SAR Images Change Detection Based On Feature Learning Of Unsupervised Neural Networks

Image change detection is the procedure that analyzes multiple images over the same scene taken at different time to find the changed regions over that scene.SAR is synthetic aperture radar which could produce high resolution images.The image obtained by the synthetic aperture radar coherent imaging system is called SAR image.At present,SAR image is becoming the main data source for the remote sensing image change detection because it has the incomparable advantage of optical image.At the same time,SAR image change detection has increasingly become an important topic in scientific research both at home and abroad.In recent years,deep learning has attracted more and more researchers' attention because of its powerful feature learning ability.This paper mainly improves some traditional methods with neural networks,and presents two new change detection methods based on the feature extraction of auto encoder machine.The focus of this paper is as follows:1.The traditional spectral clustering method combined with neural network feature extraction is studied and improved.Due to the way of generating similarity matrix in the traditional spectral clustering algorithm concerns only the pixels of the image itself which doesn't make full use of image information,an improved spectral clustering algorithm based on neural network feature extraction is proposed and applied to the field of SAR image change detection.In the proposed method,the construction of similarity matrix of traditional spectral clustering is mainly optimized.Firstly use sparse denoising auto encoder to extract image features which combine pixel neighborhood information and other feature information to optimize similarity formula mode in order to improve classification accuracy.Then use Nystrom algorithm to avoid calculating the whole similarity matrix,instead estimate the whole part by sampling part to reduce the computational complexity of the algorithm.The performance of this method on different data sets is tested,and satisfactory results are obtained.2.The way of neural network feature extraction has been studied and improved.The proposed method aims to deal with some problems existing in the combination of single scale sparse auto encoder and FLICM algorithm applied in the field of SAR image change detection.In the proposed method,firstly,the Gauss weighted sparse auto encoder input window is used to reduce the influence of neighboring pixels on the center pixels in the extracted features in order to add some local details to the result map.Then use joint feature concatenated by features extracted from multi-scale sparse auto encodes to make full use of the information difference brought by different scale features in order to improve algorithm classification accuracy.Considering the contribution rate of the different scale features is different to the result,use simple linear weighting method to weight different scale features and then concatenate them to get joint feature in order to get more reasonable difference image feature representation.The performance of this method on different data sets is tested,and satisfactory results are obtained.