Change Detection For Optical And SAR Images By Considering Deep Information Extraction And Feature Space Unification

The surface change information provided by remote sensing image change detection plays an important role in the study of the development of the nature world and of human society.Compared with homogeneous image change detection using the same data source,heterogeneous image change detection can break the limitations of single data source and make full use of the respective advantages of different data.However,due to the differences of imaging mechanisms and image radiation characteristics,heterogeneous images can not directly apply the traditional change detection methods of homogeneous images.Among the existing change detection methods for heterogeneous images,the results of Post classification comparison methods are easily affected by the classification algorithm and are prone to error accumulation,similarity methods need to construct efficient "features" to highlight the differences between changed and unchanged pixels,while feature space unification methods are not subject to aforementioned restrictions but the current methods still have the problems of manual selection of training samples and low efficiency of feature mapping.To solve these problems,considering the advantages of feature space unification method,this paper takes optical images and SAR images as the main research object and studies the change detection method between them.The specific work is as follows:(1)Aiming at the problems of manual selection for training samples and of mapping randomness,a change detection method based on deep feature change information for heterogeneous remote sensing images is proposed.The proposed method firstly extracts multiple feature images of heterogeneous images by stacked denoising autoencoder,which transforms heterogeneous images from the original incomparable feature space to the same feature space for comparison.By introducing similarity measurement,the most similar feature map is determined by calculating the structural similarity index between the feature maps,and the difference map is obtained by the subtraction method.Then the original heterogeneous images are used to construct feature vectors,from which training samples are automatically selected according to the difference map and used to train the support vector machine(SVM).Finally the trained SVM is used to classify the feature vectors to obtain the change detection map,and the morphological operation is used to reduce the noise.This paper conducts an experiment for a group of optical and SAR images to verify the feasibility of the proposed method,which provides idea for the subsequent method.An experiment for a group of SAR images verifies the applicability for homogeneous change detection.Experiment results show that this method can detect changed areas on a whole but still has the problems of low efficiency and can not distinguish types of changes.(2)Aiming at the problems of the aforementioned method that has low mapping efficiency and can not distinguish types of changes,a change detection method based on symmetrical network for heterogeneous remote sensing images is proposed.The proposed method uses multi-layer forward encoder to construct a symmetrical network for features extraction of heterogeneous images,which increases the mapping efficiency.The network is initialized by similarity measuring,guiding the direction of network feature mapping and updates iteratively so as to extract better feature maps used to generate difference map.The change detection map is obtained by clustering analysis of difference map.Finally,change vectors are constructed by the change area of original images,considering the neighbourhood information.Multi-type change detection is consequently realized by clustering analysis of change vectors.Experiments for two groups of optical and SAR images with different scene are conducted.Experiment results show that comparing existing methods,this method increases the accuracy by over 13.47% and decreases the running time by 67.12% at least,which therefore confirms the effectiveness,robustness and efficiency of the proposed method.Both of the proposed methods extract change information by unifying heterogeneous images into the same feature space for comparison and belong to unsupervised change detection,which are in line with the development trends of change detection.18 figures,5 tables,69 references...