2, and 4 d i d not present a large var iabi l i ty . Therefore, there is the poss ib i l i ty that a s ingle threshold w o u l d g ive an acceptable detection performance. 88 In the coherence test ( in M e t h o d o l o g y 2) the R i c i a n threshold seems to be more sensitive to p i x e l intensities. Decreas ing the threshold values seems to inc lude fewer bright samples. T h i s is expected as the R i c i a n threshold is related to the signal-to-clutter ratio. Table 16 shows numer ica l results for Methodo log ies 1, 2 and 4. To ta l false a larm count, total false a larm rate and false a larm count per vegetation type are presented. Resul ts f rom the scenes conta ining Targets 1, 2, 4, 12 and 20 were used to calculate the l o w vegetation totals. Resul ts f rom the scenes containing Targets 14 and 21 were used to calculate the h i g h vegetation totals. T h e clutter surrounding Target 7 was considered neither h i g h nor l o w because o f its vegetation size and roughness. Therefore, results f rom Target 7 scene were shown i n a separate row, w h i c h was named m e d i u m vegetation. To ta l false a larm results presented i n Tab le 16 show that the overa l l performance o f M e t h o d o l o g y 1 was better than M e t h o d o l o g y 2, but M e t h o d o l o g y 4 presented the best results. It is important to notice that, g iven the sma l l s ize o f the scenes and the re la t ively smal l number o f targets, these numbers should be analyzed w i t h care. F o r example, app ly ing M e t h o d o l o g y 2 resulted i n 7 false alarms i n Target 7 alone, w h i c h had a h igh influence on the total result. A n a l y s i s o f Tab le 16 also shows that M e t h o d o l o g y 1 was ve ry effective when appl ied to scenes w i t h l o w vegetation clutter, w h i l e M e t h o d o l o g y 2 was less effective than M e t h o d o l o g y 1. W h e n the clutter was composed o f both m e d i u m and h igh vegetation, M e t h o d o l o g y 1 performed less effectively than it d i d w i t h l o w vegetation clutter, al though it s t i l l performed better than M e t h o d o l o g y 2. T h i s supports the assumption that the D C A methodology w o u l d be less effective i n forested areas. In these cases, M e t h o d o l o g y 4 performs better than both Methodolog ies 1 and 2, contr ibut ing to decrease the false a larm rates. T h i s leads to the m a i n conc lus ion o f this thesis: combined results f rom the Coherence Test f rom the Symmet r i c Scattering Character izat ion M e t h o d ( S S C M ) and the D C A methodology can improve the D C A methodology when the clutter has h i g h vegetation. 89 CHAPTER 7 - CONCLUSIONS 7.1 Summary This work, experimental and heuristic, illustrates a practical approach to target detection and is intended to serve as a reference for people building operational SAR remote sensing systems. Its purpose was to: 1. Review an existing target detection methodology that was developed to detect crashed airplanes in a specific clutter; 2. Test this methodology with a more diverse set of targets and clutter types; 3. Examine other available target detection algorithms and compare their performance with the methodology above; 4. Develop improvements to these methodologies to give good detection performance for a wide range of target and clutter types. The starting point of this work is the methodology we call Detection of Crashed Airplanes (DCA). This methodology was applied by Lukowski in experiments to detect crashed airplanes in a low vegetation clutter. The DCA methodology uses the following algorithms: Polarimetric Whitening Filter (PWF), Even Bounce Analysis and Cameron Decomposition. In this thesis the D C A methodology was reproduced and re-applied to the same dataset (referred to as the Ottawa Dataset in Chapter 5 ) and then applied on two other datasets (the Westham Island dataset and the Gagetown dataset) containing different man made targets and different vegetation clutters. 