BLOCK-REGRESSION-BASED FUSION OF OPTICAL AND SAR IMAGERY FOR FEATURES ENHANCEMENT

Transcription

1 BLOCK-REGRESSION-BASED FUSION OF OPTICAL AND SAR IMAGERY FOR FEATURES ENHANCEMENT Zhang Jxan, Yang Jnghu*, Zhao Zheng, Zhang Yonghong, L Hatao Chnese Academy of Surveyng and Mappng, Betapng Road 16, Bejng , P. R. Chna - Tel: Fax: jhyang@casm.ac.cn. * Correspondng author. KEYWORDS: SAR, Optcal Imagery, Block-Regresson-based Fuson, Block-SVR ABSTRACT: Ths paper focuses on fuson of optcal and Synthetc Aperture Radar (SAR) mages to combne these two types of remotely sensed magery for features enhancement. We have proposed a new fuson technque, namely Block-based Synthetc Varable Rato (Block-SVR), whch s based on block multple lnear regresson to fuse optcal and SAR magery. In order to nvestgate the effectveness, the fuson results of hgher resoluton arborne SAR mage and lower resoluton multspectral mage have been gven. Accordng to the fuson results, the fused mages have enhanced certan features, namely spatal and textural contents as well as features not vsble n multspectral whle preservng the color characterstcs. Also the spectral, spatal and textural effects of the presented algorthm were evaluated manly by vsual and quanttatve methods comparng to those of IHS, PCA and Wavelet-based methods. Durng the mplementaton of the block-regresson-based technques there are at least two advantages. One s that the block-regresson-based technque drastcally decreases amount of computaton, however regresson of the whole scene mage s almost mpossble. The other advantage, also the most mportant, s that adjustment of regressed-block sze can result n dfferent emphases between preservaton of spectral characterstc and enhancement of spatal and textural contents. The larger s the block of regresson, the more are the spatal and textural detals enhanced. In contrast, the smaller s the block of regresson, the more are spectral features preserved. 1. INTRODUCTION Data fuson s capable of ntegratng dfferent magery data to produce more nformaton than can be derved from a sngle sensor (Pohl and Genderen, 1998). Optcal and SAR magery are two mportant types of remotely sensed data. The optcal magery presents characterstcs dependng on the multspectral reflectance of the target llumnated by sunlght, whereas the SAR magery presents dfferent characterstcs from optcal magery due to the magng mechansm of actvely transmttng mcrowave pulses and recevng the backscattered echoes. Thus, n order to enhance some features lke spatal and textural features as well as features that are not vsble n optcal mages, data fuson of optcal and SAR magery s effectve measure. Another case where people need fuson of optcal and SAR magery s that n dffcult areas hgh-resoluton optcal magery s dffcult to acqure, whereas the SAR magery s easy to obtan because of ndependence of daylght and cloud cover. However, fuson of optcal and SAR magery s more dffcult than the ordnary pan-sharpenng because the grey values of SAR magery do not correlate wth those of multspectral magery. In table 1, correlaton coeffcents between multspectral bands and SAR band are almost zero. In contrast, the correlaton between panchromatc and multspectral bands of the same sensor s hgh, for nstance, SPOT panchromatc and XS1, XS2 bands (Shettgara, 1992, Pellemans, 1993). So the classcal methods assumng the hgh-resoluton band has a hgh correlaton wth the low-resoluton bands such as IHS, PCA methods (Chavez, 1991, Carper, 1990, Shettgara, 1992) are not sutable to fuse optcal and SAR magery. Recently Landsat ETM+ and SAR mage fuson based on generalzed ntensty modulaton (Alparone, et al, 2004) was llustrated where the pan-sharpened multspectral mage (15m resoluton) was modulated by textures of ERS mage (12.5m pxel sze). The dstncton of resoluton s so small that the enhancement of spatal features s neglgble. Table 1. Correlaton between multspectral bands and SAR band Band 1 Band 2 Band 3 Average SAR We propose the block-regresson-based technque for fuson of the optcal and SAR magery,.e. Block-SVR. The Block-SVR derves a synthetc block, as a lnear functon of blocks of multspectral mages, whch has maxmum correlaton wth the relevant block of SAR mage for every block of mages. The maxmum correlaton wth the block of SAR mage results n strong ablty of spectral preservaton whle spatal and texture features of SAR mage are mported nto the fused mage. The Block-SVR has two advantages that wll be dscussed n secton SVR FUSION METHOD Before presentng the Block-SVR, we concsely revew the SVR technque. The SVR method s a type of arthmetc combnatons fuson methods. The results of the SVR method of Munechka et al. (1993) show mprovements over those from the methods of Pradnes (1986) and Prce (1987). However, determnaton of the parameters s very complcated and only specfed classes are used n the calculaton. Y. Zhang (1999) had developed a new SVR ϕ method that can drectly calculate parameters. In order to merge SPOT panchromatc mage (10m resoluton) wth Landsat TM mage (30m resoluton), the paper gave a set of equatons: ϕ

