Advances in SAR Change Detection

Transcription

1 Lesle M. ovak Scetfc Sstems Compa, Ic. 500 West Cummgs Park, Sute 3000 Wobur, MA 080 UITED STATES E-mal: ABSTRACT SAR chage detecto performace usg coheret chage detecto CCD) ad o-coheret chage detecto CCD) algorthms s demostrated usg hgh-resoluto sthetc aperture radar SAR) mager gathered b the Geeral Damcs Data Collecto Sstem. CCD performace comparsos usg phaseol mager versus comple ampltude ad phase) mager are also preseted. A ew mage qualt metrc, the Uversal Image Qualt Ide s descrbed ad used to detect chages betwee a SAR test referece versus test) mage-par; the chage mage s show to be ver smlar to the correspodg CCD mage. Studes of SAR coheret chage detecto usg the Mamum Lkelhood Estmate MLE) of coherece are also preseted; detecto performace ROCs PD vs. PFA curves) are preseted comparg coheret ad o-coheret chage detecto algorthms CCD, MLE, ad CCD).. ITRODUCTIO Fg. shows a tpcal SAR mage gathered b the Geeral Damcs Data Collecto Sstem DCS). The mage sze s pels ad the resoluto of the data s ft b ft. The bo supermposed o the mage shows a rego of terest cotag two terestg chage detecto scees that wll be vestgated. These tal o-coheret chage detecto studes wll focus o a scee cotag parked vehcles the vehcle scee ); the tal coheret chage detecto studes wll focus o a scee cotag a subtle mamade dsturbace due to people that walked a grass area the racetrack scee ). Fgures ad 3 show the 0404 rego of terest cotag the chage detecto scees; Fg. s the SAR referece mage ad Fg.3 s the correspodg SAR test mage. These referece ad test mages wll be processed usg the coheret chage detecto CCD) algorthm [] ad o-coheret chage detecto CCD) algorthm [] as defed Table. Table : SAR Chage Detecto algorthms: Coheret Chage algorthm left), o-coheret Chage algorthm rght). CCD k k k * k k + k k + CCD k k k k + k k k + k + STO-E-SET

2 Fgure : SAR Referece mage, sze pels; area of terest sze pels. 9 - STO-E-SET-7-03

3 Fgure : SAR referece mage. STO-E-SET

4 Fgure 3: SAR test mage. Fgures 4 ad 5 show the CCD ad CCD chage mages obtaed from comparsos of a test mage ad a prevousl gathered referece mage; several detected chages are poted out o the CCD ad CCD mages. ote that ol coheret chage detecto has detected the racetrack the grass area ad although the chage ampltude betwee the referece ad test mages s too small to be detected b the CCD algorthm, the chage phase.e., the coherece ) betwee the referece ad test mages s suffcet to permt detecto of ths subtle chage b the CCD algorthm. 9-4 STO-E-SET-7-03

5 Fgure 4: Coheret Chage mage. STO-E-SET

6 Fgure 5: o-coheret Chage mage. 9-6 STO-E-SET-7-03

7 . A AALYSIS OF THE CCD ALGORITHM Fgure 6 presets a smplfed block dagram of the basele coheret chage detecto CCD) algorthm we use these studes. The referece ad test mages are comprsed of comple pels, deoted as mages X m, ad X m,. As dcated the fgure, the algorthm calculates the coherece "γ " betwee the referece ad test mages.e., the magtude of the comple cross-correlato betwee the referece ad test mages). Also these studes, the coherece s calculated usg a 33 cluster of the comple mage data thus, M ). IMAGE X m, IMAGE ˆ, X m γ M m 0 0 M m 0 0 X m, X m, X ˆ M * m, m 0 0 X ˆ * m, γ > 0.9 γ < 0.9 Fgure 6: Block dagram of the CCD algorthm. A aalss of coheret chage detecto algorthm s gve as follows. We wrte the comple pel data ampltude ad phase format: jϕ jϕ m, * m, X X e m, m, ad X m, X m, e The coherece equato defed Fgure 6, epressed ampltude/phase format, s gve as follows: γ M m 0 o M m 0 0 X m, X m, j m X m e m ˆ ˆ * ϕ, ϕ, ), M m 0 0 Xˆ * m, et we make the assumpto that the magtude of the comple referece ad test pels are equal, mplg that the coherece betwee the mages wll deped ol o the phase dffereces betwee pels. Wth ths smplfg assumpto, the followg result s obtaed: f X m, Xˆ m, X m, the γ M m 0 0 We are terested comparg the performace of the CCD algorthm usg the orgal comple mage data versus usg the phase-ol data; ths stud we tall focus o a sub-mage of the grass area cludg the "racetrack feature". Fgure 7 shows a 5656 pel coherece mage cotag the racetrack wth a selected patch of grass outled; ote that the average coherece of ths 5656 pel sub-mage s e j ϕ m, ˆ ϕ m, ) STO-E-SET