90 F o u r methodologies were tested and compared: M e t h o d o l o g y 1 is the D C A methodology. M e t h o d o l o g y 2 comprises the Coherence Test proposed b y T o u z i and Charbonneau combined w i t h C a m e r o n Decompos i t ion . M e t h o d o l o g y 3 is the appl ica t ion o f Freeman-D u r d e n D e c o m p o s i t i o n w i t h a threshold appl ied to the dihedrals percentage. M e t h o d o l o g y 4 is the appl icat ion o f the Coherence Test i n combina t ion w i t h M e t h o d o l o g y 1. A l l data sets used i n this research were fu l ly polar imetr ic , C-band data acquired b y the C V -580 S A R system. G r o u n d reference data consis t ing o f photographs and coordinates o f potential targets were acquired on f ie ld trips o n September, 2004 i n Wes tham Island, south o f V a n c o u v e r ( B C ) . G r o u n d reference data for the Ot tawa and Gage town datasets were also available. Resul ts for the four methodologies appl ied to 9 sub-images o f the polar imetr ic data sets have been described. F o r the images processed i n this work , M e t h o d o l o g y 1 produces more false alarms when appl ied to higher vegetation clutter than w h e n appl ied to the Ot tawa dataset. M e t h o d o l o g y 2 presented a marg ina l ly higher number o f false alarms. M e t h o d o l o g y 3 presented an extremely h igh number o f false alarms i n a l l situations. M e t h o d o l o g y 4 generated the best detection results overa l l (i.e., it presented the smallest number o f false alarms w h i l e detecting the k n o w n targets). In the processing o f each image, m a n y different threshold combinat ions were tested i n each o f the methodologies. T h e opt imal threshold values are the ones that y i e l d the lowest false a larm rate for each image w h i l e detecting the k n o w n targets. T h e opt imal thresholds were reported for each image and overa l l range o f thresholds was determined. F o r Methodo log ies 1, 2 and 4 there was li t t le var ia t ion i n the threshold values used across the var ious images, w h i l e M e t h o d o l o g y 3 needed a w i d e range o f thresholds i n order to detect the target. Further analysis showed that for l o w vegetation clutters, Methodo log ies 1 and 2 typ ica l ly detect the target w i t h no or few false alarms. F o r h i g h vegetation clutters there are a lways 91 false alarms and i n these situations the appl ica t ion o f M e t h o d o l o g y 4 results i n the lower false a larm rates. These results support the w o r k carried out b y L u k o w s k i for clutters o f l o w vegetation. T h e results also extend the D C A methodology to different k inds o f clutter and targets, propos ing a new approach that performs better for a var iety o f situations. 7.2 Research Contr ibut ions W e have developed a new methodology that improves the detectabili ty o f m a n made targets i n h i g h vegetation clutter condit ions. In pursu ing this goal , the f o l l o w i n g contributions can be l is ted: 1. T h e assessment o f the effectiveness o f the Detec t ion o f Crashed Ai rp l anes ( D C A ) methodology w h e n appl ied to different data sets, different k inds o f clutters and different k inds o f targets. 2. The assessment o f the effectiveness o f us ing the Freeman-Durden D e c o m p o s i t i o n i n the classif icat ion o f targets that are supposedly coherent. 