Lithological classification using hyperspectral remote sensing, particularly the Advanced Hyperspectral Imager (AHSI) onboard the Gaofen-5B (GF-5B) satellite, plays a crucial role in geological mapping and mineral exploration due to its rich spectral information. In regions characterized by diverse lithological types and complex spatial distributions, lithological classification has traditionally relied on algorithms based on machine learning (ML) and convolutional neural networks (CNNs). However, these approaches are unable to comprehensively extract the multidimensional features required for lithological classification, resulting in limited performance in distinguishing multiple lithological units. To address these challenges, we propose a Multiscale-Edge-Global Fusion Network (MEGFNet) that employs a dual-branch architecture. The multiscale-edge feature learning branch extracts rich spatial and edge features of lithological units, while the global feature learning branch captures long-range contextual semantics. Subsequently, a feature fusion module (FFM) integrates complementary information from both branches to enhance the model’s representational capability. The Shibanjing area, located in Ejin Qi, Alxa Meng, Nei Mongol Zizhiqu, was selected as the experimental study site due to its diverse and complex lithological distributions. Experimental results based on GF-5B AHSI imagery demonstrate that MEGFNet effectively integrates multidimensional features to distinguish lithological types and delineate their spatial distribution. The proposed method achieves an overall classification accuracy of 95.9%, with particularly notable improvements in identifying small-scale lithological units.
Liu et al. (Mon,) studied this question.