SAR Images Change Detection And Semantic Analysis Based On Deep Learning

Synthetic aperture radar images change detection and semantic analysis are to extract change regions and identify change contents by comparing SAR data at different times in the same area.Because of the advantages of all day and all-weather imaging of SAR system,the detection and analysis of SAR image changes have shown broad application prospects in the fields of military target monitoring,key area investigation,disaster prevention and disaster relief and so on.The existing methods can effectively eliminate the influence of speckle noise in SAR images,however,they usually do not maintain well the details of the changing area.On the other hand,most researchers focus on the accurate extraction of changing regions,and there is no effective way of changing content semantic analysis,which seriously restricts the further development of SAR image application.In order to solve these problems,this paper focuses on the change detection and semantic analysis of SAR images,and studies the key steps of the acquisition and segmentation of different map and SAR image feature extraction.On this basis,we study a change analysis method based on the deep learning technology.The main works and creations are as the following:1.A different map enhancement algorithm based on dyadic wavelet transform is proposed.The gray distribution of classical log-ratio difference map is so centralized that it is very difficult to be segmented.In this paper,we use dyadic wavelet theory to study a differential map adaptive enhancement algorithm.Firstly the method decomposes the LR difference map by dyadic wavelet transform.Then we make use of the difference between the correlations of wavelet coefficients in noise and signal to enhance the wavelet coefficients adaptively,and rebuilt the difference map finally.The method can effectively improve the signal-to-noise ratio and the separability of the difference map,which is beneficial to further differential map segmentation.Experimental results show that this method can effectively enhance the contrast ratio between the changing area and the non-changing region.2.A new method which combines local edge information and fuzzy C means algorithm is studied.Classical FCM and improved algorithm are difficult to balance between noise suppression and preservation of detail information when segmenting difference maps.To solve the problem,we first analyze the principle of FCM and the improved algorithm.Then the image edge information is extracted by the robust edge operator,and it is introduced into the neighborhood window of the local information FCM algorithm to achieve the adaptive suppression of the noise in the segmentation process.The experimental results show that this method can effectively suppress speckle noise and keep the details better,and obtain better precision in the analysis of different map.3.A new unsupervised deep network model is designed to achieve high robustness SAR image feature extraction.Based on the classic unsupervised learning network,the residual idea is introduced into the convolution auto-encoder,and a new network model is designed in this paper.By adding shortcut connections,the new model changes the single cascaded structure of the traditional network,effectively solving the problem of gradient propagation in the training of deep network,and can obtain more effective characteristics under the same output dimension.4.A change semantic analysis method based on deep learning is proposed.It is difficult to distinguish different types of changes in pixel level or gray level.To solve the problem,under the framework of bag of visual word model,we combine previous research results.the SAR images are transformed into one-dimensional histogram representation by unsupervised learning network.By calculating the feature difference vector,we can represent the change of SAR image on the grayscale and texture.On this basis,we construct different map and extract change region.Finally,the result of semantic analysis is obtained through the classification network.