Abstract DGP2026-87

Boulder movement is an important indicator of geomorphic activity and slope instability. However, boulder detection and monitoring are still challenging, especially with high-resolution remote sensing data due to complex shapes, shadows, and surface heterogeneity. Furthermore, in many cases, the large spatial extent of boulder fields with hundreds of thousands of boulders makes their investigation even more challenging. In this study, we present a segmentation-based deep learning approach for automated boulder detection and movement analysis using the segment every grain framework.

High-resolution optical imagery over a defined study area was used to identify individual boulders through an instance segmentation method. In this method, an AI model was trained to detect boulder boundaries as polygons based on luminosity, geometric, and textural features. Data augmentation strategies were applied to improve robustness to variations in illumination, scale, and background conditions. Detected boulders were tracked across multi-temporal imagery to identify spatial displacement and potential movement patterns.

The presented workflow enables precise localization and shape extraction of boulders, providing a scalable and reproducible method for boulder detection over large areas. The results demonstrate that segmentation-based approaches offer significant advantages over traditional object detection methods and hand drawing by preserving boulder geometry and enabling detailed change analysis. This methodology has potential applications in hazard assessment or landscape evolution and can be applied to Earth or other planetary bodies