Research on ore classification based on multimodal information fusion of LIBS spectral and image features.

Reliable ore classification is essential for mineral exploration and intelligent beneficiation and directly determines the efficiency of mineral resource utilization. Conventional single-modal detection methods exhibit inherent deficiencies: laser-induced breakdown spectroscopy (LIBS) can provide elemental fingerprint information but is vulnerable to sample surface contamination, whereas machine vision acquires mineral morphological features yet fails to effectively distinguish minerals with similar chemical compositions. Moreover, existing multimodal fusion frameworks lack adaptive and robust fusion strategies for heterogeneous modal data, severely degrading model stability and classification accuracy under complex industrial mining conditions. To overcome these limitations, this work proposes a micro-destructive ore classification method integrating LIBS elemental spectroscopy and machine vision morphology. A DPSE adaptive multimodal fusion module is developed for one-dimensional heterogeneous concatenated embeddings. Different from vanilla SE and CBAM, DPSE introduces stacked one-dimensional, depthwise-separable convolution layers prior to the attention stage to explicitly extract local cross-modal correlations, together with a post-attention feature-reconstruction branch for dimensional alignment with dual-stream feature extractors, so as to suppress environmental interference and strengthen feature discrimination. Experimental validation on 17 natural mineral types achieves a classification accuracy of 96.13%, outperforming standalone LIBS and machine vision methods by 10.47% and 6.79%, respectively. The proposed method provides an accurate, rapid, and robust in situ mineral analysis strategy for intelligent mining applications.