8 Fgure 7: CCD mage of "racetrack" area; average mage coherece ; a grass patch s outled. Fgures 8 ad 9 preset a sde-b-sde comparso of the grass patch CCD mages calculated usg the orgal comple SAR data versus usg the ampltude-ormalzed) phase-ol SAR data. Ths small patch of grass has average coherece usg the orgal comple SAR data, whereas the average coherece usg phase-ol data. Thus, ths eample seems to dcate that both ampltude ad phase.e., the comple pel data) should be used mage eplotato usg the CCD algorthm; there appears to be a loss the level of coherece usg phase-ol mages. 9-8 STO-E-SET-7-03

9 Fgure 8: SAR CCD mage of grass patch. Fgure 9: CCD mage of grass patch from phase-ol mages. Fgure 0 valdates the cojecture that the best CCD mage s obtaed usg both the ampltude ad phase of the data formg the chage mage; the fgure dcates that the average coherece usg comple data whch was see fgure 7) has bee reduced to usg phase-ol data -- ad the chage mage Fgure 0 also shows that a larger umber of low coherece pels have bee obtaed usg phase-ol data. STO-E-SET

10 Fgure 0: CCD mage,"racetrack", phase-ol data. 3. RESULTS USIG THE "UIVERSAL IMAGE QUALITY IDEX" Ths secto presets a summar of some terestg results that were obtaed usg a approach developed Refereces 5 ad 6; these authors have proposed a mage qualt metrc called the "Uversal Image Qualt Ide" ad the have demostrated the applcato of ther ew metrc to photographs such as the well-kow "Lea" ad others. Although our SAR mages are comprsed of comple pel values, t was of terest to appl ths ew metrc to SAR test mages. Wth ths goal md, we gve a bref descrpto of the ew metrc ad the preset results of applg the approach to the SAR mager show the prevousl descrbed CCD ad CCD studes. Table presets detals of the Uversal Image Qualt Ide. There are two test mages, deoted as mage X ad mage Y; the cotet of SAR chage detecto, X deotes the referece mage test mage) ad Y deotes the test mage test mage). The table shows a par of 33 clusters of test values to be compared, ad our goal s to fd the chages betwee the referece ad test test mages. The mea, varace, ad covarace of the test values are calculated as dcated the Table. 9-0 STO-E-SET-7-03

11 Table : Defto of statstcs used calculatg the "uversal mage qualt de". ) ) ) ) ) ) ) ) X mage X mage mage Y mage Y ) ) ) ) ) ) As preseted Table 3, these mea, varace, ad covarace values are used to form three mage qualt factors: 3,, Q ad Q Q. Q s a measure of structural smlart, Q s a measure of the smlart of the meas, ad 3 Q s a measure of the smlart of the cotrasts. STO-E-SET

12 Table 3: Defto of the "uversal mage qualt de". The mage qualt de, Q, s calculated from 33 clusters of test data at each pel locato the mage, resultg a ew mage deoted as the Uversal IQ Ide Image; ths ew mage s a represetato of the chages that est betwee the referece ad test mages. The SAR referece ad test mages show prevousl Fgures ad 3 were coverted to test mages ad processed as descrbed above. The resultg Uversal IQ Ide Image we obtaed s show Fgure ; a average IQ de of was obtaed from the mage show. The terestg observato gleaed from the mage show Fgure s that ths chage mage vsuall appears to be a CCD mage but ths chage mage was obtaed from test-ol SAR referece ad test mager. ], [ ) ) Q 3 Q Q Q Q ], [ + Q ], [0 ) ) + + Q ] 0, [ Q Smlart of Image Meas A measure of Image Structural Smlart : Q : Q Smlart of Image Cotrasts : 3 Q 9 - STO-E-SET-7-03