3. T h e appl icat ion o f the Coherence Test f rom the S S C M algor i thm as a detection a lgor i thm replacing both the P W F a lgor i thm and the E v e n B o u n c e A n a l y s i s a lgor i thm i n the D C A methodology and the assessment o f its effectiveness w h e n compared to the or ig ina l D C A methodology. 4. T h e appl ica t ion o f the Coherence Test a lgor i thm i n combina t ion w i t h the three algori thms used i n the D C A methodology ( P W F , E v e n B o u n c e A n a l y s i s and C a m e r o n Decompos i t ion) as a new approach to decrease false 92 alarm rates when the D C A methodology is appl ied for h i g h vegetation clutter. 5. T h e p rov i s ion o f a source o f informat ion on the appl icat ion o f radar polar imetry for target detection for users o f data avai lable f rom current and future spaceborne polar imetr ic S A R miss ions such as T e r r a S A R , R A D A R S A T - 2 and C O S M O S - S k y m e t . 7.3 Future W o r k Based on the experience obtained w i t h this study, the f o l l o w i n g issues cou ld be suggested for future w o r k i n this f ie ld : 1. A n a l y s e more target and clutter types to so l id i fy our approach to setting the detection threshold, establishing and va l ida t ing threshold models related to specif ic vegetation types. A p p l y a s ingle threshold value for a l l scenes (apply ing Methodolog ies 1, 2 and 4) and analyse its performance. 2. Assess the effect o f look direct ion i n the effectiveness o f m a n made target detection methods b y acqui r ing different data sets us ing fl ight l ines w i t h different orientations. 3. Longe r wavelengths cou ld potent ial ly be helpful for the detection o f m a n made targets under vegetation cover. Exper iments w i t h mult i - f requency S A R acquisit ions cou ld assess the effectiveness o f the target detection algori thms when targets are total ly or par t ia l ly covered b y vegetation. 4. S A R experiments related to classif icat ion are t yp i ca l l y carr ied out on flat terrain, as were the ones w e studied. Ter ra in effects due to the geometry o f 93 radar data collection such as foreshortening, layover and very low local incidence angles usually lead to poor classification performance if only intensity is used. Man made targets located on undulated terrain might still be detected by their distinctive polarimetric signature even when the target-to-clutter ratio is low. Experiments should be performed to examine terrain effects. The value of thresholds and the performance of the algorithms depend on the resolution of the radar system and on the signal/noise ratio of the data. In general, satellite SAR data has a lower resolution and a lower signal/noise ratio than airborne data. Experiments should be done with satellite data (or satellite data simulated from airborne data) to examine the effectiveness of the methodologies under the lower resolution and signal/noise ratios. 94 BIBLIOGRAPHY [1] U l a b y , F . , and C . E l a c h i , "Radar Polar imet ry for Geosc ience A p p l i c a t i o n s " , 1990. [2] T o u z i , R . , W . M . Boerner , J .S . Lee , and E . Lueneburg , " A rev iew o f polar imetry i n the context o f synthetic aperture radar: concepts and informat ion extraction", C a n . J . Remote Sensing, V o l . 30, N o . 