Feature Learning And Classification Based On Deep Neural Networks For Change Detection In SAR Images

Change detection in remote sensing images aims to highlight the changed regions by analyzing two images over the same area acquired at different times.At present,it has been widely used in urban planning,environmental monitoring,hazard assessment.In recent years,synthetic aperture radar(SAR)images gradually play an important role in the field of remote sensing and attract more and more attention.This is mainly because of its independence on atmospheric and weather condition.In this paper,we have studied and analyzed the change detection methods in SAR images,and some change detection methods have been improved.Combining some novel techniques,we have come up with some new change detection methods.The focus of this thesis includes as follows: 1.Feature learning method which is based on deep neural networks has been studied.Combining deep learning and some problems in SAR image change detection,this part proposes a new SAR image change detection model based on feature learning.In the proposed method,feature learning is consist of unsupervised feature learning based on sparse auto encoder and supervised feature learning based on convolutional neural networks.The unsupervised feature learning based on sparse auto encoder aims to transform the difference image generating by comparing two SAR images into the feature space,and it can suppress the impact of noise and provide the classification labels for the following supervised feature learning.The supervised feature learning based on convolutional neural networks mainly aims to extracted more abstract features from the feature maps acquired by sparse auto encoder and make use of these features to detect the changed region and unchanged region more accurately.We have experimented the proposed method on many data sets,and compared with the best classification result in the four compared algorithms,the best classification result of our algorithm can be improved by 4.76% on Kappa coefficient.2.KI algorithm has been studied and improved.The improved method is proposed to deal with some problems in the original KI algorithm that the algorithm fails to make full use of the context information between the pixels and the feature information of the pixels and is easy to be influenced by the noise.In order to make full use of the neighborhood information and feature information,and reduce the impact of the noise,the improved method mainly learns the features by the sparse auto encoder from difference image and transforms the difference image into the feature space.According to the extracted features,we can get the membership degrees of every pixel that belongs to the changes and unchanges.Combining the membership degrees and the discrimination formula of the KI algorithm,we can get a more reasonable threshold.Compared with the original KI algorithm by the experiments,the best classification result of the improved method can be improved by 12.78% on Kappa coefficient.