Gf-former: an accurate UAV-based remote sensing image network for high-precision automatic segmentation of ground fissures in mining regions.

Meng, Jiangtao; Xu, Xinying; Li, Pengyue; Zhang, Zhe; Zhao, Wenjing; Ren, Jinchang; Li, Yuchen

Abstract

Ground fissure information is critical for ensuring the safety of mining operations and preventing geological disasters. Challenges include obscuration by vegetation or shadows and varying fissure sizes. To address these, we introduce GF-Former, a specialized deep learning network for precise segmentation of ground fissures in remote sensing images. GF-Former utilizes a Mix Transformer encoder (Mit) to capture long-range dependencies, enhancing global perception. An Adaptive All Feature Fusion (AAFF) module dynamically adjusts feature weights according to interference conditions, effectively combining semantic information with edge details. A Dense Spatial Pyramid Pooling (DSPP) module extracts and aggregates multi-scale spatial information, improving detection of fissures of various sizes. A Focal Dice Loss is designed to enhance recognition capabilities in challenging conditions. To advance deep learning in ground fissure extraction, we created a dataset (GFD) including data from 27 mining faces in Liliu. Experiments on GFD demonstrate that GF-Former achieves an mIoU of 75.02%, mDice of 83.46%, and mPA of 86.15%, outperforming other models. Testing on public datasets DeepCrack and Crack500 confirms the adaptability and reliability of GF-Former in fissure detection. GF-Former provides a reliable solution for the extraction of ground fissures in mine remote sensing imagery.

Citation

MENG, J., XU, X., LI, P., ZHANG, Z., ZHAO, W., REN, J. and LI, Y. 2025. Gf-former: an accurate UAV-based remote sensing image network for high-precision automatic segmentation of ground fissures in mining regions. International journal of machine learning and cybernetics [online], Latest Articles. Available from: https://doi.org/10.1007/s13042-025-02555-7

• Journal Article Type: Article
• Acceptance Date: Jan 16, 2025
• Online Publication Date: Feb 4, 2025
• Deposit Date: Feb 14, 2025
• Publicly Available Date: Feb 5, 2026
• Journal: International journal of machine learning and cybernetics
• Print ISSN: 1868-8071
• Electronic ISSN: 1868-808X
• Publisher: Springer
• Peer Reviewed: Peer Reviewed
• DOI: https://doi.org/10.1007/s13042-025-02555-7
• Keywords: Ground fissure; Semantic segmentation; Deep learning; Remote sensing; Transformer; Mining regions
• Public URL: https://rgu-repository.worktribe.com/output/2702574