Unsupervised Change Detection From Multichannel ... - IEEE Xplore

280 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 4, NO. 2, APRIL 2007and pixel labels (q, t =0, 1, 2,...; the index q is dropped forease of notation), the following operations are performed [15].1) Compute for each kth pixel the current estimate ofP {l k = H i |v k , Ck t ,θt } (k =1, 2,...,N; i =0, 1) using(2) and (3).2) Update the label of each kth pixel according to the MRFminimum-energy rule [12] by setting l t+1kas the label H ithat corresponds to the lowest value of U(H i |v k , Ck t ,θt )(k =1, 2,...,N; i =0, 1).3) Update the parameter estimates as follows:⎧⎪⎨⎪⎩∑ Nκ t+1 k=11i =wt ik ln v k∑ Nk=1 wt ik∑ Nκ t+1 k=12i =wt ik (ln v k−κ t+1 ) 21i∑ Nk=1 wt ikλ t+1 =arg maxλ>0, i=0, 1, i=0, 1[]N∑ ∑λ 1 wik t mt ik − ln ∑ 1 exp(λm t ik )k=1i=0i=0(4)Fig. 1. XSAR channels acquired on (a) April 16, 1994 and (b) April 18, 1994after histogram equalization.where wik t = P {l k = H i |v k , Ck t ,θt } if the kth pixel hasbeen assigned to H i in step 2) and wik t =0 otherwise(k =1, 2,...,N; i =0, 1).The maximization problem to compute λ t+1 in step 3) is solvedby the Newton–Raphson method [17]. LJ-EM is a modifiedversion of the EM algorithm. We recall that EM is known toconverge, under mild assumptions, to a local maximum of thelog-likelihood function (even though convergence to a globalmaximum is not ensured, usually a good solution is obtained)[18]; a similar convergent behavior is expected for LJ-EMas well [10]. LJ-EM is iterated until the differences betweenthe parameter estimates at successive iterations go below apredefined threshold (here, equal to 0.001). At each iteration,the parameters of H i are updated only according to the samplesassigned to H i (i =0, 1), which aims to reduce the overlappingbetween H 0 and H 1 .D. Initialization by the GKIT AlgorithmThe initial change map M 0 is generated on a single-channelbasis. The LN-GKIT technique [4] is adopted here; it is stillbased on a log-normal conditional model for SAR ratio data.LN-GKIT automatically computes the optimal threshold to beapplied to a single-channel ratio image in order to distinguish“change” and “no-change,” by minimizing a “criterion function”defined according to a Bayesian approach and related tothe probability of error [4]. Let J h (·) be the criterion functioncomputed by LN-GKIT when applied to the hth channel inR, τh ∗ = arg min τ J h (τ) the corresponding optimal threshold,and M ∗ hthe resulting change map (h =1, 2,...,n) [4]. M0is defined as the map M ∗ hcorresponding to the lowest optimalvalue J h (τh ∗ ) of the criterion function (h =1, 2,...,n).According to the relationship between the criterion functionand the probability of error [4], [19], this change map isexpected to be, at least, a suboptimal choice among the singlechannelmaps.Fig. 2. Semisimulated data set. (a) Test map, XSAR channels for the(b) April 16, 1994 and (c) April 18, 1994 dates (after equalization), and changemaps generated (d) by the proposed method and by a noncontextual variant ofthe method applied (e) to nonfiltered and (f) to speckle-filtered images. Legendfor the change maps: White = “change,” gray = “no-change.”III. EXPERIMENTAL RESULTSTwo 700 × 280 pixel sized coregistered multipolarizationand multifrequency SAR images, acquired over an agriculturalarea near the city of Pavia on April 16 and 18, 1994, were usedfor the experiments. At each date, a four-look XSAR image(X-band, VV polarization) and three four-look SIR-C channels[C-band, HH, HV polarizations and TP (total power) channels]were available (Fig. 1). Ground changes occurred in the scene,as several rice fields were artificially flooded (for cultivationpurpose). A test map, presenting 11 287 “no-change” test pixelsand 1870 “change” test pixels was available. As usual, inremote sensing, this map consisted of disjoint homogeneousregions with no test pixel at the interface between “change” and“no-change.” To assess the behavior of the proposed method atthe border between different regions, a 130 × 100 pixel sizedtest map with two classes was created [Fig. 2(a)], and a multichannelimage was generated by filling the resulting regionswith pixels drawn from the test areas of the SIR-C/XSAR dataset [Fig. 2(b) and (c)]. A semisimulated multichannel imagewas thus obtained for each date, such that every pixel was drawn

