Mapping Afghanistan’s Mineral Diversity With NASA EMIT Hyperspectral Data: Spatial–Spectral Transformer Versus CNN for Geo-Mineralogical Classification

Hyperspectral Imaging (HSI) is a crucial technology in mineral exploration, offering the potential to record and analyze spectral bands that reveal the actual composition of the features of the earth's surface. Although significant progress has been made in remote sensing–based mineral exploration, the availability of mineral data derived from hyperspectral sensors remains limited and not publicly available. Most existing data are available in raw unprocessed form, requiring complex, computationally intensive preprocessing to clean and refine that. Consequently, there is a pressing need for refined, geologically validated hyperspectral datasets to enable robust model training and accurate mineral identification. A novel Afghanistan-based hyperspectral dataset derived from the National Aeronautics and Space Administration (NASA) Earth Surface Mineral Dust Source Investigation (EMIT) was developed and rigorously pre processed for mineral classification. Deep Learning (DL) appli cations on hyperspectral data remain notably limited in state-of the-art literature; this study addresses this challenge by creating two DL models, a Convolutional Neural Network (CNN) and a Spatial-Spectral Transformer (SST), for hyperspectral image data applied to geological mapping in Afghanistan, a mineral rich yet underexplored diverse region. The model results validate that while CNN extracts spectral values, SST outperforms by using a multihead self-attention mechanism to capture long range dependencies and spatial details, achieving an overall accuracy of 89.24% compared to 87.71% for CNN. The SST shows superior performance and scalability, which is essential for accurate mineral classification. The advanced SST model proposed produces an adaptable framework for mineral explo ration in data-scarce terrain. Future research can investigate the potential of explainable Artificial Intelligence (AI) to improve interpretability and extend this framework to broader domains of geological surface exploration.