3, pp. 380-407, 2004. [3] Boerner , W . M . , H . M o t t , C . L iv ings tone , B . B r i s c o , R . B r o w n , J .S . Paterson, E . Luneburg , J .J . van Z y l , D . Randa l l , and P . B u d k e w i t s c h , Po la r imet ry i n Remote Sensing: B a s i c and A p p l i e d Concepts , Chapter 5 i n T h e M a n u a l o f Remote Sensing, 3rd edit ion, Pr inc ip les and App l i ca t i ons o f Imaging Radar , edi tor- in-chief R . A . Rye r son , A m e r i c a n Socie ty for Photogrammetry and Remote Sensing, Bethesda, M d . , 1998. [4] O l ive r , C . J . , and S. Quegan, "Unders tanding Synthetic Aper ture Radar Images", A r t e c h House , 1998. [5] E l - S a b a , A . M . , "Potent ia l A p p l i c a t i o n o f V i s i b l e Pass ive Imaging Po la r imet ry i n the D i s c r i m i n a t i o n o f R e a l Targets and D e c o y s " , Proc . S P I E V o l . 5816, 2005. [6] V a n der Sanden, J . J . "An t i c ipa t ed applications potential o f R A D A R S A T - 2 data", C a n . J . Remote Sensing, V o l . 30, N o . 3, pp. 369-379, 2004. [7] Staples, G . , J . Hornsby , W . Branson , K . O ' N e i l l , and P . R o l l a n d , " T u r n i n g the sc ient i f ica l ly possible into the operat ionally pract ical : R A D A R S A T - 2 commerc ia l i za t ion p lan" , C a n . J . Remote Sensing, V o l . 30, N o . 3, pp. 408-414, 2004. 95 [8] L u k o w s k i , T . I. and F . J . Charbonneau, "Synthet ic aperture radar and search and rescue", In Proceedings o f the I E E E International Geosc ience and Remote Sensing S y m p o s i u m ( I G A R S S ' O O ) , H o n o l u l u , H a w a i i , pp. 2374 - 2376, 2 4 - 2 8 J u l y 2000. [9] L u k o w s k i , T . I. and F . J . Charbonneau, "Synthet ic aperture radar and search and rescue: detection o f crashed aircraft u s ing imagery and interferometric methods", Canad ian Journal o f Remote Sensing, V o l . 28, N o . 6, pp. 770-781, 2002. [10] L u k o w s k i , T . I., B . Y u e , F . J . Charbonneau, F . K h e l l a h , and R . K . H a w k i n s , "Detec t ion o f crashed aircraft i n polar imetr ic imagery" C a n . J . Remote Sensing, V o l . 30, N o . 3, pp. 484-495, 2004. [11] N o v a k , L . M . , and B u r l , M . C . , " O p t i m a l speckle reduct ion i n polar imetr ic S A R imagery" , I E E E Transactions o n Aerospace and E lec t ron ic Systems, V o l . 26, pp. 2 9 3 - 3 0 5 , 1990. [12] N o v a k , L . M . , B u r l , M . C . , and I rv ing , W . W . , " O p t i m a l polar imetr ic processing for enhanced target detection", I E E E Transactions o n Aerospace and E lec t ron ic Systems, V o l . 29 , N o . l , p p . 2 3 4 - 2 4 3 , 1993. [13] Evans , D . L . , T . G . F a i r , J .J . V a n Z y l , and H A . Zebker , "Radar polar imetry: analysis tools and applicat ions", I E E E Transactions o n Geosc ience and Remote Sensing, V o l . 26, N o . 6, pp. 774 -789 , 1995. [14] C u m m i n g , I. " A d v a n c e d Polar imetr ic Tu to r i a l " , N o e t i x Research, 2003 . [15] D o z o n o , H . , R . K . H a w k i n s , K . Murnaghan , " R e a d i n g Conva i r -580 Polar imet r ic S A R Data" , C C R S - T N - 2 0 0 4 - 0 0 6 , 2005. 96 [16] T o u z i , R . and F . Charbonneau, "Character izat ion o f target symmetr ic scattering us ing polar imetr ic S A R s " . I E E E Transactions o n Geosc ience and Remote Sensing, V o l . 40, pp.2507-2516, 2002. [17] Egan , W . G . , "Photometry and polar iza t ion i n remote sensing", E lsev ie r , 1985. [18] Freeman, A . and S. Durden , " A three-component scattering m o d e l to describe polar imetr ic S A R data", I E E E Trans. Geosc ience and Remote Sensing, v o l . 36, no. 3, pp. 9 6 3 - 9 7 3 , M a y 1998. [19] Y e r e m y , M . , J . C a m p b e l l , K . Mat tar and T . Potter, " O c e a n survei l lance w i t h polar imetr ic S A R " , C a n . J . Remote Sensing, v o l . 17, no. 4, pp. 328 -344 , A u g . 2001 . [20] Y e r e m y , M . , G . G e l i n g and M . R e y , "Resul ts o f the crusade ship detection t r ia l : Po la r imet r ic S A R , " i n Proc. IGARSS, Jun. 2 4 - 2 8 , v o l . 