Multi-scale Feature Fusion and Threshold-based Attentional YOLO for Tailings Ponds Detection of Remote Sensing Images

Tailings ponds are a significant source of environmental pollution and present potential danger, making it a top priority to obtain accurate location information. Traditional object-based remote sensing techniques suffer from low efficiency and lack of automation. Furthermore, deep learning research on tailings ponds lacks practical models and datasets. To address these gaps, this letter proposes a novel approach - Multi-scale Feature Fusion and Threshold-based Attentional YOLO (MFTA-YOLO) for detecting tailings ponds in remote sensing images. Specifically, CSPDarknet53 is used as the feature extractor in MFTA-YOLO, and it combines the Multi-scale Feature Fusion Module (MF) and a new attention mechanism Threshold-based Attention (TA). MF effectively integrates multi-scale semantic features to detect varying sizes of objects. TA recalibrates input contributions based on the threshold gap, allowing better focus on regions and channels of interest. Additionally, a large-scale tailings ponds dataset, consisting of 3619 high-resolution (1.07m) maps sourced from Google Earth, has been constructed. Extensive experiments on this dataset have demonstrated that MFTA-YOLO outperforms most existing methods for tailings pond detection in remote sensing images, achieving an F1-score of 0.799 and AP@.5 of 0.851.