Abstract: Mine scene data is a vital foundation for the construction of smart mines and their intelligent management. The rapid extraction and recognition of complex mine scenes using multi-source data, including remote sensing images, is a current research hotspot. This study employs Sentinel-2, GF6, GF2, and Google image data in conjunction with the deep learning target recognition algorithm Faster R-CNN to establish a mine scene recognition model. The key findings of this research include: (1) the collection of 2369 mine scene samples and the creation of 3989 labels for them, as well as the proposal of a methodological process for the intelligent recognition of mine scenes through multi-source data fusion; (2) the establishment of both a mine site scene recognition model (MSSRM) and a mining resource scene recognition model (MRSRM), with mean average precision values of 0.607 and 0.428 respectively, demonstrating the accuracy of the mining scene model built using Faster R-CNN as the base model framework; (3) a comparison of the precision to build a mine scene recognition model using single data versus multi-source fused data, with the latter showing a 17.2% improvement in precision; and (4) a comparison yielding a 9.1% higher accuracy for models built from remote sensing images with better than 2m spatial resolution compared to those built from images with 10m spatial resolution. This study proposes a novel recognition method for the intelligent development of mines, with significant theoretical and practical implications for scene recognition and sustainable development throughout the entire life cycle of mines.

Keywords: mine scene; multi-source data; deep learning algorithm; intelligent recognition; AI model