Estimation of the number of endmembers in hyperspectral data using a weight-sequence geometry method

Abstract

The terrestrial reflection or emission spectrum obtained by the remote sensor is recorded in units of pixels. In most cases, a pixel usually contains many types of terrains. This pixel is a mixed pixel, and each of the terrains in the mixed pixels is called “endmember”. Estimating the number of endmembers is a significant step in many hyperspectral data mining techniques, such as target classification and endmember extraction. The paper proposes a separative detection method by the use of a weight-sequence geometry to estimate the number of endmembers. This method projects the spectral matrix into the orthogonal subspace by eigenvalue decomposition at first. Then, on the basis of the normalized eigenvalue sequence, the separative detection method innovatively uses a geometric criterion to find the separation point between the main factors and minor factors. Finally, the number of endmembers is determined by the sequence of the “separation point”. Validation through a series of simulated and real hyperspectral data, it indicates that the proposed method can accurately and rapidly detect the number of endmembers in the hyperspectral data without any prior information. In addition, the new method is also applicable to the ultra-high resolution remote spectral data in the future.