MIRN: A Global Context and Progressive Feature Alignment Network for Oriented Object Detection of Multimineral in Mining Areas

Intelligent mining area detection is fundamental for ecological monitoring and sustainable development of mining areas. However, current remote sensing detection lacks high-quality multimineral rotated object detection datasets, hindering deep learning model training. This article constructs a high-resolution remote sensing dataset (CUG-MROD) covering six mineral types in Australia, comprising 1911 images and 2581 annotated instances. The dataset features extensive coverage, high resolution, and precise rotated annotations, validated by ten mainstream detection algorithms. To address challenges of large-scale variations and complex background interference, we propose a mining intelligent recognition network (MIRN). The network incorporates multihead self-attention for global multiscale feature extraction and designs an innovative neck structure multilevel progressive feature alignment with two key modules: first, progressive multilevel feature fusion module, fusing hierarchical features through convolution and deconvolution operations to alleviate cross-scale feature loss, second, adaptive perception feature alignment block, achieving feature alignment via feature selection and dynamic convolution with gating mechanisms for semantic information weighting. Experimental results demonstrate MIRN achieves 69.30% mean average precision, improving up to 14.89% over mainstream methods. This study constructs a high-resolution rotated object detection dataset encompassing multiple mineral types and innovatively proposes MIRN network, providing essential data foundations and algorithmic support for intelligent ecological monitoring and sustainable development assessment of mining areas.