A Positional Knowledge-Guided Multiscale Gaussian Detail Enhancement Deep Learning Network for Ground Fissure Extraction

Abstract

The ground fissures caused by coal mining may pose a potential threat to mine safety. When applying deep learning to the task of extracting ground fissures in mining areas, challenges often arise due to limited sample data, poor performance in transfer learning, and the complex surface conditions that hinder surface monitoring in mining regions. This article proposes a multiscale Gaussian detail enhancement deep learning network, PF-Unet3+, guided by prior positional knowledge. This network improves upon the encoding section of Unet3+ by introducing dual-branch down-sampling, ${{X}_{En}}$ and ${{Y}_{En}}$. At the input side, a prior knowledge generator is designed to obtain the prior positional information ${{I}_p}$ of ground fissures and the prior map of ground fissures ${{I}_{pc}}$. ${{I}_p}$ is applied as a constraint to the ${{X}_{En}}$branch to improve the transfer learning effect; Input ${{I}_{pc}}$ into the ${{Y}_{En}}$branch and incorporate a Gaussian detail enhancement module to enhance the network's focus on ground fissures and its ability to perceive details. This article constructs a target domain dataset using drone images from the Lingquan mining area in Inner Mongolia, China. Experiments on the dataset show that the PF-Unet3+ network achieves a precision of 94.05%, recall of 93.85%, F1-Score of 93.94%, and mean Intersection over Union (mIoU) of 94.01% in ground fissure extraction, significantly outperforming deep learning networks such as Deeplabv3+, U-Net, Unet3+, SegNext, and SeaFormer. The deep learning network proposed in this study provides a reliable method for monitoring ground fissures in mining areas, serving the safety and ecological restoration of mining areas.

Keywords

• Deep learning
• Gaussian detail enhancement
• ground fissure
• prior knowledge