Research On Hyperspectral Image Dimension Reduction And Endmember Extraction Algorithms

Remote Sensing(RS)is a comprehensive technology of earth observation developed in the 1960 s.It is widely used in agriculture,ocean,geological and military,etc.Using RS technology in agriculture encourages agricultural scientific decisions to a new level,provides high quality service for agricultural production.Modern agriculture RS technology lacks of effective means in disaster monitoring and production assessment.,which crop classification can make them possible.Hyperspectral remote sensing is one of the frontier technology means in agriculture.Hyperspectral data's high dimension causes time consuming and low classification accuracy when directly used.Therefore,this paper makes a in-depth study on hyperspectral data preprocessing.This paper summarizes the results of previous studies and a new algorithm based on tabu search algorithm and Compactness-Separation Coefficient(CS Coefficient)is developed to perform hyperspectral image's feature reduction(TSFR).It uses less optimization time to calculate optimal feature reduction number,and through experiment simulation,it shows that the proposed method is better than Monte Carlo feature reduction method(MCFR)in reduction time,feature reduction and classification accuracy.According to mixed pixel's existence,which influence classification accuracy,N-FINDR algorithm and endmember extraction algorithm based on high dimensional simplex volume are studied in this paper.Two improved endmember extraction algorithms are proposed based on endmember independence and thinning technology.One improved algorithm combines simplex volume calculation criteria and endmember independence as thinning criteria.The other improved algorithm uses high dimensional simplex volume calculation criteria and endmember independence as thinning criteria.Through several simulation experiments,results show that proposed algorithms can obtain endmembers with higher classification accuracy and lower RMSE accuracy compared with the two existed algorithms.