Image-projected apparent particle size distribution estimation in rock avalanches using direct deep regression from UAV imagery

Due to the rapid downslope movement of fragmented rock masses, rock avalanches are among the most destructive geomorphic hazards in mountainous regions. Particle size characteristics observed at the surface provide useful information for interpreting fragmentation, transport, and depositional patterns. Recent advances in unmanned aerial vehicle (UAV)-based remote sensing and deep learning have enabled novel methodologies for PSD estimation. However, achieving rapid acquisition and reliable PSD characterization under complex field conditions remains challenging. To address these challenges, we constructed and preprocessed a dedicated UAV dataset for PSD analysis from high-resolution RGB imagery acquired with a DJI Matrice 300 RTK platform. Based on this dataset, we proposed a direct deep regression pipeline to predict characteristic image-projected apparent PSD descriptors, including D 20 , D 40 , D 60 , D 80 , and D max , directly from UAV image patches. The proposed approach was compared with two baseline methods: handcrafted feature-based regression and semantic segmentation-based contour pipelines. Experimental results showed that the direct deep regression pipeline achieved better overall PSD prediction performance than the baseline approaches, with ConvNeXt-Tiny obtaining the lowest mean absolute error (MAE) of 2.23 cm and resolution-normalized MAE of 16.77 ground sampling distance (GSD) units. The model exhibited more stable predictive performance for the intermediate percentiles (D 40 –D 80 ), whereas relatively higher errors were observed for D 20 and D max . The trained model was applied to representative areas within the study deposit to generate spatial maps of surface-visible apparent PSD, demonstrating its potential for UAV-based characterization of surface particle size patterns in fragmented rock deposits. Additional transfer validation on tunnel boring machine (TBM)-excavated rock images provides preliminary evidence of cross-scenario transferability across different material and imaging conditions. True cross-event validation on independent rock avalanche deposits remains to be conducted.