Abstract: To systematically elucidate the research status and technical pathways of near-infrared hyperspectral mineral
intelligent identification driven by deep learning,this paper begins with a bibliometric analysis of 1 362 publications (2014—
2024),revealing the research hotspots and growth trends in this field,with ″hyperspectral imaging″ and ″convolutional neural
network″ as the core. Furthermore,it comprehensively reviews the technological evolution from traditional machine learning to
deep learning,and dissects the key workflow encompassing feature extraction,model construction,and evaluation. Through performance
comparisons and case studies of mainstream models such as CNN,ResNet,U-Net,and MineralNet,it is found that targeted
algorithm improvements (e. g. ,introducing attention mechanisms) can significantly enhance model performance,while
multi-modal fusion strategies effectively improve identification accuracy. However,current techniques still face challenges including
the scarcity of high-quality labeled data,and insufficient model generalization and interpretability. Finally,future directions
are prospected,including developing few-shot learning paradigms,constructing integrated ″spatial-spectral″ end-to-end systems,
and enhancing the physical interpretability of models,to promote the deeper application of this technology in fields such as geological
exploration and extraterrestrial detection.