MOSER et al.: UNSUPERVISED CHANGE DETECTION FROM MULTICHANNEL SAR IMAGES 281TABLE ICHANGE-DETECTION PERFORMANCE OF THE PROPOSED METHOD, OFTHE INITIALIZATION PROCEDURE, AND OF THE NONCONTEXTUALVARIANT APPLIED BOTH TO THE ORIGINAL (NONFILTERED)DATA(NC-NF) AND TO SPECKLE-FILTERED IMAGES (NC-SF)from a real image and was given a test label (in particular,14.2% of the image area consisted of “change” pixels).In order to focus on the role of the MRF model, two experimentalcomparisons were performed. First, a noncontextualvariant of the method was considered: in the “EM-based classificationstage,” LJ-EM was applied by assuming the labelsnot to be Markovian but rather independent and identicallydistributed (i.i.d.), and by still modeling each conditional PDFas a log-normal. In this case, the parameter vector θ includedthe conditional log-means and log-variances and the prior probabilitiesof the two hypotheses [9]. Second, the application ofspeckle-filtering before the noncontextual variant was carriedout as an alternative to MRFs. In this case, the contextual informationwas exploited not by the classifier but by the spatial prefiltering.The Gamma-MAP filter was chosen as a usually goodtradeoff between speckle removal and computation time 2 [8].By focusing first on the semisimulated data set, two stepsof the method were sufficient to reach convergence, and asexpected, LJ-EM exhibited convergent behavior at both steps.The detection accuracy (i.e., the percentage of correctly labeled“change” test pixels), the false-alarm rate (i.e., the percentageof erroneously labeled “no-change” test pixels), and the overallerror rate (i.e., the percentage of erroneously labeled test pixels)of the resulting change maps were computed according to theexhaustive test map in Fig. 2(a) and are shown in Table I. Avery accurate map was obtained even though the initializationexhibited a lot of missed alarms. A visual analysis confirms thiscomment and also points out that, at least for this data set, theproposed contextual approach does not introduce loss of detailin the edges between “change” and “no-change” [see Fig. 2(d)].The application of the noncontextual approach to nonfiltereddata was not effective [Fig. 2(e)], and the resulting changemap presented just a minor improvement as compared withthe single-channel initialization (Table I). This is interpretedas due to the presence of speckle and confirms the importanceof contextual information. Better results were obtained by thenoncontextual variant when applied to speckle-filtered images:A very good detection accuracy was achieved but along withmany false alarms (Table I). A numerical comparison betweenthis contextual strategy and the one adopted by the proposedmethod suggests that the latter gives more accurate results2 Two Gamma-MAP filtering iterations with a 7 × 7 window were adopted.Such values for the filter parameters were selected by “trial and error,” with atradeoff between speckle removal and detail loss.(Table I). Many false alarms can be noted at the interfacebetween “change” and “no-change” in the map generated bythe noncontextual variant after despeckling [Fig. 2(f)], whereasthe MRF-based method provided a more accurate result in thesecritical areas as well.Concerning the real data set, four steps of the method weresufficient to reach convergence, and at each step, LJ-EM convergedin 14 or 15 iterations. A good detection accuracy (above93%) is obtained by this method with very low error rates(below 1%; see Table I). A visual analysis of the correspondingmaps confirms these comments [Fig. 3(a); note that no falsealarm was obtained in the test areas, although some false alarmscan be noted in other image areas, often due to small misregistrationerrors between the images acquired at the two dates].A large improvement is obtained by the proposed method, ascompared with the initialization map. LN-GKIT, applied tothe single channel selected by the procedure in Section II-D(which turned out to be SIR-C-HV), provided a poor result dueto the strong presence of speckle. Owing to the joint use ofmultichannel and contextual information, the proposed methodsharply improved the detection accuracy and the error rateduring the four steps.Less accurate results were obtained by the noncontextualvariant. When applied to nonfiltered data, it