13 Fgure : Uversal mage qualt de mage obtaed usg test mages show fgures & 3. Further aalss of the mages produced b each of the factors Q, Q, ad Q3 showed that the mage produced b the factor Q was the domat mage, ad ths factor s smpl the cross-correlato of the SAR test mages. Ths observato has resulted our researchg the lterature o prevous mathematcal aalses of the cross-correlato of SAR test mages ad ts relatoshp wth the SAR coherece parameter see Refereces 7 ad 8). Table 4 presets two fuctoal relatoshps derved the refereces. ρ A correspods to test cross-correlato wthout mea removal Referece 7) ad ρ B correspods to test cross-correlato wth mea removal Referece 8). STO-E-SET

14 Table 4: Coherece relatos vs. cross-correlato of SAR test mages. Addtoal SAR chage detecto studes usg the test cross-correlato deoted as ρ B Table 3 were performed. Fgure presets a sde-b-sde comparso of the Coherece Image left) versus the correspodg chage mage obtaed usg the test mage cross-correlato deoted as ρ B Table 4. Vsuall these mages look qute smlar -- ad the absolute value of the dfferece betwee these mages s preseted Fgure 3. The dfferece error mage shows reasoabl small dffereces betwee the actual coherece values, γ, ad the coherece estmates γ ρ. B B Fgure : Coherece mage vs. appromato. 9-4 STO-E-SET-7-03

15 Fgure 3: Magtude of dfferece mage. STO-E-SET

16 4. PERFORMACE COMPARISO OF CHAGE DETECTIO ALGORITHMS I the prevous sectos we preseted some prelmar comparsos of the CCD vs. CCD chage detecto algorthms. I ths secto we wll vestgate the detecto performace of the Mamum Lkelhood Estmate MLE) of the SAR coherece parameter. We wll quatf ad compare the chage detecto performace of the MLE versus the CCD. I Table 5 below we preset the deftos of these SAR coheret chage detecto algorthms. We wll also compare the detecto performace of these coheret chage detecto algorthms wth the basele o-coheret chage detecto CCD) algorthm defed prevousl Table. I our prevous chage detecto studes we foud that the MLE verso of the coheret chage detecto algorthm gave better detecto performace results tha the comple cross-correlato) CCD verso of the algorthm. These prevousl obtaed results were agreemet wth a paper preseted b Mram Cha of MIT Lcol Laborator at the IEEE Statstcal Sgal Processg Workshop, A Arbor, Mchga Referece [9]). I Mram Cha s paper, t was cojectured, based o theoretcal aalses of the MLE ad CCD coheret chage detecto algorthms, that the MLE verso should provde better coheret chage detecto performace.e., PD vs. PFA ROCs) tha the CCD verso f the referece ad test mages have appromatel equal uderlg varaces. Clearl, ths s the case for the accuratel calbrated SAR mager used our studes [0]. I ths secto we summarze our studes of these two coheret chage detecto algorthms. Our goal s to determe the sestvt of the MLE algorthm whe the test mage varace s ot comparable to the referece mage varace. We am to show that the MLE outperforms the CCD over some rage of calbrato ga offsets betwee the referece ad test mages, thus, we am to verf the cojecture of Mram Cha. Table 5: Coheret chage algorthms CCD MLE k k k k k k * k * k + k + k k k + k + k + Fgure 4 shows a aeral photo of the "Area-of-Iterest" selected for our SAR chage detecto studes. The area of terest s comprsed of several parkg lots whch are occuped b umerous parked.e., statoar) vehcles. Aalss of the SAR referece ad test mages of ths area were foud to cota a total of 33 vehcles that chaged durg the tme terval betwee the gatherg of the referece ad test mages. Ths set of 33 vehcles were ether "arrvals" or departures" that occurred durg the tme terval betwee the gatherg of the referece ad test mages. 9-6 STO-E-SET-7-03

17 Fgure 4: Aeral photograph of the "area of terest" selected for CCD studes. Fgure 5 presets the SAR referece ad test mages used ths chage detecto stud. The locatos of the chage-detected vehcles are show as crcles supermposed o the mages. There are a total of 33 vehcle detectos these SAR mages; both arrvals ad departures are cotaed the 33 crcles. These chage detectos were verfed b vsuall flckerg betwee the referece ad test SAR mages. Fgure 5: SAR referece mage left), SAR test mage rght). STO-E-SET

