Change Detection Technology For Remote Sensing Image Based On Feature Learning With Auto-encoder

With the development of society and progress of technology,the requirements for people to remake nature and exploit natural resources are also growing.The ground information and resource overlay are changed by human activities and climate changes.Change detection based on remote sensing technology with advantage of real-time strong,wide coverage and multi spectral is also a hot spot in the research of remote sensing application.The current change detection methods can basically be divided into two categories: postclassification comparison and post-comparison analysis.The results of detection rely heavily on results of classification with methods of post-classification comparison which make it difficult to guarantee the accuracy of change detection.At this stage,the way of postcomparison analysis is adopted by the mainstream algorithm framework.In this way,change information is extracted to generate a difference image and then we analyze the difference image to get the detection result.Remote sensing images,especially synthetic aperture radar images,are often accompanied by speckle noise due to their own imaging mechanism.The noise will seriously interfere the produce of changing information,and reduce the quality of the difference image.How to suppress interference of noise is a difficult problem for many algorithms.It is also the key problem to be solved in this article.In this thesis two methods are designed.Multilayer neural network is introduced to implement an auto-encoder network.The result of feature learning is used for denoising.Change detection results are improved.1.For high probability of error detection because of noise in multi-temporal images,a stacked sparse auto-encoder network is implemented based on multilayer neural network.The high order feature expressions of original images are obtained by the nonlinear mapping of auto-encoder layer by layer.By adding sparsity constraints to the network,the features are compressed.Then images are reconstructed based on higher-order features.In reconstructed image,the geomorphic information of the original image is retained and the influence of noise is effectively alleviated.By using the images after denoise,the difference images are generated by the log-ratio operator.At last,the binary classification results are obtained by methods of clustering or adaptive threshold.Experimental results show that the performance of the proposed method has improved significantly compared with the traditional methods of change detection.2.In order to solve the problem that the heavy uncertainty of difference analysis when the quality of difference images are poor,a multi-layer classification neural network based on supervised training is implemented.It makes full use of neural network with a good performance as nonlinear classifier.To obtain more accurate training markers,multi-metric difference information is calculated,and the fuzzy C clustering algorithm is used to screen out the marked areas with high credibility.Then the neighbor-feature samples are generated from the multi-temporal images as the data set for training network.In order to get better training effect and improve the accuracy of the classification network,the network is pretrained by contractive auto-encoder to complete the initialization of the network hidden layer weight.Then the soft Max layer is added and the training set is input to finish training of network.Finally,a trained network is used to classify all the neighborhood samples to get the final detection results.It is proved that the method is effective in the case of poor quality of the difference image.The effectiveness of the method is verified by tentative experiments.