YOLOv5-RF: a deep learning method for tailings pond identification in high-resolution remote sensing images based on improved loss function

Abstract

Tailings ponds are critical facilities in the mining industry, and accurate monitoring and management of these ponds are of paramount importance. However, conventional object detection methodologies, including recent advancements, often face significant challenges in addressing the complexities inherent to tailings pond environments. This is particularly due to deficiencies in their loss function design, which can result in protracted convergence times and suboptimal performance when detecting smaller targets. In this study, we introduce an innovative loss function termed the Rapid Intersection over Union (RIoU) loss function, which incorporates a focal weight and is integrated into the YOLOv5 object detection framework to develop the YOLOv5-RF model. This approach aims to enhance both convergence speed and improve convergence accuracy in the tailings pond identification process by comprehensively addressing the specific challenges posed by complex environmental conditions, thereby enhancing the precision and robustness of tailings pond target detection. It integrates the concepts of the central triangle and the aspect ratio of the circumscribed rectangle, assigning specific weights and penalty terms to optimize the model’s performance in object detection tasks. We validated the efficacy of YOLOv5-RF through simulation experiments and high-resolution remote sensing images of tailings ponds. The experimental results indicate that RIoU facilitates faster convergence rates. Specifically, YOLOv5-RF achieves accuracy and recall rates that are 2% and 2.1% higher than those of YOLOv5, respectively. Furthermore, it completes 120 iterations in 1.08 hours less time compared to its predecessor model while exhibiting an inference time that is 11.7 ms shorter than that for YOLOv5. These findings suggest that our model significantly enhances processing speed without compromising accuracy levels. This research offers novel technical approaches as well as theoretical support for monitoring tailings ponds using computer vision and remote sensing technologies.

Keywords

• Object detection
• YOLOv5
• loss function
• tailings ponds
• remote sensing identification