Research On Dimension Reduction And Classification Algorithm For Hyperspectral Remote Sensing Image Data

With the development of remote sensing technology,hyperspectral remote sensing as a new way of remote sensing has become an important means for people to obtain surface information,and in military and civilian areas are playing an extremely important role.Hyperspectral image classification is an important direction of hyperspectral information processing.The classification algorithm with high precision is a prerequisite for realizing various applications.However,hyperspectral images have high resolution,multi-band number and large data volume bring great challenges to traditional image classification technology.Aiming at the problems of hyperspectral remote sensing image classification,this paper mainly studies hyperspectral remote sensing image data dimensionality reduction and classification algorithm.Random forest is classification and prediction model proposed in 2001 by Breiman.As one of the integrated learning,the random forest algorithm makes full use of single classifier and statistical sampling technique to solve the problem that the single classifier can not be improved in performance.But the algorithm has some shortcomings to be perfected,and it can also improve the performance of different data sets and classification accuracy.According to the shortage in the random forest,this paper will propose an improved algorithm,from the data dimension reduction aspects to further improve the accuracy of random forest classification.In the early stage,we use the efficient dimension reduction algorithm to select the feature subsets with large information,small correlation and strong separability between classes,and then classify them in the feature subset space.This classification process can not only improve the efficiency of object recognition and classification,but also greatly improve the speed of processing data.In this paper,the concrete scheme of the algorithm is divided into two steps.The first step is to construct the feature space,combine the projection pursuit and the principal component analysis to get the new feature which form a new feature combination.The principal component analysis mainly reveals the linear distribution of hyperspectral data,which represents the whole distribution direction of the original data.While the projection pursuit reveals the local distribution of the data.The combination of global andlocal will improve the overall classification accuracy of hyperspectral remote sensing data.The second step is to use random forests for classification in new feature spaces.Experiments show that,compared with other classification algorithms,this algorithm has obvious advantages.Based on the study of random forest algorithm,this paper proposes an improved algorithm for the shortcomings of the algorithm.The improved algorithm not only reduces the dimension of the sample data,but also retains the main information and improves the classification accuracy of the random forest.Finally,the article analyzes the advantages and disadvantages of this algorithm in hyperspectral remote sensing image classification,which laid a solid foundation for further research in the future.