CHANGE DETECTION FOR AERIAL PHOTO DATABASE UPDATE Yongdae Gweon* and Yun Zhang Dept. of Geodesy and Geomatcs Engneerng, Unversty of New Brunswck, P. O. Box 4400, Fredercton, NB, Canada, E3B 5A3 x93pn@unb.ca and yunzhang@unb.ca Commsson VII, WG VII/5 KEY WORDS: Remote Sensng, Change Detecton, Classfcaton, Land Cover, Multsensor ABSTRACT: Wth rapd new technologcal development durng the past 50 years, aeral photos have been ncreasngly and commonly used not only n geographc nformaton systems but also n varous spatally related applcatons. Many muncpaltes and government agences have constructed aeral photo databases all over the world. Keepng these databases up to date s the most mportant part of makng them effectve so that aeral photo databases are expected to be updated as frequently as possble. However, n practce, some of them are barely updated because of hgh cost. In case of detectng where changes occur, aeral photo databases could be frequently and partally updated only n the changed areas.in ths study, medum spatal resoluton magery s proposed to detect changes. Instead of usng aeral photos, medum spatal resoluton magery could be a good alternatve because of ts wder coverage and lower prce than aeral photos. Free accessble Landsat ETM+ and orthophotomaps are used for change detecton. In order to compare wth 15 m pan-sharpened Landsat ETM+, 1 m orthophotomaps are decomposed, segmented, and classfed through wavelet transform and obect-orented classfcaton. Although the detected changes are rough, the result shows that the method s qute costeffectve and practcal. Moreover, t could support decson makng for updatng aeral photo databases. 1. INTRODUCTIOIN Aeral photos have been proved to be useful for extractng spatal nformaton or performng spatal analyss snce they have been wdely used for mappng purposes. In addton, dverse aeral photo products such as topographc maps, dgtal elevaton models, and orthophotos play an mportant role n varous spatally related applcatons. These are reasons why many muncpaltes and government agences have constructed aeral photo databases n Canada and also n other natons. In Canada, for example, n federal case, the Natonal Arphoto Lbrary (NAPL) by Natural Resources Canada archves over sx mllon aeral photos coverng all of Canada. In provncal case, the Alberta Sustanable Resource Development collects over 1.4 mllon aeral photos coverng whole provnce, and the Provncal Softcopy Orthophotomap Database (SODB) by the Servce New Brunswck (SNB) was created from aeral photos for the whole provnce. Consderable amounts of aeral photo databases have been bult all over the country. In general, most databases have an updatng problem and t has been a contnuous ssue n ths feld. Aeral photo database s no excepton, and up-to-date data s ndeed necessary to extract and analyze trustworthy spatal nformaton. In practce, the SNB has receved numerous nqures nto when the SODB wll be updated. Although most aeral photo databases are expected to be updated as frequently as possble, some are barely updated because of hgh cost. Human actvtes and natural phenomena would cause landcover changes, and n practce they are not occurred n whole area. Hence, n case of cost-effectvely detectng where changes occur, aeral photo databases could be frequently and partally updated only n the changed areas. Multtemporal mages from same sensor are generally used n change detecton, but t s too expensve to use aeral photos. Medum spatal resoluton magery, such as Landsat and SPOT usually has wder coverage and lower prce than aeral photos, so t could be a good alternatve for change detecton. In ths study, a cost-effectve and practcal method s suggested to detect land-cover changes to support decson makng for updatng aeral photo databases. It nvolves wavelet transform and obect-orented classfcaton to compare two dfferent types of mages, and then post-classfcaton comparson s appled to fnd changes. 2.1 Wavelet Transforms 2. BACKGROUND Wavelets bascally ntended to address shortcomngs of the short tme fourer transform (STFT). Instead of usng the fxed tme and frequency resoluton n the STFT, wavelet transforms are able to provde the tme and frequency representaton of the sgnal. In order to use the dea of multresoluton, wavelet transforms can be defned as a scalng functon and a wavelet functon. These functons composed of translatons and scalngs of a scalng functon and a wavelet functon are: / 2 ϕ ( ) 2 (2 ) (1), k t = ϕ t k / 2 ψ ( t) = 2 ψ (2 t ) (2), k k * Correspondng author 927
