A New Approach for Mineral Mapping Using Drill-Core Hyperspectral Image

Abstract

Hyperspectral remote sensing technology has been successfully applied to geological fields. Drill-core hyperspectral imagery has the characteristics of segmented processing and large data volume. Due to simple principle and high accuracy, the Spectral Angle Mapping (SAM) has become the most commonly used method for mineral mapping using drill-core hyperspectral images. However, SAM analyzes the entire spectral form of minerals, and is not sensitive enough to small differences in drill-core mineral spectra. Compared with traditional machine learning methods, deep learning has more powerful feature learning and feature expression capabilities. In order to improve the mineral mapping accuracy, this paper proposes a new approach called Graph Convolutional Neural Networks-Spectral Angle Mapping (GCNNSAM), which integrates the advantages of deep learning and spectral matching to extract mineral information from drill-core hyperspectral images. Taking the drill-core hyperspectral data near the depth of 240m as an example, this paper compares the performances of SAM, GCNN and GCNNSAM mapping methods. The results show that the overall accuracy of the GCNNSAM mapping is 89.23%, and the overall accuracies of SAM and GCNN mapping methods are 80.25%, 83.58%, respectively. Comparing the mineral mapping statistical results of GCNNSAM with the measured geological statistical results, the maximum statistical error of mineral relative content is 1.4%, and the errors are all less than 2%, which verifies the reliability of the proposed method in this study. The method provides a new idea for mineral information acquisition in geological research.