2, pp. 711 - 713, 2002. [21] Chotoo , K . , B . Huxtab le , A . M a n s f i e l d and H . Ra i s , "P robab i l i t y o f detection o f downed aircraft us ing S A R polar imetry" , In Proc. SPIE, v o l . 4050, pp. 163-168 , A p r . 2000. [22] Cameron , W . L . , N . N . Y o u s s e f and L . K . L e u n g , "S imula ted polar imetr ic signatures o f p r imi t ive geometrical shapes," I E E E Trans. Geosc ience and Remote Sensing, v o l . 34, no. 3, pp. 793 - 803, M a y 1996. [23] Kre i then , D . E . , S. Ha lversen and G . O w i r k a , " D i s c r i m i n a t i n g Target F r o m Clut te r" i n The Lincoln Laboratory Journal, v o l . 6, number 1, 1993. [24] Kennaugh , K . , "Po la r i za t ion properties o f radar reflections", M . S c . thesis, T h e O h i o State Un ive r s i ty , C o l u m b u s , O h i o , 1952. 97 [25] H u y n e n , J . R . , "Measurement o f the target scattering mat r ix" , Proceedings o f the I E E E , V o l . 53, N o . 8, pp. 936 -946 , 1965. [26] V a n Z y l , J .J . "Unsuperv ised classif icat ion o f scattering behavior us ing radar polar imetry data", IEEE Trans, on Geoscience and Remote Sensing, V o l . 27, N o . 1, pp. 37 - 4 5 , 1989. [27] C loude , S .R. , and Pottier, E . , " A rev iew o f target decompos i t ion theorems i n radar polar imetry" , I E E E Trans, on Geoscience and Remote Sensing, V o l . 34, N o . 2, pp. 498 -518, 1996. [28] L a r s o n , V . , L . N o v a k and C . Stewart, "Joint spatial polar imetr ic whi ten ing filter to improve S A R target detection performance for spat ial ly distributed targets" In Proc . SPIE, V o l . 2230, pp. 285-301, 1994. [29] Henry , C . , J . C . Souyr is and P . Mar thon , "Target Detec t ion and A n a l y s i s Based on Spectral A n a l y s i s o f a S A R Image: a S imula t ion A p p r o a c h " , Proceedings o f I G A R S S '03, Toulouse , France, pp. 2 0 0 5 - 2 0 0 7 , 2003. [30] G . D e Grand i , J . S. Lee , D . Schuler , and E . N e z r y , "Texture and speckle statistics i n polar imetr ic S A R synthesized images," I E E E Trans. G e o s c i . Remote Sens., v o l . 4 1 , no. 9, pp. 2070-2088, Sep. 2003. 98 APPENDIX A - SUPPORTING PROCESSING RESULTS T h e processing o f the four methods generated a large amount o f plots. In order to make the presentation more clear, w e decided to keep i n chapter 6 a l l the results o f targets 7, 14 and 21 and the f inal detection map o f each method for the remain ing targets. T h e intermediate results for each o f these remain ing targets (Targets 1, 2, 4, 5, 12 and 20) were transferred to this appendix. T h e plots are presented i n the f o l l o w i n g manner: E a c h target is i n one section o f this appendix and contains four figures. E a c h o f these figures contains the intermediate plots for one methodology. Capt ions indicate the methodology number and the contents o f each plot. A l l numer ica l results o f this thesis are presented i n Chapter 6. 99 A . l - Target 1 P W F Target Map (K = 3) Even Bounce Target Map (K = 8 ) 100 200 300 400 500 600 100 200 400 500 Cameron , P W F and Even B o u n c e 5 -10 -15 -20 30 -35 • AO -100 200 300 400 500 600 Figure 48: Target 1 - P lots f rom M e t h o d o l o g y 1. U p p e r left: P W F target map. U p p e r right: E v e n B o u n c e target map. B o t t o m : combined result o f C a m e r o n (dihedrals p lus narrow dihedrals), P W F and E v e n Bounce . K is the C F A R constant. 