Non-destructive, fast and intelligent identification of coal and gangue via spatial-spectral fusion of hyperspectral images

Abstract

Coal and gangue identification is a crucial part of the intelligent and green coal industry. Traditional coal and gangue sorting has problems such as low efficiency, poor accuracy, and limited applicability. This article proposes a coal and gangue identification method based on the MambaHSI model, leveraging the non-destructive, fast, and information-abundant attributes of hyperspectral imaging technology. Firstly, the coal and gangue hyperspectral images undergo preprocessing procedures, including image stretching enhancement, data dimensionality reduction, and normalization, with the aim of enhancing image quality and data processing efficiency. Next, in order to fully utilize the advantages of the “spectrum integration” of hyperspectral imaging technology, the model takes the hyperspectral images of the entire coal and gangue as input. Using an end-to-end approach for training, explore the differences in texture or local distribution of coal and gangue in two-dimensional space, as well as the differences in reflection characteristics in different bands of one-dimensional spectra. The Spatial Feature Extraction Module is dedicated to discerning the long-distance dependence relationships at the pixel level, thereby enabling the capture of the spatial distribution coherence of coal and the interrelationships among neighboring pixels. The Spectral Feature Extraction Module segments the spectral vectors of coal and gangue into multiple spectral groups and delves into the relationships between disparate spectral groups. The Spatial-Spectral Feature Fusion Module adaptively integrates the spatial and spectral information of coal and gangue. Finally, the proposed method is applied to hyperspectral image datasets of coal and gangue originating from three distinct sources. The results show that the classification performance based on the MambaHSI model is excellent, and the overall accuracy of coal and gangue classification can reach up to 99.65 %; The highest average accuracy can reach 99.62 %; The Kappa coefficient can reach up to 100 %; The mean Intersection over Union can reach up to 99.56. This method has the characteristics of high identification accuracy, good real-time performance, and strong robustness. The results of this study can be used for in-situ, non-destructive, and intelligent identification of coal and gangue underground in mining areas, promoting the rapid development of intelligent coal gangue separation.