Change Detection In Synthetic Aperture Radar Images Based On Multiobjective Optimization And Deep Learning

Synthetic Aperture Radar(SAR),which is a kind of advanced active microwave earth observation equipment,is able to penetrate cloud vegetation and not easily affected by outside environment such as weather conditions.Also,the SAR has the characteristics of long-distance,all-weather and all-time.Change detection of remote sensing image is based on these applications.Remote sensing image change detection firstly quantitatively analyzes the remote sensing images obtained from multiple time phases over the same region,and then obtains the interested information of the region change.This paper firstly discusses the SAR image change detection methods and then proposes three change detection algorithms.The main contributions of this paper can be summarized as follows:1.A change detection algorithm for SAR image based on neighborhood ratio and self-paced learning is proposed.The algorithm combines neighborhood ratio and self-paced learning algorithm.The neighborhood ratio method can effectively utilize the neighborhood pixels information of each pixel in SAR image.The self-paced learning mechanism could make the deep belief network extract feature from “simple samples” to “complex samples”,which gradually promote the learning ability of the network.Simultaneously,the effect of noise on change detection results could be weakened by adopting self-paced learning mechanism.In order to test the performance of this algorithm,we use three sets of real SAR images to test,and some comparisons between proposed algorithm and the other three change detection algorithms are conducted.The experimental results show that our algorithm has high detection accuracy and our experimental results are the closest to the standard reference graph.2.A change detection algorithm for SAR image based on network cost function optimization is proposed.Our algorithm constructs a single objective optimization model,which sets the automatic encoder network weights as the decision variables,the energy loss and the hidden layers' sparse as the multi-objective optimization goal.Then each hidden layer is optimized through particle swarm multi-objective optimization algorithm to guarantee that each hidden layer could extract effective feature.To validate the effectiveness of our proposed algorithm,we choose three sets of real SAR images to test,and we conduct some comparisons between proposed algorithm and the other five change detection algorithms.The experimental results show that our algorithm has higher detection accuracy.3.A change detection algorithm for SAR image based on network weight grouping is proposed.In training each hidden layer,the model is not directly using the particle swarm optimization,but to group the network weights and optimize the network with improved particle swarm multi-objective optimization algorithm,strengthening information exchange between groups.It could avoid premature convergence when optimizing the hidden layer by using the standard particle swarm optimization algorithm.In order to test the performance of the algorithm,we also have made some comparisons between our proposed algorithm with the other six change detection algorithms on three groups of real SAR images.The experimental results show that the proposed algorithm is more able to reflect real change over the surface of the earth.