Research on surface crack segmentation and quantitative degradation assessment in alpine meadows based on UAV imagery and the YOLO-AMSC model

Cracks in the mattic epipedon of alpine meadows are a key morphological indicator of ecosystem degradation. However, the accurate extraction of these minute and irregular crack features remains a significant challenge in complex natural environments characterized by vegetation cover, drastic fluctuations in light intensity, and soil texture interference. To address these challenges, a novel deep learning network, YOLO-AMSC, is proposed for the high-precision segmentation and morphological quantification of cracks in the mattic epipedon of alpine meadows under complex field environments. A Spatial to Depth Attention Fusion (SDAF-Block) module is proposed to preserve fine-grained spatial details of minute cracks without information loss and to reconstruct their broken topological continuity. Simultaneously, a Mixed Local Channel Attention (MLCA) mechanism is introduced to adaptively enhance crack textures while suppressing background noise. Furthermore, a hierarchical focused loss function, termed HFP-IoU (Hierarchical Focal-Penalty IoU), is designed to impose strict geometric constraints on the elongated crack boundaries via a non-monotonic focusing mechanism. The experimental results demonstrate that YOLO-AMSC achieves superior segmentation performance, improving mAP50 by 4.27%, 2.96%, 2.43%, 4.58%, 8.86%, 12.51%, 22.03%, 55.77%, and 52.19% compared with YOLOv5n-seg, YOLOv10n-seg, YOLO-hyper-seg, YOLOv12n-seg, SOLOv2, SparseInst, DeepCrack, CrackFormer-II, and CrackSegDiff, respectively. Using the high-fidelity segmentation masks from this model and a skeleton-based geometric constraint method for crack width estimation, the relative error relative to field measurements is less than 10%. Employing a system dynamics framework, this study quantitatively determines the topological width thresholds of 0.6 cm and 3.1 cm, which signify a sudden increase in connectivity and indicate an accelerated degradation process of alpine meadows. This method directly links pixel-level image analysis with regional early warning systems, delivering a cost-effective, non-destructive digital framework for dynamic ecological monitoring.