The Internatonal Archves of the Photogrammetry, Remote Sensng and Spatal Informaton Scences. Vol. XXXVII. Part B7. Beng 2008 where ϕ ( ) and ψ ( ) are the scalng functon and the, k t, k t wavelet functon respectvely (Burrus et al, 1998). Accordng to the basc requrement of multresoluton analyss (MRA), the scalng functon space, can be formulated: V a factor of two at each level. These processes are called the flter bank llustrated n Fg. 1. wth V V +1 for all Z (3) V = {0}, V = { L 2 ( R)} (4) Orgnal Image LL 2 LH 2 LL 1 LH 1 LH 1 HL 2 HH 2 HL 1 HH 1 HL 1 HH 1 2 where Z s the set of ntegers and L ( R) s the space of all functons f (t) wth a well defned ntegral of the square of the modulus of the functon. Then, ϕ (t) can be smply expressed: ϕ ( t ) = h ( n) 2ϕ (2t n) n ϕ where h ϕ (n) s the scalng functon coeffcents. There are some advantages that a scalng functon and a wavelet functon are orthogonal. Therefore, the related wavelet functon space, W s the orthogonal complement of n. Then, V V + 1 the wavelet functon space can be defned: V +1 = V W n general 2 { L ( R)} = V W W +1 Λ (5) (6) (7) Fgure 2. Two-level two-dmensonal mage decomposton By applyng the DWT flter bank to two-dmensonal mages, an mage s actually decomposed nto four mages at each decomposed level. It frst apples to the columns and then to the rows. Consequently, t produces one approxmaton coeffcent (LL) and three detal coeffcents (LH: horzontal, HL: vertcal, HH: dagonal). To obtan the next level of detal coeffcents, the DWT flter bank apples the LL alone, and t results n two level decompostons as shown n Fg. 2. 2.2 Obect-orented Classfcaton A pxel s the basc processng unt n conventonal classfcaton whereas an obect n obect-orented classfcaton. The term, obects, refers to ndvdually resolvable enttes located wthn an mage that are perceptually generated from pxel groups (Hay et al., 1997). Obects have been composed of a group of pxels parttoned by mage segmentaton, and then dverse classfers (e.g., Maxmum Lkelhood, ISODATA, Neural Network, Fuzzy Logc and combnaton of these) operate on obects ndvdually. where s the drect sum. Then, ψ (t) can be expressed: where ψ ( t ) = h ( n) 2ϕ (2t n) (n) h ψ f (n) h (n) n ψ s the wavelet functon coeffcents. 2 level 1 detal (8) Dscontnuty and smlarty are two basc propertes of mage segmentaton algorthms. Rapd changes n ntensty called dscontnuty could be used n the edge-based segmentaton. It s based on nformaton about edges n the mage and dverse edge detectng operators are utlzed. Accordng to the smlarty specfed by a set of predefned crtera, an mage can be parttoned nto obects or regons that are smlar. It s used n thresholdng and regon-based segmentaton. By applyng threshold to an mage hstogram, thresholdng can smply dvde an mage. However, regon-based segmentaton, such as regon growng and regon splt-mergng approach, can drectly construct regons usng predcated measurements to satsfy the followng condtons: l (n) 2 h(n) l(n) 2 2 level 2 detal level 2 approxmaton Fgure 1. Two-level one-dmensonal DWT flter bank Whle our nterest s n dscrete mages, dscrete wavelet transforms (DWT) s commonly used to mplement wavelet transforms. It s easy to apply and reduce a computatonal burden. The process of DWT can be represented as two operatons; flterng and down-samplng. The flterng mplements a low-pass flter extractng the approxmate coeffcents, and a hgh-pass flter extractng the detal coeffcents. Then, the down-samplng reduces the resoluton by S R =ΥR R Ι = Ø =1 R (9) H ( R ) = TRUE = 1,2, Κ, S (10) H ( R Υ R ) FALSE R adacent to R (11) = where S s the total number of regons n an mage, and H ( R s a bnary smlarty (or homogenety) evaluaton of ) the regon (Sonka et al., 1999). R 928
