Accurate lithological mapping in semi-arid and geologically complex regions is crucial for advancing geoscientific understanding and guiding mineral exploration, yet it remains a technically challenging task. Here, we introduce the Inception-LSTM Hyperspectral Mapper (IL-HS), a deep learning framework designed to enhance lithological classification from hyperspectral satellite imagery. The model integrates InceptionV2 for multi-scale spatial feature extraction with a bidirectional long short-term memory (Bi-LSTM) module to capture sequential spectral information. Using EnMAP hyperspectral data over the Kerdous inlier, Anti-Atlas, Morocco, IL-HS achieved an overall accuracy (OA) of 98.05% across twenty-six lithological units, substantially outperforming state-of-the-art models, e.g., support vector machines (OA = 91.17%) and 3D convolutional neural networks (OA = 94.61%). The proposed model demonstrated strong class-wise performance, including perfect recall for copper and manganese-bearing formations in the Proterozoic and infra-Cambrian basement-cover border mineralizations, and reliably distinguished spectrally similar formations such as multi-age carbonates and volcanic units. These results show that IL-HS effectively mitigates spectral redundancy and mixing artifacts in heterogeneous extended age and altered terrains. Our findings pinpoint IL-HS as a robust and scalable approach for hyperspectral lithological classification and mapping using hyperspectral satellite data, with broad potential for applications in geoscientific research, mineral resource assessment, and sustainable exploration research.
Khandouch et al. (Sun,) studied this question.