18 Fgure 6 left) shows a bar mage of the locatos of the chage-detected vehcles. Ths bar mage s used as groud truth for the targets these CCD/MLE/CCD chage detecto studes. We use the groud truth mage to score the performace of the coheret ad o-coheret chage detecto algorthms evaluated the studes. Detectos that occur the black areas are scored as target detectos. Detectos that occur the whte backgroud area are scored as false alarms. I Fgure 6 rght) we show a "do't care" bad costructed aroud each of the targets to prevet clutter false alarms from beg declared targets. Fgure 6: Locatos of CD targets left); masked locatos of CD targets rght). Table 5 above gves mathematcal equatos defg the SAR chage detecto algorthms to be evaluated ad compared these studes. Two versos of the coheret chage detecto algorthm are gve -- these are deoted as MLE ad CCD, depedg o the algorthm ormalzato used. ote that the MLE ormalzes the umerator b a sum of the referece ad test pel powers, whereas the CCD ormalzes the umerator b a product of the referece ad test powers. The CCD algorthm s the well-kow comple cross-correlato algorthm used for calculatg estmatg) the coherece betwee comple-valued referece ad test mages; we refer to ths algorthm as the "CCD". Sce the MLE algorthm s actuall the Mamum Lkelhood Estmate of the coherece parameter, we refer to ths algorthm as the "MLE". The o-coheret SAR chage detecto algorthm evaluated these studes was gve above Table ; we refer to ths algorthm as the "CCD". A comparso of the performace of these SAR chage detecto algorthms s gve the followg Fgures 7-0; these fgures show PD versus PFA obtaed va chage detecto processg usg CDalgorthm bo szes 33, 55, 77, ad 99, respectvel. Fgure 7 shows the PD/PFA ROC curves obtaed usg a 33 CD-algorthm bo sze; the curves show that coheret chage detecto usg the MLE algorthm acheves sgfcatl better detecto performace tha the classcal CCD algorthm. At 70% PD, CCD gave 700 target-szed FAs, whereas the MLE gave 0 FAs. We also observe that the o-coheret CCD chage detecto provdes better performace tha both of the coheret chage detecto algorthms. We evaluated detecto performace versus the percetage %) of detected pels a target-se bo, ad we foud that detecto performace was ot ver sestve to ths parameter; the ROC curves obtaed for 5% to 30% are tghtl clustered. 9-8 STO-E-SET-7-03

19 Fgure 7: Chage detecto performace ROCs; CD algorthm bo sze 33. Fgure 8 summarzes chage detecto performace usg a CD algorthm bo sze of 55. The ROCs show are smlar to those show the prevous Fgure 7. Fgure 8 shows that CCD performace s deftel mproved usg the larger 55 bo sze ad CCD performace s also mproved wth the larger bo sze. The MLE algorthm shows slghtl degraded performace wth the larger bo sze. Fgure 8: Chage detecto performace ROCs; CD algorthm bo sze 55. Fgure 9 summarzes CCD, MLE ad CCD algorthm performace usg a CD algorthm bo sze of 77 to costruct the chage mages. The ROCs show are smlar to those show the prevous fgures. CCD performace s aga somewhat mproved usg a larger 77) bo sze. The CCD performace s also mproved wth the larger bo sze; for eample, PD > 90% s acheved wth PFA 0. appromatel 700 FAs). The MLE algorthm has better ROC performace tha the classcal CCD algorthm, but aga CCD gave the best performace. STO-E-SET

20 Fgure 9: Chage detecto performace ROCs; CD algorthm bo sze 77. Fgure 0 summarzes CCD, MLE ad CCD performace usg a 99 bo sze. Coheret chage detecto usg the MLE provdes sgfcatl better detecto performace tha the classcal CCD algorthm. At 60% PD, the CCD gave appromatel 350 target-szed FAs, whereas MLE gave ma fewer FAs. Aga we observe that the CCD algorthm gave better performace tha both coheret chage detecto algorthms. Fgure 0: Chage detecto performace ROCs; CD algorthm bo sze STO-E-SET-7-03

