Research on Core Hyperspectral Mineral Identification Technology Based on Spectral-Spatial Hybrid Convolutional Neural Networks

The traditional core logging method based on manual experience has a low recognition efficiency, while the spectral matching method based on spectral similarity in core scanning ignores the spatial structure characteristics of the core. To address these issues, this paper takes the hyperspectral images of core samples from the Baertu copper mining area in Xinjiang as the research object and introduces a hybrid convolutional neural network (HybridSN) based on spectral-spatial joint modeling to conduct research on mineral identification and alteration information extraction at the core scale. Firstly, hyperspectral data obtained by the hyperspectral core scanning system are utilized, and combined with the TSG interpretation results and rock and mineral identification data, a pixel-level sample set is constructed, including 11 categories of samples such as major minerals and background types. Based on this, a HybridSN model based on a 3D-2D hybrid convolutional structure was constructed to automatically extract and classify the spectral-spatial joint features of core hyperspectral images. The experimental results show that the overall accuracy (OA) of the model on the test set is 96.15%, the average accuracy (AA) is 93.52%, and the Kappa coefficient is 0.9568. Compared with 2DCNN, HybridSN shows obvious advantages in prediction accuracy for most mineral categories, and the accuracy fluctuation between categories is small, demonstrating more stable classification performance. The comparison with the rock and mineral identification results indicates that the classification results of the HybridSN model are highly consistent with the actual geological features in terms of mineral type identification and the spatial distribution of alteration zones, indicating that this method has good applicability and application potential in the fine mineral identification and automatic extraction of alteration information of core samples, and can provide reliable technical support for digital core analysis and intelligent geological logging.