100 Coherence Test Target Map Cameron and Coherence Test Map 100 200 300 400 (a) (b) Figure 49: Target 1 - Plots f rom M e t h o d o l o g y 2 . (a) Coherence Test target map. (b) C o m b i n e d result o f C a m e r o n (dihedrals plus narrow dihedrals) and Coherence Test. Target Map Cameron, P W F , E B and Coherency Test Target Map Figure 50: Target 1 - Plots f rom Methodolog ies 3 and 4. (a) F reeman-Durden target map showing p ixe l s that present a percentage o f dihedral scattering above the chosen threshold, (b) C o m b i n e d results o f C a m e r o n (dihedrals plus narrow dihedrals) , P W F , E v e n B o u n c e and Coherence Test. 101 A.2 - Target 2 P W F Target Map (K = 2) Even Bounce Target Map (K = 6 ) MI i i i i 1 n ' • i» ' . !' l f l | l ' l l l I •. i . i i i 1 i i , '. 1 11 i-: i 1 ' ' i i 1 i , • 1 I i i i 11 50 100 150 200 250 300 350 400 450 500 50 100 150 200 250 300 350 400 450 500 Cameron, PWF and Even Bounce 50 100 150 200 250 300 350 400 450 500 Figure 51: Target 2 - Plots f rom M e t h o d o l o g y 1. U p p e r left: P W F target map. U p p e r right: E v e n B o u n c e target map. B o t t o m : combined result o f C a m e r o n (dihedrals plus narrow dihedrals), P W F and E v e n Bounce . K is the C F A R constant. 102 Coherence Test Target Map Cameron and Coherence Test Map 25 . 35 I, l.ll II . I I I . M l . II I I I I, 11 I I \ :. I' 1 -IV v ; • „ i.i 50 100 150 200 250 300 350 400 450 500 50 100 150 200 250 300 350 400 450 500 (a) (b) Figure 52: Target 2 - Plots f rom M e t h o d o l o g y 2. (a) Coherence Test target map. (b) C o m b i n e d result o f C a m e r o n (dihedrals p lus narrow dihedrals) and Coherence Test. Target Map Cameron, P W F , E B and Coherency Test Target Map Figure 53: Target 2 - Plots f rom Methodolog ies 3 and 4. (a) F reeman-Durden target map showing p ixe l s that present a percentage o f dihedral scattering above the chosen threshold, (b) C o m b i n e d results o f C a m e r o n (dihedrals p lus narrow dihedrals) , P W F , E v e n B o u n c e and Coherence Test. 103 A.3 - Target 4 P W F Target Map (K = 2) E w n Bounce Target Map (K = 2 ) Vy,i • V,',, • ',• ' 1 1 i i 5 • 10 -I ® 0 ' 20 | 25 -''I'l,,'!, V 20 40 60 80 100 120 140 160 180 200 220 20 40 60 80 100 120 140 160 180 200 220 Cameron , P W F and Even B o u n c e 20 40 60 80 100 120 140 160 180 200 220 Figure 54: Target 4 - Plots f rom M e t h o d o l o g y 1. U p p e r left: P W F target map. U p p e r right: E v e n B o u n c e target map. B o t t o m : combined result o f C a m e r o n (dihedrals plus narrow dihedrals), P W F and E v e n Bounce . K is the C F A R constant. P i x e l s i n the rectangle are other m a n made targets i n this complex scene. C i r c l e indicates locat ion o f a tree. 104 Coherence Test Target Map Cameron and Coherence Test Map 20 40 60 80 100 120 140 160 180 200 220 20 40 60 80 100 120 140 160 180 200 220 (a) (b) Figure 55: Target 4 - Plots f rom M e t h o d o l o g y 2. (a) Coherence Test target map. (b) C o m b i n e d result o f C a m e r o n (dihedrals plus narrow dihedrals) and Coherence Test. P ixe l s i n the rectangle are other m a n made targets i n this complex scene. C i r c l e indicates locat ion o f a tree. Target Map Cameron. PWF, EB and Coherency Test Target Map 15 20 i i r -II i i i - I I ! I II ,l I, II 1 i 1 HI ' M i i i i I, n . ; i-v . 1 ^ •• ' . II 1 1 1 1 20 40 60 80 100 120 140 160 180 200 220 20 40 60 80 100 120 140 160 180 200 220 (a) (b) Figure 56: Target 4 - Plots f rom Methodolog ies 3 and 4. (a) Freeman-Durden target map showing p ixe ls that present a percentage o f dihedral scattering above the chosen threshold, (b) C o m b i n e d results o f C a m e r o n (dihedrals plus narrow dihedrals) , P W F , E v e n B o u n c e and Coherence Test. P i x e l s i n the rectangle are other m a n made targets i n this complex scene. C i r c l e indicates loca t ion o f a tree. 105 A.4 - Target 5 P W F Target Map (K = 1) Even Bounce Target Map (K = 4 ) 25 20 40 60 80 100 120 140 160 • 180 Figure 57: Target 5 - Plots f rom M e t h o d o l o g y 1. U p p e r left: P W F target map. U p p e r right: E v e n B o u n c e target map. B o t t o m : combined result o f C a m e r o n (dihedrals plus narrow dihedrals) , P W F and E v e n Bounce . K is the C F A R constant. 