Change Detection Of SAR Images Based On Deep Learning And Model Fusion

Change detection in remote sensing image is used to detect the change area in the same place at different time periods by appropriate algorithm.Because of the advantages of Synthetic Aperture Radar(SAR),which are not affected by illumination,weather and other conditions SAR image has been widely used in many fields such as environmental detection,disaster assessment,urban planning and monitoring.This paper mainly studies and analyses some original methods of SAR image change detection and improves them.At the same time,combined with the novel theory,we proposed some new methods to improve the accuracy of change detection.The main contents of this paper are as follows:1.Because the difference image(DI)generated by different operators have difference advantages and disadvantages,this paper will complement the advantages of the three difference DI produced by the difference method,log-ratio method and mean-ratio method.It is different from the method of manual selection strategy to fuse DI.In this method,three DI are input to three-channel convolutional neural networks as different feature maps.And then the convolutional neural network will automatically learn and extract the different features of each DI.The algorithm will select the pixels and their neighborhood space of different DI to generate training and test samples to train the convolution network.Then the trained network predicts the change class and the invariant class.2.To solve the problem of low accuracy of edge detection,we propose a multi-scale model fusion algorithm.Firstly,in this part a new DI generation algorithm is proposed.A new DI is obtained by gamma transformation of the DI generated by the Ratio operator.Both the transformation and log transformation can enhance the contrast of the image.The new DI is brighter than the logarithmic DI.The validity of our difference graph is proved by the analysis and comparison of several DI by FLICM algorithm.In change detection results,error rate of edge details is higher than others area.we propose an algorithm to solve this problem in this paper.The sliding window size of the generated sample is changed to one.That is to say,when generating samples,only the pixel information of four difference maps is used,and the corresponding neighborhood information is not extracted.Then classify by Multilayer Perceptron(MLP).This method is noisy because it does not use neighborhood information,but the advantage is that it can detect heterogeneous regions such as edges effectively.The edges of the test results are clear.Finally,weighted fusion is used between the model and the model of algorithm 1.The accuracyt of edge detection is improved after fusion.3.There are many parameters in the neural network.Aiming at the difficult problem of structure optimization and super-parameter selection of neural network.A neural network structure and hyperparametric optimization based on genetic algorithm is proposed.The algorithm optimizes the convolutional neural network and MLP used in change detection.There are two ways to combine evolutionary algorithm with neural network.One way is to use evolutionary algorithm to determine the connection weight when the network topology is fixed.In this paper,we use another method to optimize the structure of the neural network directly by genetic algorithm,and then train the network by back propagation algorithm.Firstly,the super-parameter optimization problem is built into an optimization model,and then the genetic algorithm is used to optimize the model.This method can find relatively optimal structure and hyperparameters for the model of change detection.And improve the accuracy of the algorithm.