Species Identincation Study Based On Dimensionality Reduction Hyperion Data

Experiments based on the Hyperion images have been performanced to study the species identification both in the XiaYang and ShaXi countrys of MingXi County in Fujian province.Firstly, the characteristics of Hyperion hyperspectral data were analyzed, preprocessing methods based on the Hyperion hyperspectral defects were proposed. Due to the large amount and high data redundancy of hyperspectral data, MNF and IC A transforms were performed to reduce dimension.Then,20principal component of dimension reduction were to classify data using MLC, SAM and SVM methods in order to three kinds of classification methods and6kinds of composed methods. Confusion matrix analysis was used to analysis classification result.Finally, that classification accuracy with the change of number of principal components of6composed methods was studied. The results showed below:(1) When the number of principal components is equal to20, classification accuracy affects of MNF and ICA transform on the same classification method can be ignored. Classification accuracy of three classification methods and Kappa coefficient declined in the order of SVM (77%,0.76)> SAM (64%,0.63)> MLC (52%,0.50), the the decline of classification accuracy is approximately12%.(2) Accuracy of the SVM and SAM increased with increasing quantities of principal components, nevertheless the MLC continued to decrease. When the principal components increased to a certain degree, the classification accuracy of SVM and SAM tended to saturation, the precision of SVM was higher than SAM.(3) Because the MNF transformation of the classification accuracy are dominantly decided by the first principal component, the effect of MNF transformation on data classification result was more stable, however stability was not very good after ICA transform.(4) The accuracy of MNF under three classification methods was higher than ICA when the first three principal components were took.