A Study Of Remote Sensing Image Change Detection Method Based On Hierarchical Structure

With the acquisition techniques of remote sensing image become more and more advanced and the massive accumulation of remote sensing data, the temporal remote sensing image change detection techniques are applied in many applications, such as monitoring resources and environment, updating geographic data, assessment of natural disasters, urban planning and post-disaster reconstruction and so on. The process of change detection method are broadly splitted into four stages, which consists of preprocessing, difference image acquisition, difference image analysis and accuracy evaluation. Among them, the difference image acquisition and analysis are two critical factors affecting the accuracy of change detection. This paper mainly focused on the difference image analysis step and conducted an in-depth research. The author's major contributions are outlined as follows:1. For questions the difference image contains a large number of speckle noise and may causes interference for subsequent analysis and processing, we proposed a noise suppression method for remote sensing image based on the directional characteristics of the non-subsampled contourlet transform(NSCT). This method can extract the changed area effectively and ignore the influence of speckle noise. It obtains the decomposition coefficients of the difference image at multiple scales in multiple directions by NSCT. Then it calculates the energy in each direction on each scale of the decomposition coefficients and sorts by descending order and takes a thresholding process on the coefficients in each direction on each scale. After finish the processing in each direction on each scale, inverse NSCT achieves the effect of image de-noising. This method can retain the target pixel area while suppressing the interference noise.2. Based on the shortcomings of the unsupervised change detection method which detection accuracy not high and robustness weak, and the supervised change detection method which requires a lot of training samples, we proposed a remote sensing image change detection based on online-learning. This method abandoned the traditional methods of analysis with the whole difference image processed pixel-by-pixel classification, but the difference image is divided into image blocks and a similar form of video frames is obtained. Then we classify the single frame and update the sample database constantly through the online-learning mechanism, and constantly optimizing the performance of the classifiers. Finally, splicing classification detection results of each frame into a whole image, we complete the analysis of the difference image. Compared with the traditional unsupervised change detection method, the method proposed in this chapter has higher detection accuracy and more universal applicability.3. For this paper, based on the degree of automation of the remote sensing image change detection method based on online-learning is not high enough, an adaptive remote sensing image change detection method based on NSCT de-noising and support vector machine(SVM) classification was proposed. This method first constructed the difference image and conducted image de-noising step based on NSCT, then on the de-noised difference image conducted an overlapping 2×2 block of scanning and "coarse classification" with the mean-classifier step. At the same time, according to the threshold of the mean-classifier the sample library based on the original coarse scales could be constructed and the SVM classifier could be trained. Finally, after the "coarse classification" step, the remainder pixel points were tested by SVM — the "fine classification". And after getting the remaining pixel points' categories, then the final change detection result could be got. The proposed method while maintaining high detection accuracy, avoid human interventions, improving the degree of change detection methods' automation. Compared with the traditional unsupervised remote sensing image change detection method, this method has higher detection accuracy.