Tailings dam management is increasingly critical as mining operations extend to greater depths and generate more complex waste streams. To support scalable monitoring, this study establishes a harmonized benchmark of YOLOv5, YOLOv6, YOLOv8, and YOLOv10 for automated detection of tailings dams using high‐resolution satellite imagery. All models were trained and evaluated under aligned protocols, eliminating inconsistencies that hinder comparability in prior work. Building on these baselines, we introduce YOLOv10 Refined, which incorporates a VoVGSCSP backbone, a multi‐channel attention (MCA) mechanism, and mpDIoU‐style supervision to enhance feature representation and improve robustness in cluttered environments. On the test set, YOLOv10 Refined achieved a precision of 0.851, recall of 0.837, mAP50 of 0.923, and mAP50–95 of 0.722, surpassing the YOLOv10 baseline in mAP50–95 and F1 while maintaining competitive recall. These improvements are stable and reproducible rather than dramatic leaps, underscoring their practical value in workflows where consistency is paramount. The findings clarify trade‐offs across YOLO versions: YOLOv10 offers slightly higher recall for wide‐area screening, while YOLOv10 Refined provides stronger balance and reduced overfitting, making it more suitable for confirmatory analysis. Detection is positioned as the first stage in an early‐warning pipeline—supporting triage of large image streams and integration with downstream inspection, temporal analysis, and risk assessment—rather than a stand‐alone solution. Limitations include the modest number of unique source scenes and the absence of ablation experiments; future work should extend cross‐sensor generalization and multi‐temporal integration. Together, the benchmark and refined model provide a reproducible reference for applying YOLO‐family detectors to safety‐critical monitoring of tailings dams.
Yu et al. (Thu,) studied this question.