The Improvement And Optimization Of Endmembers Extraction In Hyperspectral Remote Sensing Image

Coming up with the conception of digital earth, hyperspectral remote sensingtechnology plays an increasingly important role in modern science and technologyand our daily life. Through hyperspectral remote sensing, it could get the informationof the surface, to process the image and explore the distribution of surface feature,supporting the first materials for the geographical mining, the geographical disasters,the environmental monitoring and the urban traffic. According to the hyperspectraldata, with extracting the pure pixel respecting single features, it could classify thesepixels. Therefore, the extracting endmembers is one of keys to hyperspectral remotesensing technology. The data feature of hyperspectral data having several hundreds ofbands result the huge amount computing. So how to analysis the data and extract theendmembers become one of key problems in hyperspectral remote sensing imageprocessing.Recently the mixed hyperspectral includes the linear and the nonlinear models,the usual algorithms of extracting endmembers always base on the linear model. Andthe linear model consists automatic methods and human computer interaction methods.The automatic methods always are easy to realize through computer, avoiding somemistakes caused by people. The human computer interaction methods could controlsome situations needing accordance. In total, both of theses two methods achievesome outcome and apply into these applied examples.This article analysis the high dimension data structure of hyperspectral,according to the requirement of reducing dimension. It studies PCA, MNF, SVD threealgorithms from different reducing dimension criterions, and illustrates the effectiveresult of reducing dimension through the procedure. On the foundation of needs toreducing dimension, it also studies these algorithms of extracting endmembers,figures out the defects, improves PPI, N-FINDR and IEA methods. It transforms therandom project direction into fixed project direction, solves the unstable matrix determinant of N-FINDR, reduces the proceeding time of IEA. Resultantly itimproves the outcome of the extracting endmembers, compares these differencesamong these three algorithms, figures out that the PPI extracts more endmembers thanN-FINDR and IEA. Finally, it designs some parallel algorithms of some extractingendmembers methods.