Neural Network-Based Determination of Coarse-Grained Material Size in Debris-Flow Deposits of the Southern Baikal Region

This study is devoted to automation of the assessment of the dimensional structure of coarse-grained material on orthophotomaps obtained from unmanned aerial vehicles (UAVs). Orthomosaics with a Ground Sample Distance (GSD) of ~2–4 cm were used on the material of ten debris-flow deposits of the Baikal region; an annotated dataset was prepared in SedimentAlyzer. The algorithm combines object detection by computer vision methods with subsequent classification and Feret diameter estimation using convolutional neural networks (CNN; architectures: residual convolutional network (ResNet) and densely connected convolutional network (DenseNet), transfer learning). The output is performed on overlapping cells; preliminary determinations are merged (non-maximum suppression, NMS / weighted boxes fusion, WBF), probabilities are calibrated (temperature scaling) and shape filters are applied. The implementation is integrated into SedimentAlyzer. The software is registered (state registration certificate for computer program No. 2025616929). The following values were achieved in the retained areas: F1-score = 0.84±0.03; mean absolute Feret error = 4.8 cm; RMSE=7.9 cm; correlation with field measurements R=0.89; agreement with GOST size classes averaged 82 % across basins. Compared to manual interpretation, labor time is reduced by a factor of 6–8. Typical patterns include dominance of clasts 10–100 cm (≈30–40 %) and a variable share of boulders >200 cm (1–13 % with maxima in small tributaries); in transit zones, the proportion of very large fragments is higher than in accumulation zones. The approach enables reproducible mapping of fractions and is suitable for regular monitoring of debris-flow cones and support of engineering decisions.