2 XSP = (1) (2) HSyn H H XS HSyn = ϕ XH = ϕ XH (3) Where XSP s the grey value of the th band of the merged hgh-resoluton TM-SPOT mage, H s the grey value of XS the orgnal SPOT PAN mage, and s the grey value of the th band of the TM mage modfed to have the same pxel sze as the orgnal SPOT panchromatc H mage. HSyn s the grey value of the hgh-resoluton synthetc panchromatc mage smulated through equaton 2 where these ϕ parameters are calculated drectly through multple regresson analyss of the orgnal panchromatc mage H XS and the orgnal multspectral bands formulated n equaton 3. Accordng to the results of small test area (600 x 380, 10m), the spectral effect of the new SVR s good n comparson to those of IHS and PCA methods (Carper et al. 1990, Chavez et al. 1991, and Shettgara 1992). 3. BLOCK-SVR FUSION METHOD The man dea of Block-SVR s multple lnear regresson of block regon between SAR and optcal mages nstead of multple regresson of whole mage n the SVR method. The block-based tactcs at least have two merts. Frstly, the multple lnear regresson of block drastcally decreases the requrements of computaton. Secondly, the adjustment of block sze of regresson can result n dfferent emphases between preservaton of spectral characterstcs and enhancement of spatal and textural contents. In other words, the larger s the block of regresson, the more are the spatal and textural contents enhanced. In contrast, the smaller s the block of regresson, the more are the spectral features preserved. The algorthm procedure of the Block-SVR s shown n fg. 1 and the explanatons of steps are as follows: 1) The bands of multspectral mage are re-sampled to have the same sze of arborne SAR mage. Before the resample processng, the multspectral mage and arborne SAR mage must be co-regstered. 2) A block and ts neghbor are fetched from SAR mage and re-sampled multspectral bands n sequence. 3) Lnear multple parameters are calculated accordng equaton 3 drectly through multple lnear regresson between the 3 x 3 blocks fetched from SAR mage and these 3 x 3 blocks fetched from multspectral bands. 4) The central block of 3 x 3 blocks s selected for the SAR mage as well as these multspectral bands. 5) Accordng to the coeffcents of multple lnear regresson, the synthetc block s produced by equaton 2. 6) The SAR block, three multspectral blocks and synthetc block are used to do rato-calculaton by equaton 1. After these steps the fused blocks are produced. 7) The result blocks after step 6 are added to the fused bands. 8) Repeatng steps (2) ~ (7) untl the last block. After above steps, three fused bands are generated. From these steps we can regard SVR algorthm as a specal case of Block-SVR n whch the block sze of regresson s the sze of whole mage to be merged. In Secton 4 the spectral and spatal effects wll be assessed through experment results and comparson to those of IHS, PCA and wavelet-based methods. Also, the trats and flexblty of the method wll be analyzed n secton TEST RESULTS AND DISCUSSIONS The above technque was tested on an AeS-1 arborne SAR mage of area around the new far n Munch, German and ts relevant arborne multspectral mage. The wavelength of the arborne SAR mage s X-Band and the resoluton of the mage s 1.5m. The multspectral mage was degraded to about 3.0m resoluton. Thus, the vsual and quanttatve assessments wll be used to valdate the effectveness of features enhancement compared to those of IHS (Chavez, 1991 and Carper, 1990), PCA (Chavez, 1991, Shettgara, 1992) and Wavelet-based method (Yocky, 1996). Because of the purpose of features enhancement, our attentons wll concentrate on the spatal, textural characterstcs and features not vsble n multspectral mages. 4.1 Vsual Comparson The apparent resoluton of Plate 1e and Plate 1f (Block-SVR fused results) s mproved whch s close to those of Plate 1c and Plate 1d (the arborne SAR mage). Note that the textural characterstcs avalable n the SAR mage are mported nto the fused mage and the color characterstcs of the multspectral mage are preserved bascally. An especal attenton should be pad s that the features, whch are not vsble n the orgnal multspectral mage, are enhanced by SAR mage. In a word, through the fuson of SAR mage and multspectral mage, some features lke spatal and textural features can be enhanced n multspectral mage, and some features that are not avalable n multspectral mage can be mported to the fused mages. In Plate 1 the enhancement results of the multspectral true color mage are shown. The slces of the fused mage of Block-SVR n Plate 1e and Plate 1f are much closer to the orgnal multspectral n Plate 1a and Plate 1b than those of PCA n Plate 1g and Plate 1h and IHS n Plate 1 and Plate 1j n color characterstcs. The color characterstcs of wavelet-based method n Plate 1k and Plate 1l are a lttle better than those of Block-SVR. However, the spatal and textural nformaton mported from SAR mage n Plate 1k and Plate 1l s far lesser than those of Block-SVR, especally the features n plate 1k whch not avalable n orgnal multspectral mage n plate 1a s obscure whereas that of Block-SVR n Plate 1e s relatvely clear. The spatal and textural features of Block-SVR and PCA whch are mported from SAR mage are almost the same, and are much better than those obtaned from IHS and Wavelet-based methods.consderng these above fuson methods we fnd that the Block-SVR can acheve a satsfyng tradeoff between preservaton of spectral characterstcs and enhancement of spatal and textural contents. 4.2 Quanttatve Assessments In order to evaluate the effectveness of the proposed technque quanttatvely, correlaton coeffcents are calculated. The correlaton coeffcent denotes the smlarty between structures (Wald and Ranchn, 1997). Hence, the correlaton coeffcents between the multspectral and the relevant fused mage can ndcate the ablty of spectral preservaton to some extent; the hgher are the correlaton coeffcents, the more are the spectral characterstc preserved. In table 2 t can clearly be seen that

