Extracting Tailings Ponds from High Spatial Resolution Remote Sensing Images using Improved YOLOv5 and SegFormer

Abstract

Dam failures in tailings ponds pose severe threats to nearby ecosystems, residents' lives, and property. Therefore, accurately and efficiently extracting information on tailings ponds is essential. Remote sensing technology has become a crucial tool for periodic and precise detection of these ponds. However, tailings ponds vary significantly in color, scale, and shape, often blending with their surroundings, which limits the effectiveness of traditional remote sensing methods. In this paper, we propose a framework for extracting tailings ponds from high-resolution remote sensing images using an improved YOLOv5 and SegFormer. Our improved YOLOv5 incorporates the coordinate attention (CA) and Transformer attention mechanisms into the C3 module of the backbone, creating new C3CA and C3TR modules that form a hybrid attention mechanism backbone. For the neck network, we build on YOLOv6's Bi-directional Concatenation (BiC) module, replacing the nearest-neighbor interpolation with transposed convolution, and designing a new BiCT module to create the BiC Transposed Convolution Path Aggregation Network (BiCTPAN). Following detection by the improved YOLOv5, SegFormer is used to accurately delineate tailings pond boundaries. The results show that the improved YOLOv5s achieves an mAP@0.5 of 90.10%, a 4.8% increase over the original YOLOv5s, with minimal impact on parameters and Floating-Point Operations per Second (FLOPs). The SegFormer model achieves an Intersection over Union (IoU) of 87.45% and an accuracy of 94.1%, demonstrating excellent extraction performance.