21 I ths paragraph we brefl summarze the fdgs of our chage detecto algorthm comparso studes descrbed above. For the target ad clutter data used our studes, we foud that the o-coheret chage detecto CCD) algorthm acheved the best overall detecto performace; ths was true for all bo-szes tested, ad larger bo-szes gave margall mproved ROCs. PD/PFA curves were ot ver sestve to the detector "Fll" parameter from 5% to 30%. The MLE verso of the coheret chage detector gave better detecto performace tha the CCD verso of the algorthm, ad the smaller bo sze 33) gave the best MLE performace. 4. Vsual Comparsos of the MLE ad CCD Algorthms I ths secto we wll evaluate ad vsuall compare chage mages from the CCD ad MLE versos of the coheret chage algorthm wth the test mage scaled ampltude relatve to the gve referece mage. I these tal studes we scaled the test mage b 0 db, +3 db ad +6 db relatve to the orgal referece mage. Fgure compares CCD ad MLE chage mages obtaed usg the Gotcha SAR referece ad test mages wth K,.e., o test mage ga offset). Usg Gotcha SAR mager [0] smlar to that used our studes, MIT Lcol Laborator also showed that a MLE chage mage has hgher cotrast tha the correspodg CCD chage mage; ths pheomeo was demostrated va chage mages smlar to those show Fgure. ote that the average coherece values of the MLE ad CCD chage mages show Fgure are vs , respectvel; evertheless, we show that the chage vehcles the MLE mage are more easl detected tha the chage vehcles the CCD mage. Fgure : Basele performace; left, CCD mage; rght, MLE mage. Fgure shows addtoal eamples of MLE chage mages obtaed usg ampltude scaled test mages; these chage mages were obtaed b comparg the referece mage wth a scaled test mage, K*test mage)), where the scale factor, K, was ether sqrt), or. The MLE basele K ) average coherece of has bee reduced from ths value to ad for ga offsets of +3 db ad +6 db, respectvel. Although these chage mages have less average coherece, t remas to be determed what these test mage ga offsets wll do to the detectablt of the chage vehcles these mages. STO-E-SET

22 Fgure : Left, MLE mage, K sqrt ); rght, MLE mage, K. We cotued these studes order to quatf the actual chage detecto performace PD vs. PFA) acheved as a fucto of the test mage ga offset parameter, K. Frst we determed the average coherece values obtaed from the CCD ad MLE algorthms versus the test mage ga offset for offsets as large as 0dB; Fgure 3 presets these results. The data gve that fgure dcates that the CCD algorthm s uaffected b the ga mbalace, whereas the average coherece estmated from the MLE algorthm s reduced magtude as the ga mbalace s creased. Thus, the PD/PFA ROC curves obtaed usg the CCD verso of the coheret chage algorthm wll be uaffected b the ga offset, ad the correspodg chage mages for the CCD algorthm wll be detcal to the left mage Fgure. Fgure 3: Average chage mage coherece vs. test mage ga offset K db. 9 - STO-E-SET-7-03

23 We evaluated the chage detecto performace ROCs for ths rage of ga offsets these ROCs are show Fgure 4. The curves Fgure 4 show that MLE chage detecto performace s superor to CCD chage detecto performace for PFA < ths s true over the rage of ga offsets smulated. It s of terest to vsuall compare the false alarms ad mssed targets whe a detcal umber of targets are detected b each algorthm. To ths ed, we have selected detecto thresholds for the CCD ad MLE algorthms, resultg a detcal umber of targets detected see operatg pots at PD 0.8 Fgure 4). The correspodg PFA values at these operatg pots are 0.05 vs for the CCD ad MLE algorthms, respectvel. The MLE PD/PFA curve used ths comparso s the 3 db ga offset case. Fgure 4: Coheret chage detecto ROCs for basele CCD vs. MLE; ote the CCD & MLE 3dB) operatg pots at pd 0.8. Fgure 5 below shows the CCD chage mage at the selected operatg pot. The detecto threshold used to obta PD 0.8 was I ths fgure, detected targets are overlad wth Crcles, false detectos are overlad wth Squares, ad mssed targets are overlad wth Damods. At PD 0.8 there are 6 mssed targets Damods) ad 73 false detectos Squares). ote that there are a total of 64 false detectos the shadow rego below the large buldg, whereas there are ol 9 false detectos the remag areas of the mage. I the et secto we wll cosder the problem of mtgatg false detectos SAR mage shadow regos ad we wll preset a smple, robust algorthm for performg ths task. STO-E-SET

