The Research Of Endmember Extraction Approaches From Hyperspectral Image Based On The Linear Mixed Model

The linear mixed mixture is a commonly accepted model for hyperspectral data processing. As the important parameter of the linear mixed mixture, endmember represents a certain ground component whose spectral is changeless comparatively. Recently, endmember has been playing a key role in spectral unmixing, anomaly detection, classification and so on. This thesis focuses on analyzing the characteristics of the linear mixed model, and puts great emphasis on the study of extracting endmembers from hyperspectral image based on it.According to the mechanism of hyperspectral image, the linear mixed model has proved to be appropriate for the purpose of representing the synthesis of mixed pixels from distinct endmembers. Based on the principle of the linear mixed model, the physical and geometric meanings of endmembers are analyzed. The performances of extracting geometry vertices and extracting mean spectrums are compared, and then, an endmember extraction algorithm based on RMS error analysis is presented. In this way, more abundant information can be provided in the following data processing. Though, these methods focus exclusively on the spectral nature of the data, few has exploited the existing correlation between neighboring pixels. In the view of the importance of analyzing spatial and spectral patterns simultaneously, a new approach based on mathematical morphology to perform endmember extraction from hyperspectral data is proposed. It uses both spectral and spatical information in a simultaneous manner and improves the reliability of endmember extraction. At last, two different strategies are presented respectively aimed at whether the prior knowledge exists or not, a comparative study between these methods is accomplished. The obtained results indicate that endmember extraction of the linear mixed model is more successful when the spatial and spectral information are combined.