Enhanced hyperspectral classification for geological mapping based on spatial–spectral novel framework of vision transformers and mamba models on PRISMA data at Arabian–Nubian Shield

Abstract Lithological mapping of arid Precambrian basement underpins mineral exploration across the Arabian-Nubian Shield (ANS), where rugged, poorly accessible exposures make field campaigns slow and costly, and where serpentinites, ophiolitic mélange, granitoids and hydrothermal alteration of economic interest occur in close spatial association. Spaceborne hyperspectral imaging resolves the narrow mineral absorptions that distinguish these units and offers a route to rapid, reproducible regional reconnaissance. We present an open, end-to-end framework that converts PRISMA Level-2D imagery into an analysis-ready, MNF-augmented data cube and benchmarks four supervised classifiers Random Forest, XGBoost, SpectralFormer (a spectral transformer) and S 2 Mamba (a spatial-spectral state-space model) under a single preprocessing and evaluation protocol over the Wadi El-Gemal area, Central Eastern Desert of Egypt, for the first time in the ANS. Twelve lithological classes were trained from field- and petrography-constrained polygons. S 2 Mamba produced the most accurate and spatially coherent maps (overall accuracy 99.44%, mean IoU 98.45%, Cohen’s κ = 0.994; macro-averaged F1 = 0.992), with the clearest improvement over classical models in spectrally ambiguous alteration zones and deformed gneisses, where spatial context is decisive; a well-tuned XGBoost remained competitive with the transformer once paired with high-SNR MNF features. We show that such accuracies are strongly sensitive to spatial autocorrelation under pixel-random sampling and argue that spatially blocked validation is required before landscape-scale claims are made. The documented Python pipeline is released as open-source software and transfers readily to other imaging spectrometers and shield terranes, supporting operational mineral exploration targeting across the wider ANS.