The Internatonal Archves of the Photogrammetry, Remote Sensng and Spatal Informaton Scences. Vol. XXXVII. Part B7. Beng 2008 3.1 Image Data 3. METHODOLOGY Two orthophotomaps from SODB and Landsat ETM+ from GeoBase are used n ths study. The SODB s a hgh qualty dgtal map product whch conssts of orthometrcally corrected aeral photos of New Brunswck. It s a scanned, orthorectfed and mosacked aeral photo clpped n a 1:10,000 topographc map wndow. It has three bands (RGB) and 1 m spatal resoluton. It was acqured n the summer of 1996. Landsat ETM+ was acqured on September 2000. It has one 15 m panchromatc, sx 30 m multspectral bands. It s dstrbuted from GeoBase whch s a free accessble geospatal database for all of Canada. These mages are shown n Fg. 3. 3.2.1 Image Fuson: Image fuson combnes hgh spatal resoluton of panchromatc mages and hgh spectral resoluton of multspectral mages. The spatal resoluton dfference between orthophotomaps and Landsat ETM+ s qute hgh. In order to enhance the mage and reduce the resoluton dfference, the panchromatc and multspectral mages of Landsat ETM+ can be fused by an automatc fuson method called UNB Pan- Sharpenng. The method based on least squares s employed for a best approxmaton of the gray value relatonshp between the orgnal multspectral, panchromatc and the fused mage bands for a best color representaton (Cheng et al., 2003). 3.2.2 Wavelet Transform: In change detecton, t s mportant to use the same sensor, same radometrc and spatal resoluton data wth annversary acquston dates (Lu et al., 2004). Although mage fuson reduced the spatal resoluton dfference, the dscrepancy between orthophotomaps and pansharpened Landsat ETM+ s stll subect. Four levels wavelet transform decompose orthophotomaps nto one approxmaton and three detal coeffcents from 1 m to down-sampled 16 m spatal resoluton. Proper spatal resoluton, therefore, can be appled to mplement convncble change detecton. Fgure 3. Image data: top: orthophotomap(sodb no. 46006480), bottom: correspondng subset of Landsat ETM+, top: orthophotomap(sodb no. 46106480), bottom: correspondng subset of Landsat ETM+ 3.2 Methodology The steps assocated n ths study are summarzed n Fg. 4. (c) Orthophotomaps RGB Wavelet Transform (Approxmaton & Resoluto n matchng Landsat ETM+ Pan & Mult Imagery Image Fuson (Pan & Mult) Obect-Orented Classfcaton Obect-Orented Classfcaton Result Result (d) (e) Change Detecton Fgure 4. Overvew of methodology Fgure 5. Wavelet transform: 4 th level coeffcents of Red band of orthophotomap n 16 m spatal resoluton approxmaton, horzontal detal, (c) vertcal detal, (d) dagonal detal and (e) sum of detal coeffcents 929
The Internatonal Archves of the Photogrammetry, Remote Sensng and Spatal Informaton Scences. Vol. XXXVII. Part B7. Beng 2008 After a few levels of wavelet decomposton, the approxmaton coeffcent loses ts hgh frequency component because of a low-pass flter. The low-pass flter not only removes nose n an mage but suppresses overall varaton n brghtness. It means the brghtness of the shadow n forest regons gradually decreases as close to the brghtness n water regons. Therefore, t s qute dffcult to dstngush whch regon s forest or water. By addng texture propertes to a classfcaton process, the drawback of low-pass flterng could be compensated. Whle there s no formal defnton of texture, the texture of an mage s related to the spatal dstrbuton of the ntensty values n the mage, and as such contans nformaton regardng contrast, unformty, rugosty, regularty, etc (Ruz et al., 2004). For the purpose of text descrpton, detal coeffcents were used n ths study because they contan the hgh frequency component whch s the relevant texture nformaton. The summaton of each detal coeffcents was