Research On Change Detection Techniques In Multitemporal Remote Sensing Images

As an important application of remote sensing image analysis, change detection provides effective techniques in many application fields, such as environmental surveillance, resource exloration, disaster relief and management, etc. Although human interpretation is a kind of relatively reliable change detection method, this visual interpretation based method has many problems, for example, it suffers from heavy work load, low efficiency, strong subjectivity, high error rates, etc. In the past20years, change detection methods for remote sensing images have been updated constantly, however, change detection is still affected by multiple factors. For example, remote sensors is affected by seasons, weather, sun height, illumination and atmospheric conditions, etc., which results in spectral differences between different temporal samples. These differences are mixed with the differences of real geographical objects, increasing the difficulties in change detection. Numerous studies attempt to find all kinds of new methods for change detection in remote sensing images. But so far, no general method has been found which has totally satisfactory results for all different conditions and applications. In this thesis, the following study is conducted in view of the shortcomings and limitations of several existing methods:In order to deal with the problem existing in ordinary MRF based methods that the weight between prior energy and the likelihood energy is set to an identical value in different image areas, an adaptive weight Markov random field (MRF) based change detection method is proposed. The proposed method firstly extracts the image detail features, judging the locations of such features, and set relatively big weight to locations with no image detail features while set relatively small weight to locations with image detail features. Based on the above idea, a change detection method based on adaptive weighted MRF model is proposed. First, edge pixels are extracted by8-neighborhood line process; then the necessary conditions of adaptive weighting functions (AWFs) are defined. Eight examples of AWFs are provided; At last, the method is tested on multitemporal remote sensing images to verify the effectiveness.In order to deal with the problem existing in ordinary EM parameter estimation method that it does not consider the pixel neighborhood information and is vulnerable to noise which results in inaccuracy in parameter estimation, an expectation maximization parameter estimation method is proposed based on evidence theory and is applied to change detection. To use neighborhood information in parameter estimation, Dempster-Shafer’s theory of evidence is integrated into the ordinary EM algorithm in this study, so that each parameter updating step depends not only on the intensity of the current pixel, but also on the neighboring pixels’, which yields an evidence theory based EM algorithm (EEM). To further improve the accuracy of change detection, the maximum a posteriori (MAP) labeling method is used to enhance the result of EEM method. The class labels of the difference image is assumed to be smooth, then the labels generated by the EEM method is iteratively updated by the MAP labeling method. Experimental results show that the noise suppression ability of MAP labeling is superior than the EEM.To deal with the problem that the ordinary active contour models (ACM) model is not suitable for change detection in synthetic aperture radar (SAR) images, a change detection method for multitemporal SAR images is proposed based on generalized Gaussian distribution (GGD) and ACM. Recently, ACM is used for change detection. However, SAR images are usually polluted by multiplicative noise. And the conventional C-V active contour model assumes an image to be piecewise smooth, which violates the data property of SAR images. Therefore the C-V model cannot be directly used for change detection of SAR images. In this study, the C-V model is generalized under the assumption of generalized Gaussian mixture model (GGMM), resulting in a change detection method for multitemporal SAR images based on GGD and ACM. Experiments are performed to evaluate the effectiveness of the method.In order to deal with the problem that the commonly used CVA method loses the information in the spectral feature space, a change detection method for multispectral and multitemporal remote sensing images is proposed based on stationary wavelet transforms (SWT) and integrated active contour (IAC) models. The method view the spectral change vector space as a2-dimensional Riemannian manifold embedding into a2+B dimensional manifold, where B is the number of spectral bands. The segmentation of the change vector image is performed by the curve evolution on the manifold, i.e. IAC. The IAC model combines the advantages of geodesic active contours (GAC) and edgeless active contours (C-V model), increasing the accuracy of change detection. To overcome the problem of noise in remote sensing images, two measures are adopted:Use SWT to generate the multiresolution representation of the change vector image; Use the metric of manifolds to define a regional homogeneity measurement, and use this measurement to choose the reasonable scale for each pixel.The above methods are tested on different synthetic data sets or real multitemporal remote sensing images. Experiments results show that the above methods are comparable to other prevalent methods, most of the results of the methods proposed in this study are better than other prevalent methods.