MA-YOLOv8 Algorithm for Mining Area Object Detection Based on High-Resolution Remote Sensing Images

Remote sensing object recognition has been applied in the ecological environment of mining areas, exploring its potential applications in remote sensing target classification, detection, and monitoring of land surface cover and land use changes. However, existing target detection algorithms have not fully utilized the rich detailed information in remote sensing images. To exploit the feature information of remote sensing targets at different scales and achieve precise detection of ground objects, we propose the MA-YOLOv8 algorithm for ground object detection in open-pit mining areas based on high-resolution remote sensing images. Firstly, addressing the issue of insufficient multi-scale target detection capability of the YOLOv8 model on the MAOD dataset, this paper proposes two improvement structures: serialized ASPP structure and Dilated Spatial Pyramid Position Pooling structure (ASPP-SP). These two improvement structures can capture features of ground objects at different scales. Secondly, addressing the problem of missed and false detections of deformable targets in the MAOD dataset, this paper proposes the DC2f structure, which significantly improves the model's detection capability for deformable targets. By integrating these three improvements, the MA-YOLOv8 model is proposed, which effectively enhances the performance of multi-scale target detection, utilization of positional information, and detection of deformable targets. This study employs remote sensing methods such as unmanned aerial vehicles and satellites to collect data on open-pit mining areas and surrounding areas and constructs the MAOD dataset based on open-pit mining areas. Compared with YOLOv8, our proposed algorithm improves mAP@0.5 from 0.880 to 0.894 and mAP@0.5:0.95 from 0.658 to 0.692.