utlzed n followng classfcaton process, and decomposed 4 th level coeffcents are shown n Fg. 5. As many segmentaton algorthms rely on user assgned parameters, the mult-resoluton segmentaton also requres defnng parameters, such as scale, color/shape, and smoothness/compactness. It means the sze and shape of mage obects could be affected by dfferent values of parameters as shown n Fg. 6. In addton, t s tme-consumng and tedous to fnd the most sutable sze and shape of obects for classfcaton. For example, the approprate scale value s 40 to solate a forest regon, and t s revealed through many trals as shown n Fg. 6 (d). In order to avod ths tresome process, mage obects were extracted usng default values of parameters wthout any consderaton of sze and shape. Then, based on the small but meanngful mage obects (see Fg. 7) as the basc process unt, spectral statstcs and texture descrptons were calculated for classfcaton. 3.2.3 Obect-orented Classfcaton: Orthophotomaps and Landsat ETM+ have dfferent spatal resolutons, number of bands, and wavelengths. Although the problem of dfferent spatal resoluton was resolved through the above steps, some others were stll remaned. Therefore, the post-classfcaton comparson was appled to change detecton nstead of the mage-to-mage comparson. The classfcaton used n ths study s based on a new obectorented approach. As contrasted wth conventonal pxel-based classfcaton, obect-orented classfcaton can expect the semantc relaton between real-world obects and mage obects. Ths relaton mproves the value of the fnal classfcaton (Benz et al., 2004). Meanngful mage obects were extracted usng the mult-resoluton segmentaton algorthm. It s one of the basc procedures of the commercal remote sensng mage processng software Defnens Professonal 5.0 whch s formally known as ecognton. Fgure 7. the result of mult-resoluton segmentaton usng default values the basc process unt The mode value was used to represent each spectral value of obects, and the standard devaton of detal coeffcent was used to descrbe the texture nformaton of obects. In general, most text descrptors, such as the gray-level co-occurrence matrx (GLCM) and the local bnary pattern (LBP) are calculated usng a small wndow. However, the standard devaton was used to descrbe the texture nformaton n ths study because the shape of mage obects s usually rregular (see Fg. 6 ) and wndow-based texture descrptors nclude pxels on the outsde of the mage obects n ther calculaton. In addton, the standard devaton s easy to calculate and smple to represent coarse or fne. Combned classfcaton was employed after segmentaton. Frst, supervsed classfcaton was performed usng backpropagaton neural network algorthm, and then, unsupervsed classfcaton was performed usng Kohonen s self-organzng feature map (SOFM) algorthm. 4. RESULTS AND DISCUSSION Landsat ETM+ was georeferenced, fused and resampled to 16 m spatal resoluton, and then orthophotomaps were decomposed usng 4 th level wavelet transform to compare respectvely. The mult-resoluton segmentaton parttoned both mages nto meanngful mage obects, and then the combned classfcaton was appled to detect changes. (c) (d) Fgure 6. Results of mult-resoluton segmentaton usng dfferent scale parameters (a=5, b=10, c=20, d=40) Although the problem caused by dfferent spatal resoluton was solved, some extra problems appeared n classfcaton because of the propertes of orthophotomaps and wavelet transform. Frst, avalable spectral nformaton of orthophotomaps s lmted wthn only vsble bands. It affects not only an accuracy of classfcaton but also a dscrmnaton of complex 930