24 Fgure 5: CCD chage mage, PD 0.8, PFA 0.05, mssed targets 6; threshold 0.488, detecto fll percetage 5%. Fgure 6 shows the correspodg MLE chage mage at the selected PD 0.8 operatg pot. For ths chage mage the detecto threshold used to obta PD 0.8 s ote that the detecto overlas o the mages llustrate that a dfferet set of targets were mssed b the CD algorthms, ad hece, each algorthm detects a somewhat dfferet set of the targets. It s also mportat to pot out that although both algorthms obta false detectos the shadow regos below the large buldg, the CCD verso of the algorthm obtas ma more false detectos these shadow regos. I the et secto we wll develop ad evaluate the performace of a smple approach to locate the shadow regos tree ad buldg shadows) ad remove false detectos from the SAR chage mage. Our algorthm for mtgatg false detectos these shadow regos wll, of course, mprove the performace of both algorthms. 9-4 STO-E-SET-7-03

25 Fgure 6: MLE chage mage, PD 0.8, PFA 0.04, mssed targets 6; threshold 0.73, detecto fll percetage 5%. Table 6 presets a detecto performace comparso of the CCD ad MLE algorthms before ad after the buldg shadow false alarms are ecluded from the false alarm FA) calculato. Ths comparso shows that a sgfcat mprovemet coheret chage detecto performace could be acheved b mtgatg false detectos shadow regos of SAR chage mager. STO-E-SET

26 Table 6: Coheret Chage Detecto statstcs, MLE vs. CCD at PD Wth Buldg FAs Wthout Buldg FAs PD # FAs PD # FAs MLE 7/33 0 7/33 0 CCD 7/ / Coheret Chage Detecto Performace wth Shadow Regos Masked The Mamum Lkelhood Estmate MLE) of the SAR coherece parameter was derved b Charles Jackowatz Referece []. I the prevous subsecto we showed that usg the MLE algorthm to estmate SAR chage mage coherece could gve better target detecto performace tha usg the comple correlato coeffcet, CCD. Furthermore, fgures 5 & 6 we observed that ma false detectos ca occur low-coherece buldg ad tree) shadows. Low-RCS areas such as buldg shadows ad tree shadows, as well as flat asphalt roads have low coherece due to radom phase returs. Sce a X-bad SAR ca t detect target returs from targets located ad masked shadow areas, the CCD performace ROCs could be sgfcatl mproved f such areas were masked as do t care areas before performg the detecto operato. Oe smple approach for mprovg the ROC curves s to post-process the stadard CCD ad MLE mages b settg the coherece of the areas that correspod to low-rcs areas both the test ad referece mages to ut. Target detecto s the performed o the chage mages wth low RCS areas masked. The low RCS areas ca easl be detected as follows. The -th pel the CCD mage s declared as belogg to the o-terestg, low RCS area f the average power of the coheret sum ad the coheret dfferece betwee the correspodg eghborhoods the test ad referece mages s below a selected threshold, T : k f k + gk + fk g k < k T where s the umber of pels the local eghborhood,, of the -th pel. f k ad pel of the referece ad test data the -th pel eghborhood,. g k are the k-th Ths subsecto apples the above algorthm for detectg low-rcs areas the SAR scee ad removg false detectos from these regos of the scee. The ga detecto performace acheved wll be quatfed va ROC curves. We frst show the SAR referece ad test mages used these studes Fgures 7 ad 8). We also show the groud truth overla ad the low-rcs mask fgures 9 ad 30). 9-6 STO-E-SET-7-03

27 Fgure 7: SAR referece mage. Fgure 8: SAR test mage. Fgure 9: Groud truth overla. Fgure 30: Eample low-rcs mage. Coheret chage detecto performace was determed usg the groud truth overla show above Fgure 9. Fgure 3 shows the orgal basele K )MLE chage mage ad Fgure 3 shows the MLE mage after applg shadow removal usg the low-rcs mage show Fgure 30. ote that after shadow removal, the average MLE chage mage coherece has creased from to STO-E-SET

