Deep learning-based object detection on LiDAR-derived hillshade images: insights into grain size distribution and longitudinal sorting of debris flows

Abstract. Debris flows are hazardous natural phenomena characterized by rapid movements of sediment-water mixtures in steep channels, posing significant risks to life and infrastructure. Better understanding and managing these hazards requires new methods to collect and process high-resolution data. This study introduces a novel method that leverages hillshade images derived from a high temporal resolution LiDAR scanner and deep learning-based object detection models to analyze debris-flow dynamics. By transforming 3D point clouds into hillshade projections, the method enables efficient detection and tracking of key flow features, including boulders, rolling boulders, surge waves, and woody debris, independent of ambient light conditions. Outputs include object velocities, sizes, and tracks, offering high-resolution insights into debris-flow phenomena such as longitudinal sorting. Six state-of-the-art object detection models were evaluated, with YOLOv11 achieving the best balance of precision, recall, and processing speed. We used the framework to calculate dynamic grain size distributions and found that the median grain size decreased continuously throughout the event. The proposed framework is scalable, significantly reduces processing time compared to manual analysis, and sets the foundation for real-time monitoring and early-warning of debris flows across diverse locations and conditions.