Change Detection Methods For Remotely Sensed Images Based On Enhanced Spatial Information

Remotely sensed image change detection has been one of the most important technique for the earth observation. However, due to the complexity of interactions in the natural environment as well as the natural environment and remote sensing sensors, the remotely sensed images contain the problems of a large number of mixed pixels, the fuzzy boundary, the same objects with the different spectrum and the different objects with the same spectrum. In addition, uncertainties of the pre-processing and change detection methods also decrease the accuracy of change detection. The current study shows that, change detection methods combining spectral and spatial information can solve the problem above in a certain extent, but the utilization of spatial information is not accurate. Uncertainties of the existing change detection methods are deeply analyzed, then the relationship between the misregistration and change detection accuracy is studied. For different spatial resolution remotely sensed images, studies are made respectively on the pixel-, object- and feature-level. By enhancing the spatial information, reliable change detection methods combining spectral and spatial information are put forward, which reduces the impacts of uncertainties in the remote sensing data and change detection methods on the change detection results, and finally improve the accuracy of the change detection. The study will provide a new approach to change detection by combining the spectral and spatial information for remotely sensed images. The details are as follows:(1) The errors of change detection caused by misregistration and their spatial distribution are summarized by experiments: 1) The required registration accuracy is at least close to 0.6 pixels to get change-detection results with 90% accuracy when only the misregistration factor is considered. 2) The commission errors caused by misregistration values from 0 to 1 pixel are almost always within 1 pixel of the edge, regardless of image resolution. 3) In addition, the omission errors falling within 1 pixel of the edges are about 70% for medium-resolution images. The omission errors falling within 1 or 2 pixels of the edges for high-resolution images can be as much as 50% to 60%.(2) Considering the problems of the mixed pixels and fuzzy boundary in remotely sensed images, new change detection methods based on the active contour model are proposed: 1) Supposing that the difference image can be seen as a mixture Gaussian distribution, the mean gray values of not changed and changed pixels are evaluated by EM algorithm and introduced into the active contour model to establish new energy functions to strengthen the distinction between not change and change pixels, which improves the accuracy of the change detection on the pixel-level. 2) An advantage fusion strategy is put forward at the feature-level by fusing the small scale and large scale change detection map generated by the active contour model, remaining the advantages of the both change detection maps under different parameters. In a certain extent, it reduces the influence of the contour length parameter on the change detection results. 3) Building outlines provided in GIS vector data before the earthquake as the initial contour, the active contour model is implemented to the post-earthquake high resolution remotely sensed image, and the collapsed buildings are finally detected, which avoids setting a threshold and enhances the accuracy and stability of results on the object-level.(3) Change detection methods of Markov random field with enhanced spatial neighborhood information are proposed on the pixel-level: 1) The membership information of fuzzy C-mean algorithm is introduced into the Markov random field by the spatial attraction model, which enhances the accuracy of the spatial relationship between the neighborhood pixels and obtains more accurate change detection results. 2) Two thresholds T1 and T2 are set according to the estimated center gray values of changed and not changed pixels, and the difference image is then divided into not changed, uncertain and changed parts using the two thresholds. Given that, different calculation methods of spatial information weight are designed for each part to weaken the excessive utilization of spatial information and improve the accuracy of change detection.(4) Taking into account the characteristics of objects, the optimal segmentation scale is determined for different changes, and two change detection methods by using multi-scale information of the segmented objects are presented on the object-level: 1) The initial change detection results generated by the active contour model are refined using the segmentation of the SRM algorithm, which weakens the influence of the segmentation scale and the contour length parameter in the active contour model and enhances the stability of change detection results. 2) Through combining the pixel- and object-based change detection results, uncertainties caused by the segmentation scale are analyzed deeply. On the basis, more precise segmentation results are adopted to refine the uncertain objects to improve the change detection accuracy.(5) On the feature-level, an edge density matching index is proposed, and three textures of GLCM, Gabor and GMRF are introduced. These spatial features and spectral information are combined to detect changes. In the process, all features are decomposed using the wavelet transform and the change information is then extracted by DS evidence theory and advantage fusion strategy. Experimental results show that the combination of spectrum, texture and edge features can improve the accuracy of change detection to a large extent, and the improved degree depends on what features are combined.