Change Detection In Synthetic Aperture Radar Images Based On Deep Neural Networks

Image change detection detects the changes between two multitemporal images of the same scene.By the end of the last decades,with the development of remote sensing technology,change detection in remote sensing images has become an increasingly important part in the field of satellite image processing.Now,this technology has extensively applied to both civil and military fields,such as disaster evaluation,medical diagnosis and video surveillance.Deep neural network comes from the neural network in the last century.With the passage of time,the artificial neural network is replaced by shallow learning methods such as SVM,Boosting and maximum entropy method,because of the high computational complexity and the difficult of training.In this century,Hinton suggests that artificial neural networks with multiple hidden layers have very good feature learning capabilities that can learn the characterize of the data,it is good for classification and visualization.And the difficulty of training the traditional multi-layer artificial neural network,can be solved by "layer by layer initialization".Those make the deep neural networks become a new field of machine learning,and is widely used in field of academia and industry.In this dissertation,we introduce deep neural networks into the traditional change detection technology.First we use deep neural networks to solve the problem of synthetic aperture radar(SAR)image change detection.Then we propose positive and negative change detection problems,and we solve this problem by deep neural networks model and get good results.The main contributions of this dissertation can be reflected in the following two parts:(1)Aiming at the characteristics of deep neural networks technology,the change detection of synthetic aperture radar(SAR)image is studied.Based on unsupervised learning method and deep neural networks,the SAR image change detection based on sparse automatic encoder is proposed.This method mainly includes unsupervised learning feature extraction and fine tuning with supervised learning.By introducing pre-classification and sample selection techniques,the deep neural networks technique can also be applied in change detection without label data.Experiments were validated from a variety of different data sets,and satisfactory results were obtained.(2)Positive and negative change detection of SAR image are studied.Combined with the deep neural networks technique,positive and negative change detection of SAR image based on deep belief networks are proposed.In most dissertations,image change detection divides the difference image into two parts: the change part and the unchangeable part.However,the difference image of the multi-temporal SAR images has positive and negative change classes.Some image pixel brightness decrease and some image pixel brightness increase.Therefore,it is very meaningful to further divide the changed part into positive changed part and negative changed part.In this novel change detection,this dissertation uses the technology of deep belief networks to realize it.The influence of speckle noise on the detection of SAR image changes is reduced,the test results become more accurate and more robust to the noise.