Remote Sensing Image Change Detection Based On Kernel Methods

Remote Sensing images obtain information of the far and untouched targets by receiving their reflected and radiant electromagnetic waves.With the development of the sensor technology,synthetic aperture radar(SAR)images which are insensitive to atmospheric and sun illuminations and multispectral images which possess high spectral resolutions play an important role in remote sensing image interpretation.Change detection is a research hot spot nowadays.It aims at identifying changes in images obtained at the same scene but different times.It is in great demands in military and civil activities.In this dissertation,we study the remote sensing image change detection techniques.Focusing on the multispectral and SAR images,we study new change detection technologies on the basis of kernel methods.This dissertation contributes in designing advanced kernel functions and kernel-based classifiers.Several improved kernel functions and kernel-based classifiers are proposed for providing strong noise or disturbance immunity,good details preservation and effective feature extraction and fusion schemes.Main contents of this dissertation consist of the following five parts.It is known that performance of the supervised change detection methods suffers from lack of training samples.Therefore,a cluster-neighborhood(CN)kernel is proposed for SAR image change detection.In the proposed method,two composite kernels are built first.Then,samples are categorized into two neighborhoods with the kernel k-means algorithm.Next,a CN kernel is constructed on the basis of a composite kernel and the neighborhood-based statistical features.When a few labeled samples are available,the proposed CN kernel explores the information of unlabeled samples to enhance its discriminative ability and robustness against speckle noise.Finally,the support vector machine(SVM)is implemented to get the final change detection results.Experimental results on real SAR images demonstrate the high precision,strong noise resistance and low false alarm rates of the proposed method.It is known that the spatial-neighborhood information is helpful in suppressing degrading effects of the noise.Therefore,a label-information composite(LIC)kernel constructed on the basis of the spatial-neighborhood information is proposed for SAR image changedetection.The method focuses on the extraction and utilization of the spatial-neighborhood information,and the kernel-based fusion scheme proposed provides strong noise immunity as well as good preservation of edge locations of changed areas.In the proposed method,the anisotropic Gaussian kernel model is utilized for analyzing anisotropic textures of the bi-temporal images first.Then a comparison scheme is proposed whose results are used to supervise the extraction of the output-space label-neighborhood information.Next,construct the LIC kernel and update it iteratively with the newest change map outputted from the SVM until the change map converges.Experiments on real SAR images demonstrate the effectiveness of the LIC kernel method and illustrate that it has both strong noise immunity and good preservation of edge locations of changed areas for SAR image change detection.It is found that the complementary information existing in the subtraction image and the ratio image has found limited applications in real tasks by now.Therefore,a multiple wavelet fusion(MWF)kernel method which utilizes multiple wavelet kernels for fusing the complementary information of the two difference images is proposed for multispectral image change detection.First,the complementary information of the two difference images is analyzed.Then,for each difference image,wavelet kernels at multiple scales are computed followed by a reliable scale selection scheme based on correlation coefficients.After that,two difference images' wavelet kernels at reliable scales are fused to construct the MWF kernel.Finally,the MWF kernel is inputted into a classification algorithm based on the minimum Euclidean distance in the kernel space to get the final change detection result.Experiments demonstrate the effectiveness of the proposed method,and illustrate that it possesses both strong disturbance immunity and good homogeneity of changed areas for remote sensing image change detection.Performance of the k-means clustering algorithm for SAR image change detection is usually worsened by the inherent existence of the speckle noise.Therefore,in this part an unsupervised multiple kernel k-means clustering algorithm with local-neighborhood information(LIMKKM)is proposed for SAR image change detection.LIMKKM algorithm contributes in two aspects.First,it fuses various features through a weighted summation kernel by automatically and optimally computing the kernel weights.Here,the intensity and texture features of the ratio image are fused.Second,it incorporates the local-neighborhood information into its clustering objective function for providing strongnoise immunity.The LIMKKM change detection algorithm is carried out in a train-test way to lighten the computational burden.Experimental results on real images demonstrate the effectiveness,especially the strong noise immunity,of the LIMKKM method,and illustrate that it is suitable for SAR image change detection.For constructing a powerful kernel function specifically for change detection tasks and an improved extremely extreme learning machine(ELM),this part proposes a SAR image change detection method relying on a difference correlation kernel(DCK)function and a multistage ELM(MS-ELM).This method contributes in two aspects.First,a DCK function is computed for measuring the distance between any two pixels.The DCK depicts the cross-time relations between coupled bi-temporal image patches at any cyclic shifts with a kernel correlation operation and the high-order spatial relations between two different located pixels with an algebraic subtraction.Second,a MS-ELM classifier is carried out to obtain the change detection result.In MS-ELM,the hidden nodes and weights between the hidden and output layers are updated stage by stage by improving the kernel functions with the output spatial-neighborhood information of the previous stage.Both the DCK function and MS-ELM provide strong noise immunity and good discriminating ability.Experiments on real SAR image change detection demonstrate the effectiveness of the method,especially their vigorous discriminating ability and robustness against noise in dealing with SAR images.