Research On SVM Classification Optimization Of Remote Sensing Image Based On Upper Confidence Bounds Strategy

With the development of remote sensing technology,remote sensing image data is becoming more and more abundant,and the application scenarios of remote sensing image are becoming more and more diversified,which poses a higher challenge to remote sensing image classification.Two of the most important classification methods among the existing pattern recognition theory are Artificial Neural Networks(ANN)and Support Vector Machine(SVM)methods.Although ANN has made outstanding achievements in many fields,its huge networks have brought about many problems such as difficulties in training and application.Therefore,SVM still has an irreplaceable position in many practical applications.The performance of SVM is minimally influenced by the form of its kernel function.However,research shows that it is significantly affected by the setting of its key factors.In recent years,more and more artificial intelligence(AI)methods have been used for global optimization,such as evolutionary algorithms(EA).However,these algorithms are not perfect,they have some problems such as high randomness and poor performance.In order to solve these problems and improve the accuracy of SVM,using Delaunay triangulation method for objects creating,this paper proposes a global optimization method for factors of SVM(cost C and kernel parameter σ)based on the Upper Confidence Bound(UCB)policy.The policy was first constructed in the study of multi-armed bandit problem,with a concise form and solid mathematical foundation.It utilizes all the information obtained in previous cycles,makes decisions taking into account function values and potential of exploration,and strikes a balance between exploration and exploitation.The proposed algorithm was tested on 6 UCI(University of California,Irvine)datasets and a hyperspectral remote sensing image of Xiongan New Area.Comparative experiments were conducted between the proposed algorithm and original SVMs of Matlab and ENVI platforms,as well as 5 well-known(modified)algorithms:Adaptive Inertia Weight Particle Swarm Optimization Algorithm(AIW-PSO),Modification Rate Artificial Bee Colony Algorithm(MR-ABC),Genetic Algorithm(GA),Memetic Algorithm(MA),and Butterfly Optimization Algorithm(BOA).In the experiment of UCI datasets,each algorithm was executed 30 times in order to calculate their average accuracy,optimization stability and other statistical indicators.The main conclusions are as below:(1)Choosing appropriate kernel parameters is essential for SVM classification.Compared with original SVMs,the accuracies of SVMs optimized by different algorithms are all significantly improved,and the average accuracy is increased by 35.70%;(2)Existing evolutionary algorithms usually have strong randomness,and the uncertainty of their optimization results is high.The standard deviations of the results of various evolutionary algorithms are all above 4,and their accuracies can vary as much as 50%to 70%;(3)The optimization stability of the proposed UCB algorithm far exceeds that of various evolutionary algorithms,the accuracies of it’s results vary only within 5%,and it’s average standard deviation is only 1.30.In addition to the optimization stability,the average accuracy of UCB is also better than that of the evolutionary algorithms,with an average accuracy of 3.85%ahead of them.And its leading degree relative to other algorithms is positively related to the difficulty of the SVM parameter optimization of the datasets.The UCB algorithm is both intelligent and stable in parameter optimization;(4)During the process of parameter optimization,the accuracy of SVM will remain basically stable after a rapid increase,and will not fall into serious overfitting.