3 those of the Block-SVR are more hgher than PCA, IHS and SVR, whch are slghtly lower than wavelet-based method, so the color characterstcs n the fused mage usng Block-SVR are much closer to the orgnal multspectral than those of PCA, IHS and SVR method. The results are also valdated by vsual comparson. For the case of wavelet-based method, the strong ablty of spectral preservaton has been gotten on the condton that the fused mage largely lacks the spatal and textural nformaton from SAR mage. The correlaton coeffcents between SAR mage and fused mage ndcated n table 3 are also calculated for Block-SVR, PCA, IHS, Wavelet-based and SVR method. The correlaton values of Block-SVR can be accepted bascally whose average value s The values n table 3 explctly reveal the fact that spectral preservaton and enhancement of spatal and textural nformaton s a par of contradctons and the best fuson technque s to acheve a satsfyng tradeoff between spectral preservaton and enhancement of spatal and textural nformaton wth respect to a specfc applcaton. Table 2. Correlaton between the multspectral and the relevant fused mages usng Block-SVR, PCA, IHS, Wavelet-based and SVR fuson algorthms B-SVR PCA IHS Wavelet SVR Band Band Band Average Algorthm Dscussons Durng the mplementaton of the block-regresson-based technques there are at least two advantages. One advantage, also the most mportant, s that adjustment of regresson block sze can results n dfferent emphases between spectral preservaton and enhancement of spatal and textural nformaton. The correlaton coeffcents gven n table 4 show that the smaller the block sze, the hgher correlaton between the multspectral and the fused mages. Thus, the concluson can be easly drawn that the smaller s the block of regresson, the closer to those of the orgnal multspectral mage are the color characterstcs of the fused mage. Ths s due to the fact that the regresson result (Synthetc Varable n Munechka, 1993 and Y. Zhang, 1999) of small block of multspectral bands s much closer to the SAR mage block than those of large block. In contrast, the larger s the block of regresson, the more are spatal and textural features mported nto the fused mage from the SAR mage, whch can bascally be llustrated by the correlaton between SAR and fused mages ndcated n table 5. The other advantage s effcency of mplementaton of the technque. The block-regresson-based technque drastcally decreases requrements of computaton, whereas regresson of the whole scene mage s almost mpossble. The progress of ths technque makes possble n real applcatons because of ts effcency of mplementaton. Table 3. Correlaton between SAR and fused mages usng Block-SVR, PCA, IHS, Wavelet-based and SVR fuson algorthms B-SVR PCA IHS Wavelet SVR Band Band Band Average CONCLUSIONS Block-regresson-based fuson technque has proven tself avalable and flexble for fuson of optcal and SAR magery. The fuson results show that the fused mages have enhanced certan features, namely spatal and textural contents, as well as features not vsble n multspectral mage whle preservng the color characterstcs. Also, the spectral, spatal and textural effects of the presented algorthm were evaluated manly by vsual and quanttatve methods comparng to those of IHS, PCA and Wavelet-based methods. The assessments ndcate that the block-regresson-based method s the best technque among these methods, because t can acheve a satsfyng tradeoff between preservaton of spectral characterstcs and enhancement of spatal and textural contents wth respect to a specfc applcaton.

