Research On Remote Sensing Classification Based On Bayesian Algorithm Optimized Support Vector Machine

As a non-contact large-scale observation method,remote sensing technology plays an effective role in monitoring land cover change.Land cover change is considered to be one of the causes of global environmental change,so the study of land cover change has attracted extensive attention from experts and scholars.As the basis of remote sensing images analysis,the classification of ground objects is of great significance in remote sensing science and application.Among them,Support Vector Machine(SVM)shows superior performance in the task of ground object classification.However,because the SVM model cannot determine the optimal parameter combination during classification,the accuracy of ground object classification results will be affected to some extent.In this paper,the Sentinel-2 optical satellite remote sensing image data of Chongming Island,China,is used to construct Bayesian algorithm optimized SVM model to automatically obtain the optimal parameter combination for remote sensing image classification,so as to improve the classification accuracy of ground objects.Furthermore,the classification advantages and precision changes of the optimized model compared with other models were analyzed to provide scientific reference for the classification research of large area features.The specific work carried out is as follows:(1)In order to highlight the information differences between different features and enhance the ability of classifier to identify different features,the Sentinel-2 image feature information construction from 2017 to 2020 was firstly carried out in this paper.In addition to the commonly used normalized vegetation index(NDVI)and normalized water index(NDWI),three characteristic parameters were also extracted to enhance the differences between different vegetation types,namely,the ratio vegetation index(RVI),enhanced vegetation index(EVI)and color index(CI).Principal component analysis(PCA)was used to reduce the dimension of multiple texture features of Sentinel-2 four 10 m resolution images extracted by gray level co-occurrence matrix year by year,so as to enrich the information of different ground objects and reduce data redundancy,so as to improve classification efficiency and performance.(2)Based on the classification model of SVM,combined with the popular Bayesian algorithm in recent years,the optimal combination of model parameters(penalty coefficient C and kernel function parameter ?)of SVM was searched to improve the classification accuracy of remote sensing images.The results show that the precision of the SVM model trained by the optimal parameter combination reaches 94.33%,which is 3.07% higher than that trained by the default parameter combination,which provides an important technical support for its application in the classification of remote sensing images.(3)In order to study the universality of the new model and the classification effect of large area and complex features,the Sentinel-2 image bands set in Chongming Island area from 2017 to 2020 were used for classification experiments.In this paper,the performance indexes of Maximum Likelihood Method(MLC),Random Forest Method(RF),Support Vector Machine(SVM)and Optimized Support Vector Machine(BO-SVM)are compared.The results show that the classification accuracy of BO-SVM is generally higher than that of the other three methods,and the classification results of BO-SVM are 2.505%,2.935%,3.365% and 4.582% higher than the classification results of SVM from 2017 to 2020 respectively.It can be seen that this method has certain stability and universality in the classification of remote sensing images.