106 Coherence Test Target Map Cameron and Coherence Test Map 20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180 (a) (b) Figure 58: Target 5 - Plots f rom M e t h o d o l o g y 2. (a) Coherence Test target map. (b) C o m b i n e d result o f C a m e r o n (dihedrals p lus narrow dihedrals) and Coherence Test. Target Map Caneroi, PVUF.EB aidCol«e 1(3^  T««ttarget Uap (a) (b) Figure 59: Target 5 - Plots f rom Methodolog ies 3 and 4. (a) Freeman-Durden target map showing p ixe ls that present a percentage o f dihedral scattering above the chosen threshold, (b) C o m b i n e d results o f C a m e r o n (dihedrals plus narrow dihedrals) , P W F , E v e n B o u n c e and Coherence Test. 107 A.5 - Target 12 P W F Target Map (K = 2) Even Bounce Target Map (K = 6 ) 14 20 40 60 80 100 120 140 160 Figure 60: Target 12 - Plots f rom M e t h o d o l o g y 1. U p p e r left: P W F target map. U p p e r right: E v e n B o u n c e target map. B o t t o m : combined result o f C a m e r o n (dihedrals p lus narrow dihedrals), P W F and E v e n Bounce . K is the C F A R constant. 108 Coherence Test Target Map Cameron and Coherence Test Map 60 80 100 120 140 160 (a) (b) Figure 61: Target 12 - Plots f rom M e t h o d o l o g y 2. (a) Coherence Test target map. (b) C o m b i n e d result o f C a m e r o n (dihedrals plus narrow dihedrals) and Coherence Test. Target Map Cameron, P W F , E B and Coherency Test Target Map 20 40 60 80 100 120 140 160 20 40 60 80 100 120 140 161 (a) (b) Figure 62: Target 12 - Plots f rom Methodo log ies 3 and 4. (a) Freeman-Durden target map showing p ixe l s that present a percentage o f dihedral scattering above the chosen threshold, (b) C o m b i n e d results o f Cameron (dihedrals plus narrow dihedrals) , P W F , E v e n B o u n c e and Coherence Test. 109 A.6 - Target 20" * T h i s is the scene used b y L u k o w s k i et al. [10] P W F Target Map (K = 3) E w n Bounce Target Map {K = 5 ) 200 400 600 800 1000 1200 200 400 600 800 1000 1200 Cameron, P W F and Even Bounce 200 400 600 800 1000 1200 Figure 63: Target 20 - P lots f rom M e t h o d o l o g y 1. U p p e r left: P W F target map. U p p e r right: E v e n B o u n c e target map. B o t t o m : combined result o f C a m e r o n (dihedrals plus narrow dihedrals), P W F and E v e n Bounce . K is the C F A R constant. Rectangles show locations o f corner reflectors. 110 Coherence Test Target Map Cameron and Coherence Test Map 10 10 1 D 20 • 20 30 i a i ' ; , 0 • 30 0 0 40 V, 40 50 - 50 -60 • 1 i 60 -'A' 1 " , i 70 , i ' . i -i 70 -80 ,'' >i I, - 80 • ' , I',1" i 90 - ' i i i 90 i i i i i r -_1 I I J_J I I 1 » u t I I I I I [ 200 400 600 800 1000 1200 200 400 600 800 1000 1200 (a) (b) Figure 64: Target 20 - Plots f rom M e t h o d o l o g y 2. (a) Coherence Test target map. (b) C o m b i n e d result o f C a m e r o n (dihedrals plus narrow dihedrals) and Coherence Test. Rectangles show the loca t ion o f corner reflectors. Target Map Cameron, P W F , E B and Coherency Test Target Map (a) (b) Figure 65: Target 20 - Plots f rom Methodolog ies 3 and 4. (a) Freeman-Durden target map showing p ixe l s that present a percentage o f dihedral scattering above the chosen threshold, (b) C o m b i n e d results o f C a m e r o n (dihedrals plus narrow dihedrals) , P W F , E v e n B o u n c e and Coherence Test. I l l 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.831.1-0101077/manifest

Comment

Related Items