4 1 Resamplng Fetchng a block and ts 3x3 neghbors 2 2 Fetchng a block and ts 3x3 neghbors 3 Multple lnear regresson 4 Selectng the central block Lnear regresson coeffcents 5 Producng Synthetc block Selectng the central block 4 6 Rato calculatng 7 Addng to the fused bands No 8 The last block? No 9 Yes Endng Fg. 1 The procedure of Block-SVR (a) (b) Plate 1. Vsual comparson of dfferent fuson results (a), (b) two slces of the orgnal multspectral mage. (c), (d) two slces of the AeS-1 arborne SAR mage (e), (f) fuson results of Block-SVR method. (g), (h) fuson results of PCA method. (), (j) fuson results of IHS method. (k), (l) fuson results of wavelet-based method. (contnued on next page)

5 Plate 1. contnued. (c) (d) (e) (f) (g) (h) () (j) (k) (l)

6 Table 4. Correlaton between the multspectral and the relevant fused mages usng Block-SVR fuson algorthms of dfferent block sze 8 x 8 16 x x x x x x 512 Band Band Band Average Table 5. Correlaton between SAR and fused mages usng Block-SVR fuson algorthms of dfferent block sze 8 x 8 16 x x x x x x 512 Band Band Band Average References A.H.J.M. Pellemans, R.W.L. Jordans, and R.Allewjn, Merge Multspectral and chromatc SPOT Images wth respect to the Radometrc Propertes of Sensor. Photogramm. Eng. Remote Sens., vol. 59, No.1 pp B. Garguet-Duport, J. Grel, J. M. Chasseny, and G. Pautou, The use of multresoluton analyss and wavelet transform for mergng SPOT panchromatc and multspectral mage data. Photogramm. Eng. Remote Sens., vol. 62, pp C. Pohl, J.L. Van Genderen, Multsensor mage fuson n remote sensng: concepts, methods, and applcatons. Internatonal Journal of Remote Sensng, vol.19, pp.823 ~ 854. C.A. Yocky, Multresoluton wavelet decomposton mage merger of Landsat Thematc Mapper and SPOT chromatc data. Photogramm. Eng. Remote Sens., vol. 62, pp J. W. Carper, T. M. Lllesand, and R. W. Kefer, The use of ntensty hue saturaton transformatons for mergng SPOT panchromatc and multspectral mage data. Photogramm. Eng. Remote Sens., vol.56, pp magery from dual resoluton satellte nstruments. Remote Sensng of Envronment, vol. 21, pp.119 ~ 128. V. K. Shettgara, A generalzed component substtuton technque for spatal enhacement of multspectral mages usng a hgher resoluton data set. Photogramm. Eng. Remote Sens., vol. 58, pp Y. Zhang, A new mergng method and ts spectral and spatal effects. Internatonal Journal of Remote Sensng, vol. 20, pp.2003 ~ Acknowledgements It s wth great apprecaton that we thank Prof. Anke Bellmann, Department of Computer Vson and Remote Sensng, Techncal Unversty of Berln. Prof. Anke Bellmann provded us the AeS-1 arborne SAR mage and the relevant optcal mage for the purpose of EuroSDR sensor and data fuson contest Informaton for mappng from SAR and optcal mage data. Ths paper s part of work of the project Research on Mult-sensor Remotely Sensed Data Fuson supported by Key Laboratory of Geo-nformatcs of State Bureau of Surveyng and Mappng, P. R. Chna. Lucano Alparone, Stefano Baront, Andrea Garzell and Flppo Nencn, Landsat ETM+ and SAR Image Fuson Based on Generalzed Intensty Modulaton. IEEE Transactons on Geoscence and Remote Sensng, Vol. 42, No. 12, pp L. Wald, T. Ranchn, and M. Mangoln, Fuson of satellte mages of dfferent spatal resolutons: Assessng the qualty of resultng mages. Photogramm. Eng. Remote Sens., vol. 63, no. 6, pp Munechka, C. K., Warnck, J. S., Salvaggo, C., and Schott, J. R., Resoluton enhancement of multspectral mage data to mprove classfcaton accuracy. Photogrammetrc Engneerng and Remote Sensng, vol. 59, pp.67 ~ 72. Pradnes, D., Improvng SPOT mages sze and multspectral resoluton. Proceedngs of the SPIE Earth Remote Sensng usng the L andsat T hematc Mapper and SPOT Systems, 660, pp.98 ~ 102. Prce, J. C., Combnng panchromatc and multspectral