28 Fgure 3: MLE, o shadow removal. Fgure 3: MLE, after shadow removal. The ROC curves Fgure 33 show that a large ga coheret chage detecto performace s acheved after the shadow areas the SAR mage are detected ad deoted as do t care areas of the scee. From the curves we see that after shadow removal, CCD performace s sgfcatl mproved relatve to the orgal CCD performace; however, the CCD algorthm performace after shadow removal s ot as good as the orgal MLE performace ad MLE performace after shadow removal s the best overall chage detecto performace result. Fgure 33: Coheret Chage Detecto ROCs true postve rate vs. false postve rate). 9-8 STO-E-SET-7-03

29 5. SUMMARY AD COCLUSIOS Ths paper revewed some of the basc propertes of SAR chage detecto. A eample was preseted comparg CCD ad CCD performace usg a scee cotag varous tpes of chages that occurred betwee the SAR referece mage ad test mage par. Aother eample was preseted comparg CCD performace usg phase-ol SAR mages versus usg the comple ampltude ad phase) SAR mages -- for the mages used ths stud we foud that the coherece levels obtaed from the comple mager were somewhat better tha the coherece levels obtaed from the phase-ol mager. A ew mage qualt metrc [5, 6], the "Uversal Image Qualt Ide" was vestgated. Applg ths mage qualt metrc to SAR test mages referece versus test test mages) we observed that the Uversal IQ Image was vsuall qute smlar to the SAR CCD coherece mage. We the vestgated the relatoshp betwee the coherece calculated usg comple mage data ad the correlato calculated usg the correspodg test mages. We demostrated that the magtude of the dffereces betwee the actual coherece mage ad the estmated coherece mage obtaed from the test-derved correlato mage were qute small, however, further aalss s eeded to quatf the errors versus the regstrato accurac requred betwee referece ad test mages. We also compared the coheret chage detecto performace obtaed usg the mamum lkelhood estmate of the coherece betwee the referece ad test comple SAR mages versus usg the comple correlato coeffcet estmate. We showed, usg a eample chage detecto mage par from the publcall released Gotcha SAR mager [0], that the MLE algorthm gave better PD/PFA detecto ROCs tha the CCD algorthm. Fall, we preseted a smple, robust algorthm for mtgato of false chages shadow areas buldg ad tree shadows), thereb mprovg the performace of both coheret chage detecto algorthms MLE ad CCD). STO-E-SET

30 6. REFERECES [] Jackowatz, C., et al., Spotlght-Mode Sthetc Aperture Radar: A Sgal Processg Approach, Sprger ew York, pp , 996. [] Press, M. ad. Stac, "Coheret Chage Detecto: Theoretcal Descrpto ad Epermetal Results," Itellgece, Survellace ad Recoassace Dvso, Defese Scece ad Techolog Orgasato, DSTO TR 85. [3] ovak, L., "Coheret Chage Detecto for Mult-polarzato SAR," Aslomar Coferece o Crcuts, Sstems, ad Computers, Pacfc Grove, CA, October, 005. [4] ovak, L., "Chage Detecto for Mult-polarzato, Mult-pass SAR," SPIE Coferece o Algorthms for Sthetc Aperture Radar Imager, Orlado, FL, March, 005. [5] Wag, Z. ad A. Bovc, A Uversal Image Qualt Ide, IEEE Sgal Processg Letters, March 00. [6] Wag, Z. ad A. Bovc, Mea Squared Error: Love It or Leave It, IEEE Sgal Processg Magaze, Jauar 009. [7] Guarer, A. ad C. Prat, SAR Iterferometr: A Quck ad Drt Coherece Estmator for Data Browsg, IEEE Tras. G.R.S., Ma 997 [8] R.Touz, et al, Coherece Estmato for SAR Imager, IEEE Tras. G.R.S., Ja.999. [9] Cha, M,. et al, Test Statstcs for Sthetc Aperture Radar Coheret Chage Detecto, IEEE Statstcal Sgal Processg Workshop, A Arbor, Mchga, 5-8 August, 0. [0] Scarborough, S., et al., "A Challege Problem for SAR Chage Detecto ad Data Compresso," SPIE Coferece o Algorthms for Sthetc Aperture Radar Imager, Orlado, FL, Aprl, STO-E-SET-7-03