Feature Dimensionality Reduction And Hierarchical Classification Of Remote Sensing Image

With the rapid development of remote sensing technology,the information extracted during the image interpretation process is more diverse.In order to improve the accuracy of classification of features,more feature information is often added,which often results in certain information redundancy,and so that resulting in lower classification efficiency and even lower accuracy.In this paper,we use random forest(RF)and support vector machine(SVM)classifiers to explore a method of ensuring classification accuracy and reducing feature dimensions in remote sensing classification process.Taking part of Fuxing Forest Farm in Antu County of Jilin Province as research area,using the 2015 Landsat-8 image as the data source,extracting spectral information(red,green,blue,near-infrared and short-wave infrared bands),vegetation index(NDVI,enhancement vegetation index,ratio vegetation index and bare soil vegetation index),texture(homogeneity,mean,second monent,variance,dissimilarity,contrast,entropy and correlation)and topographic information(slope and aspect)a total of 19 indicators were used as classification feature variables.Taking the feature selection based on the estimation importance of RF classifier is used as a comparison,feature selection was based on the classification accuracy of a single feature in RF and SVM classifier respectively,and divided the selected features into two conditions according to whether the selected features were analyzed by principal component analysis or not.Then,using RF and SVM classifiers to classify,evaluating the classification accuracy,and determining the optimal features and classifiers combinations.Finally,the above classification results were further optimized and classified based on the classification accuracy of land types and the traditional rule-based hierarchical classification method.The results show:(1)Feature dimensionality reduction based on single feature classification accuracy,which can reduce the feature dimension while ensuring classification accuracy.In lower dimensions,the classification accuracy of features selected based on this method is more stable than that selected based on feature importance.(2)Features selected based on single feature classification accuracy are different by different classifiers,and classification accuracy is also different.Using multiple classifiers to cross-select features and classify(in this paper,only two classifiers are used to experiment,so this means using one classifier to select features and using another classifier to classify),which is better than using a single classifier to select features and classify.(3)In middle and low dimensions,classification accuracy of RF classifier may be related to the input order of features.Principal component analysis to input features is helpful to improve classification accuracy and stability of the classifier.(4)Classification accuracy of features selected based on feature importance,and then classified layered by classification accuracy of ground objects,is improved compared with the classification before hierarchy.(5)The rule-based hierarchical classification method is slightly inadequate in distinguishing coniferous forests from broad-leaved forests,and its overall classification effect is less than that of hierarchical classification based on the accuracy of terrain classification.(6)Compared with the hierarchical classification based on feature importance,the hierarchical classification based on single feature classification accuracy of ground objects is slightly less effective.