Remote Sensing Image Change Detection Based On Multi-Objective Optimization And Deep Neural Network Structure Optimization

With the continuous development of science and technology in modern society,it has become an important way to monitor ground information by using radar remote sensing to obtain earth observation information at present.And Synthetic Aperture Radar(SAR)is an indispensable way in various radar remote sensing systems.SAR image change detection refers to the process of analyzing changes occurring in multi-temporal SAR images obtained from different times in the same region.Nowadays,SAR image change detection has been widely used in municipalities such as hydrology,agriculture,forestry,atmosphere,ocean and other environmental monitoring,urban construction and other municipal monitoring,military deployment,disaster assessment and other aspects.In this thesis,the following research works are carried out on the problems of low accuracy and poor adaptability of methods about SAR image change detection:(1)A method is proposed for SAR image change detection based on generative adversarial networks and evolutionary algorithm.In the part of preprocessing,the sparse autoencoder is used to learn the SAR images of different time phases,and corresponding feature images are obtained.Then the difference map can be received by change vector analysis.On the basis of the traditional evolutionary algorithm,consider the model of generative adversarial networks to achieve effective classification of the difference map.This method avoids the artificial influence of crossover and mutation operator in the traditional evolutionary algorithm while preserving the original image feature information as much as possible.Meanwhile,this method also shows better adaptability in the whole process of SAR image change detection.(2)A method is proposed for SAR image change detection based on generative adversarial networks and multi-objective optimization methods.This method combines the multi-objective optimization algorithm with the generative adversarial networks to classify the difference map.Based on preserving the main effective features of SAR images,this method introduces the evolutionary multi-objective optimization algorithm and considers the fuzzy clustering method to obtain the optimal aggregation after updating the dominant populations selected by generative adversarial networks.It improves the automation performance of the algorithm,further maintains the local and global search ability of the algorithm,and extends the application of the algorithm to the multi-classification problem which expands the application scope of the algorithm.(3)A method is proposed for SAR image change detection based on binary autoencoder generative adversarial networks and multi-objective optimization methods.On the basis of traditional generative adversarial networks,this method considers to reconstruct the part of generative model by the autoencoder,thereby obtaining the binary autoencoder generative adversarial networks.This novel model is combined with the evolutionary multi-objective optimization algorithm to classify the difference map.This method constructs an original generative adversarial networks model,which further considers the semantic information of input samples in order to deepen the learning effect of generative adversarial networks.Besides,it also improve the stability and accuracy of generative adversarial networks and the learning ability of generative adversarial networks at the probability distribution level,which improves the performance of the multi-objective optimization method.