A Novel Method for Improving the Accuracy of Hyperspectral Remote Sensing for Detecting Surface Minerals on the Earth

Abstract

Hyperspectral remote sensing technology has made remarkable progress and shows a huge application prospect, such as in mineral distribution detection. Hyperspectral unmixing is a key step in applying hyperspectral remote sensing to detect surface minerals, which extracts endmember spectrum of minerals and detects mineral distribution. Non-negative matrix factorization (NMF) has been introduced into hyperspectral unmixing in the last decade. To reduce the influence of the non-convexity of NMF on spectral unmixing accuracy, the paper proposes a novel hyperspectral unmixing model. The proposed method uses the kernel density function to estimate the intrinsic data structure of hyperspectral images and uses regularization to establish the relation between high-dimensional hyperspectral image and low-dimensional abundance matrix. The proposed method makes the decomposed abundance matrix preserve the hyperspectral data structure, which leads to a more desired spectral unmixing performance. The experimental results on real hyperspectral image prove the superiority of the proposed method in surface mineral detection compared with other typical spectral unmixing methods.