Continental-Scale Open-Pit Mine Detection From Remote Sensing Imagery: A Lightweight Attention-Enhanced Approach With High-Resolution Benchmark Dataset

Accurate and efficient identification of open-pit mining areas from remote sensing imagery is critical for global resource management and environmental monitoring. However, existing datasets are either low-resolution, region-limited, or incompatible with deep-learning pipelines, while current detection models struggle with complex backgrounds, multi-scale targets, and real-time inference. To address these gaps, we present MAM-YOLOv10n, a lightweight yet powerful detector that integrates CBAM (Convolutional Block Attention Module) and C3TR (C3 Transformer) modules into YOLOv10n. This design boosts global context modeling and suppresses background clutter, yielding +1.71% mAP50 and +1.12% mAP75 over the baseline with negligible runtime cost. Concurrently, we release the North American Open-Pit Mine Target Detection dataset—the high-resolution (0.1–3 m), continent-scale benchmark covering 667 manually annotated mines across diverse mineral types, provided in both COCO and YOLO formats for immediate training use. Extensive experiments against YOLOv10n, RT-DETR (Real-Time Detection Transformer), Faster R-CNN, EfficientNet, and SFAF (Scale-Aware Feature Aggregation Framework) demonstrate that MAM-YOLOv10n achieves the best accuracy–efficiency trade-off.