Research And Application Of Endmember Extraction In Hyperspectral Remote Sensing

The core of hyperspectral remote sensing technology is based on feature of electromagnetic wave emission, absorption and reflection properties. Hyperspectral remote sensing is an implementation of new remote sensing technology to maximize spectral resolution,and use spectral differences as the core features for fine classification. Because of its high spectral resolution, the band number, and large amount of data, the hyperspectral remote sensing data’s spatial resolution is often not enough. Mixed pixels are Common in hyperspectral remote sensing image. It hindered the classification accuracy to some degree. To further improve the hyperspectral remote sensing classification accuracy must face the problem decomposition of mixed pixels. The procedure of unmixing is divided into two steps:extraction of endmembers and spectral unmixing. This paper is mainly about extraction of endmembers, uses them unmixing the image by model of unmixing and analyzes the result.The distribution of the pixels in n-dimensional space is convex geometry. This is the theoretical basis that endmembers are located in the convex polygons vertex position. This paper introduces the existing spectral mixture models and extraction methods, and points out the advantages and disadvantages of each models and methods. Then using PPI to extract endmembers and unmixing images by linear spectral separation.Studying the features of the endmembers and obtaining a rapid identification method of the initial components based on insufficiency of the maximum distance method. Using PPI and spectrum analysis tool verify five endmembers by this method. Then using them unmix the image.Considering the error inside image, so we introduce the concept of distance threshold. The pixels which distance between endmembers is less to the threshold make a sample. Averaging the spectrums of every sample and matching these spectrums with USGS. It proves that the mean spectrums are more similar to true spectrums than the ones from single point. The experiments show that using the mean spectrums can improve the reliability of the endmembers and the accuracy of unmixing by comparing the results of three times.