generated an evenworse change map than the initialization one (see Table I), withmany missed and false alarms, the latter ones being particularlyconcentrated in image areas covered by wet soil at the seconddate [see Fig. 3(b)]. A similar effect was also noted whenapplying the noncontextual variant after speckle filtering: eventhough the detection accuracy sharply improved as compared tothe case without despeckling, many false alarms were present inthe areas occupied by the wet soil in the April 18th image (thecorresponding false-alarm rate in Table I is still quite low, since,in this case, most false alarms are located outside the test areas).This can be interpreted as being due to the possible multimodalityof “no-change,” each mode being related to a specificland-cover subclass (a multimodal behavior is not expected for“change” because only one change typology is present in thisdata set). LJ-EM, formulated according to a two-componentmodel (“change” versus “no-change”), can be sensitive to suchmultimodality because it can erroneously merge one or more“no-change” mode(s) with the “change” one. Since the mapgenerated by the “EM-based classification stage” of each stepis used as a training set in the “feature-transformation stage”of the next step, this merging is also expected to affect theupdate of the SAR-specific Fisher transform. This undesirableeffect was not observed in the results of the proposed method,owing to the fact that the spatial regularization energy term ofthe MRF model explicitly takes into account the spatial contextin the iterative parameter update [see (4)], thus optimizing ateach iteration the separation between the two hypotheses.IV. CONCLUSIONAn unsupervised contextual change-detection technique hasbeen proposed for multichannel SAR images. Accurate changemaps were provided by the method, when applied to bothsemisimulated and real data. No preliminary despeckling wasneeded, owing to the capability of MRFs to exploit the spatialcontextual information in the classification process. A specific

282 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 4, NO. 2, APRIL 2007Fig. 3. Real data set. Change maps generated by (a) the proposed method, and by its noncontextual variant applied (b) to the original images and (c) afterdespeckling (the color legend is the same as in Fig. 2).experiment, carried out with semisimulated data endowed withan exhaustive test map, has also demonstrated the accuratebehavior of the method at the interface between “change” and“no-change,” with no significant loss of detail, at least, on thisdata set. A fully realistic experiment would need to take intoaccount the mixed pixels, which is a problem beyond the scopeof this letter. Two further versions of the method, adoptingthe standard EM [18] and the stochastic EM [15] algorithms,have also been explored and have provided very similar results(see [20] for details).A noncontextual variant of the method, even when appliedto despeckled images, has proven to be strongly sensitive tothe possible multimodality of the “no-change” hypothesis. Thisdrawback has not been observed for the proposed approach,which further confirms the effectiveness of the MRFs as contextualclassification tools. On the other hand, according tothe monomodal model for the “change” statistics, only onetypology of change (either an increase or a decrease in thebackscattering coefficient) is assumed to be present in the image.When both an increase and a decrease in the backscatteringcoefficient are present, the proposed method can be separatelyapplied to the two possible ratio images [4], [8] (preliminaryexperiments, not reported here for brevity, have confirmed theeffectiveness of this procedure). However, when many typologiesof change are present, the adopted two-component modelmay lead to an incorrect merging of some “change” and “nochange”subclasses. A multiclass extension of the method couldbe developed to overcome this difficulty.The proposed technique is iterative, and the experiments haveremarked good convergent behavior in a small number of steps,although the method was initialized with (quite) low-accuracychange maps. However, the convergence properties have notbeen analytically proven so far and represent an interestingissue worth being investigated. Future extensions of this workcould examine integration of more sophisticated MRF modelsto further reduce the error rates (e.g., by taking into account theimage nonstationarity [21] or by including edge information),or combination of this method with object-based approaches[22], to better adapt the method to the application to highresolutionimages.ACKNOWLEDGMENTThe authors would like to thank P. Gamba (Universityof Pavia, Italy) for providing the data employed for theexperiments.REFERENCES[1] J. Inglada, “Change detection on SAR images by using a parametric estimationof the Kullback–Leibler divergence,” in Proc. IGARSS, Toulouse,France, Jul. 21–25, 2003, vol. 6, pp. 4104–4106.