Depth feature extraction for mineral mixed spectrum analysis based hyperspectral data

Rocks in nature are mostly aggregates of various minerals. Because of the intimate mixing of minerals, the rock spectrum measured by hyperspectral sensors is generally the mixing spectrum of mineral components. Influenced by the noise of measuring equipment, non-linear mixing of mineral spectra, rock structure and mineral cleavage, determining and quantifying the abundances of minerals is still an open problem. In this paper, a method based on depth feature extraction for spectral unmixing of minerals is proposed to solve the problems of estimating the number of mineral endmembers, extracting the endmembers spectrum and calculating the abundance of mineral. Firstly, hyperspectral signal subspace minimum error identification (HySIME) algorithm is used to calculate the number of mineral endmembers. Secondly, the depth neural network is constructed to extract the dimension reduction feature of hyperspectral data, and obtain the mineral endmembers spectrum. Finally, the single scattering albedo of the spectrum is calculated by the Hapke model, and the abundance is estimated according to the number of endmembers and the endmembers spectrum. Aiming at the common rock forming minerals and altered minerals, the spectral data measured in the laboratory are tested. The results show that the method and technical process proposed in this paper are superior to the commonly used spectral mixed analysis method.