Fine-grained coal-density classification via hyperspectral imaging and deep learning

The pursuit of efficient and environmentally sustainable coal utilization has intensified the demand for accurate, density-based sorting methods. Conventional coal separation techniques, including dense-medium cycloning, float-sink testing, and manual picking, are hindered by limited throughput, high energy consumption, and adverse environmental impacts. Although hyperspectral imaging (HSI) has successfully improved binary coal – gangue discrimination by providing rich spectral features, its potential for direct, fine-grained coal-density classification has not been fully exploited. To bridge this gap, this study introduces an innovative hyperspectral deep-learning framework capable of classifying coal densities into five distinct categories. The proposed SpectralConvTran architecture integrates novel Gaussian Spectral Encoding (GSE) and Spectral Distance-Aware Attention (SDAA) modules, significantly enhancing the network’s sensitivity to critical wavelength intervals associated with coal density. Experimental validation on a five-class coal-density dataset demonstrates that SpectralConvTran achieves an overall accuracy of 93.94%, surpassing traditional 1D-CNNs and standard Transformer-based methods. Further spectral analysis identifies wavelength intervals of 794–803 nm and 970–1000 nm as particularly influential for density discrimination, with occlusion experiments confirming their essential roles. Consequently, the proposed method not only offers a robust and precise coal-density classification solution but also provides valuable insights into spectral mechanisms driving coal quality differentiation, paving the way for refined resource management in coal utilization.