Classification Of Remote Sensing Images Based On Multiple Classifier Ensemble

Remote sensing image classification plays a very important role in forest vegetation monitoring,climate change and ocean exploration.Because of the complexity of the distribution of ground objects in remote sensing images and the different application backgrounds,The emphases and difficulties of this study are to improve the classification of ground objects.Traditional remote sensing image classification algorithm has been mature and widely used.The classification accuracy of remote sensing image has an inseparable relationship with many factors such as samples selection,multi-temporal remote sensing images selection,characteristics used for classification and classification algorithm.Therefore,how to use existing well-rounded technique approach and mine the rich information in remote sensing data and improve the classification accuracy of remote sensing images has always been one of the hot issues in study.In this paper,two kinds of remote sensing data with different spatial resolutions are used as data sources,and Dehui,Changchun City,Jilin Province is taken as the research area to study the classification method of remote sensing images.The single temporal remote sensing image has the problem of "the same object has different spectrum,and the same spectrum has refers to the different objects",which limit the improvement of classification accuracy.To solve this problem,this paper comprehensively considers the dominant information sensitive to ground objects in each time phase,and selects multi-scene remote sensing images for classification.Jeffrerys-Matusita Distance was used to calculated the separation between different classes.The results show that the degree of separation between different objects gradually increases with the increase of remote sensing images.Which show that the multi-temporal images with different spectral characteristic of various features combination is beneficial to enhance the differentiation of various features.Considering the data dimension and computational efficiency,The vegetation index and texture features of the best time phase were selected as auxiliary data for the research of remote sensing images classification and the feature set for classification was constructed.Six classification algorithms were selected for classification,According to the comparative analysis of the classification accuracy of a single ground object,K-Nearest Neighbor(KNN)has the highest classification accuracy of building land based on remote sensing images.Support Vector Machine(SVM)and Artificial neural network(ANN)were the best for corn,rice and water classification.By comparing the overall accuracy,average accuracy and Kappa coefficient of each classifier,KNN and SVM have the best classification effect.SVM has relatively stable performance and it can be complementary to KNN algorithm in classification.Based on the classical classification algorithm and ensemble algorithm theory,By comparing and analyzing the classification results of different classifiers,KNN and SVM classifiers with better classification effect were selected,the KNN?SVM ensemble classification algorithm is constructed for remote sensing image classification by using weighted voting method.The experimental results show that the KNN?SVM algorithm improves the overall accuracy,average accuracy and Kappa coefficient of the two kinds of remote sensing data with different resolutions compared with the single classifier,random forest and Adaboost algorithm.The experiment proves the effectiveness of the algorithm in improving the image classification.However,the combination number,combination mode and performance optimization of multiple classifiers remain to be further studied.