HTransNet: A Hierarchical Transformer Network with Dual Attention for Large-Scale Mining Scene Classification Using Multi-Category HighResolution Remote Sensing Dataset

Remote sensing-based mining scene classification plays a pivotal role in monitoring mineral extraction activities and evaluating sustainable development practices. Nevertheless， this critical task confronts two principal challenges: (1) the scarcity of diversified remote sensing datasets encompassing multiple mining types， and (2) the limitations of conventional models in processing complex spatial characteristics inherent to mining landscapes， including sparse target distribution， multi-scale variations， high inter-class similarity， and heterogeneous feature mixing. To address these limitations， this study presents CUG_MA， a novel large-scale mining scene dataset characterized by multi-category coverage and high spatial resolution (0.5-2 m). Comprising 2，649 annotated samples across nine representative mining types from Hubei， Jiangxi， and Heilongjiang provinces， this dataset systematically captures diverse geological and geographical characteristics of Chinese mining regions. We further propose a hierarchical Transformer architecture (HTransNet) incorporating dual attention mechanisms for enhanced mining scene recognition. The model employs multi-head self-attention to establish global contextual relationships across varying scales while utilizing cross-level attention fusion to hierarchically integrate multi-stage features， effectively addressing scale discrepancies and enhancing discriminative feature representation. Comprehensive experiments demonstrate that HTransNet achieves superior performance with 69.4% mean accuracy on CUG_MA， outperforming eight state-of-the-art models by a 5.0% margin. Notably， all benchmark models attained acceptable classification performance， substantiating the dataset's robustness and generalization capability. This research provides both a valuable benchmark dataset and an advanced framework for intelligent monitoring of mining ecosystems.