Automated high-resolution 3D crevasse extraction and dynamic linkages: an integrated UAV-LiDAR, photogrammetry, and C-TransUNet framework

Glacier crevasses are critical indicators of ice dynamics and stability, yet their detailed monitoring is hindered by the limitations of traditional remote sensing. This study presents an innovative, integrated framework combining Unmanned Aerial Vehicle (UAV)-based LiDAR Scanning (UAV-LS), photogrammetry, and an optimized deep learning model, C-TransUNet, for automated, high-resolution, three-dimensional (3D) crevasse characterization. We conducted surveys at the terminus of the Yanong Glacier (YNG), Tibetan Plateau, acquiring centimeter-resolution orthophotos and LiDAR point clouds. The enhanced C-TransUNet model, featuring a local–global collaborative encoder and adaptive multi-scale feature fusion, significantly outperformed a suite of well-established and representative methods in crevasse extraction (mIOU = 88.04 %, F1-Score = 87.06 %) and demonstrated promising spatial transferability. A novel workflow integrating the deep learning results with UAV-LS point clouds enabled the systematic extraction of 3D crevasse geometry, including length, width, orientation, and unprecedented detail in depth (average 3.06 ± 3.91 m, max 26.69 m). Five distinct crevasse types were identified and meticulously mapped, revealing significant variations across different altitudinal zones. Furthermore, surface strain rates calculated from UAV-derived velocity data revealed preliminary quantitative links between crevasse patterns and underlying glacier dynamics. Our initial findings suggest that transitions between crevasse types correspond to changes in the local strain regime. This study establishes a powerful, automated framework for fine-scale, multi-dimensional crevasse analysis, offering a robust foundation for gaining crucial insights into glacier mechanics and stability in response to climate change.