[2] G. Mercier and S. Derrode, “SAR image change detection using distancebetween distributions of classes,” in Proc. IGARSS, Anchorage, AK,2004, vol. 6, pp. 3872–3875.[3] Y. Bazi, L. Bruzzone, and F. Melgani, “An unsupervised approach basedon the generalized Gaussian model to automatic change detection inmultitemporal SAR images,” IEEE Trans. Geosci. Remote Sens., vol. 43,no. 4, pp. 874–887, Apr. 2005.[4] G. Moser and S. B. Serpico, “Generalized minimum-error thresholdingfor unsupervised change detection from SAR amplitude imagery,”IEEE Trans. Geosci. Remote Sens., vol. 44, no. 10, pp. 2972–2982,Oct. 2006.[5] W. Dierking and H. Skriver, “Change detection for thematic mapping bymeans of airborne multitemporal polarimetric SAR imagery,” IEEE Trans.Geosci. Remote Sens., vol. 40, no. 3, pp. 618–636, Mar. 2002.[6] M. Qong, “Polarization state conformation and its application to changedetection in polarimetric SAR data,” IEEE Geosci. Remote Sens. Lett.,vol. 1, no. 4, pp. 304–308, Oct. 2004.[7] K. Conradsen, A. A. Nielsen, J. Schou, and H. Skriver, “A test statistic inthe complex Wishart distribution and its application to change detectionin polarimetric SAR data,” IEEE Trans. Geosci. Remote Sens., vol. 41,no. 1, pp. 4–19, Jan. 2003.[8] C. Oliver and S. Quegan, UnderstandingSynthetic Aperture RadarImages. Norwood, MA: Artech House, 1998.[9] F. Melgani, G. Moser, and S. B. Serpico, “Unsupervised change detectionmethods for remote sensing images,” Opt. Eng., vol. 41, no. 12, pp. 3288–3297, 2002.[10] Q. Jackson and D. A. Landgrebe, “An adaptive classifier design for highdimensionaldata analysis with a limited training data set,” IEEE Trans.Geosci. Remote Sens., vol. 39, no. 12, pp. 2664–2679, Dec. 2001.[11] K. Fukunaga, Introduction to Statistical Pattern Recognition, 2nd ed.New York: Academic, 1990.[12] R. C. Dubes and A. K. Jain, “Random field models in image analysis,”J. Appl. Stat., vol. 16, no. 2, pp. 131–163, 1989.[13] P. Lombardo, C. J. Oliver, T. M. Pellizzeri, and M. Meloni, “A newmaximum-likelihood joint segmentation technique for multitemporalSAR and multiband optical images,” IEEE Trans. Geosci. Remote Sens.,vol. 41, no. 11, pp. 2500–2518, Nov. 2003.[14] Q. Jackson and D. Landgrebe, “Adaptive Bayesian contextual classificationbased on Markov random fields,” IEEE Trans. Geosci. Remote Sens.,vol. 40, no. 11, pp. 2454–2463, Nov. 2002.[15] G. Celeux, F. Forbes, and N. Peyrand, “EM procedures using meanfield-like approximations for Markov model-based image segmentation,”Pattern Recognit., vol. 36, no. 1, pp. 131–144, Jan. 2003.[16] R. Cossu, S. Chaudhuri, and L. Bruzzone, “A context-sensitive Bayesiantechnique for the partially supervised classification of multitemporal images,”IEEE Geosci. Remote Sens. Lett., vol. 2, no. 3, pp. 352–356,Jul. 2005.[17] W. H. Press, S. A. Teukolsky, W. T. Wetterling, and B. P. Flannery,Numerical Recipes in C. Cambridge, U.K.: Cambridge Univ. Press,2002.[18] R. A. Redner and H. F. Walker, “Mixture densities, maximum likelihood,and the EM algorithm,” SIAM Rev., vol. 26, no. 2, pp. 195–239, 1984.[19] J. Kittler and J. Illingworth, “Minimum error thresholding,” PatternRecognit., vol. 19, no. 1, pp. 41–47, Jan./Feb. 1986.[20] G. Moser and S. B. Serpico, “Unsupervised change-detection from multichannelSAR data,” in Proc. NORSIG, Reykjavk, Iceland, Jun. 7–9, 2006,pp. 246–249.[21] X. Descombes, M. Sigelle, and F. Preteux, “Estimating Gaussian Markovrandom field parameters in a nonstationary framework: Application toremote sensing imaging,” IEEE Trans. Image Process., vol. 8, no. 4,pp. 490–503, Apr. 1999.[22] G. G. Hazel, “Object-level change detection in spectral imagery,”IEEE Trans. Geosci. Remote Sens., vol. 39, no. 3, pp. 553–561,Mar. 2001.