The Internatonal Archves of the Photogrammetry, Remote Sensng and Spatal Informaton Scences. Vol. XXXVII. Part B7. Beng 2008 nature. Especally t s qute dffcult to dstngush grass, mpervous and crop regons. Second, the gray level range of mages s decreasng after few wavelet transforms. It means overall brghtness of mages s gettng darker and the spectral range of each land-cover type s gradually closer. It often msclassfes some land-cover types, especally the shadow n forest and water. Fnally, the classfcaton results from two dfferent mages vary because of dfferent characterstcs. These problems practcally derved the classfcaton error, and drectly affected the classfcaton accuracy. The am of ths study was to detect changes n coarse scale, and t was not necessary to dstngush very detal land-cover types. Therefore, the four land-cover category, such as water, mpervous, forest and non-forest, was used n classfcaton. The classfcaton was done n two steps. In the frst step, the supervsed classfcaton was performed usng tranng stes. Then, the unsupervsed classfcaton was performed usng 30 classes. After the unsupervsed classfcaton was completed, the classes were re-categorzed and combned wth the supervsed classfcaton result. Fnally, map-to-map comparson was appled to detect changes, as shown n Fg. 8. 5. CONCLUSION Water Impervous Forest Non-forest In ths study, mages from dfferent sensors have been utlzed for change detecton. In order to apply map-to-map comparson, the new obect-orented classfcaton dscrmnates four landcover types from orthophotomaps and Landsat ETM+ after wavelet transform and mult-resoluton segmentaton. Through the mplementaton, the methodology suggests how to compare hgh spatal resoluton magery wth medum spatal resoluton magery and how to mprove the classfcaton result when the avalable spectral nformaton s lmted. (c) (d) Instead of usng aeral photos, medum spatal resoluton magery, such as Landsat and SPOT can be a cost-effectve and practcal alternatve n change detecton. Although the results n ths study could not dentfy changes n detal, t s enough to represent where changes occur and suggest where should be updated roughly. Moreover, t could be appled to dfferent types of magery. REFERENCES References from Journals: Benz, U. C., Hofmann, P., Wllhauck, G., Lngenfelder, I., and Heynen, M., 2004. Mult-resoluton, obect-orented fuzzy analyss of remote sensng data for GIS-ready nformaton. ISPRS Journal of Photogrammetry & Remote Sensng, 58, pp. 239-258. Water Impervous Forest Non-forest (e) (f) Hay, G. J., Nemann, K. O., and Goodenough, D. G., 1997. Spatal Thresholds, Image-Obects, and Upscalng: A Multscale Evaluaton. Remote Sensng of Envronment, 62(1), pp. 1-19. Lu, D., Mausel, P., Brondxo, E. and Moran, E., 2004. Change detecton technques. Internatonal Journal of Remote Sensng, 25(12), pp. 2365-2407. References from Books: Burrus, C. S., Gopnath, R. A. and Guo, H., 1998. Introducton to Wavelets and Wavelet Transforms. Prentce-Hall, Inc., Upper Saddle Rver, NJ, pp. 349-402. (g) (h) Fgure 8. Change detecton: classfcaton of 46006480, classfcaton of correspondng Landsat ETM+, (c) changes between and, (d) refned changes, (e) classfcaton of 46106480, (f) classfcaton of correspondng Landsat ETM+, (g) changes between (e) and (f), (h) refned changes Sonka, M., Hlavac, V. and Boyle, R., 1999. Image Processng, Analyss, and Machne Vson: second edton. Brook/Cole Publshng Company, Pacfc Grove, CA pp. 123-227. References from Other Lterature: Ruz, L. A., Fdez-Sarra, A. and Reco, J. A., 2004. Texture Feature Extracton for Classfcaton of Remote Sensng Data 931
The Internatonal Archves of the Photogrammetry, Remote Sensng and Spatal Informaton Scences. Vol. XXXVII. Part B7. Beng 2008 usng Wavelet Decomposton: a Comparatve Study. XX th ISPRS Congress, Istanbul, Turkey, Commson 4, pp. 1109. References from webstes: Cheng, P., Toutn, T., and Zhang, Y., 2003. PCI Geomatcs Techncal Paper QuckBrd-Geometrc Correcton, Data Fuson, and Automatc DEM Extracton, Canada. http://www.pcgeomatcs.com/servces/support_center/tech_pap ers/acrs03_cheng.pdf (accessed 1 Apr. 2008) ACKNOWLEDGEMENTS Ths study has been supported by the New Brunswck Innovaton Foundaton (NBIF). Orthophotomaps was acqured from SNB, and Landsat ETM+ was acqured from GeoBase. I wsh to thank my supervsor, Dr. Yun Zhang and I also thank Mr. Ben Wuest for hs assstance. 932