Assessing Flood-Influenced Sediment Dynamics Using UAV Photogrammetry and Machine Learning: Insights from River Sense, Switzerland

Abstract. Gravel-bed rivers are shaped by complex interactions between hydrological forcing, sediment sorting, and channel morphology, yet fine-scale, spatially continuous observations of these processes remain rare. We combine UAV structure-from-motion photogrammetry with machine-learning grain segmentation to quantify flood-driven sediment redistribution in a minimally disturbed gravel-bed river (Sense River, Switzerland). Two surveys of four gravel bars (2021 and 2024) mapped individual clasts in images at centimetre resolution, allowing spatial and temporal analysis of grain-size patterns. We show clear intra-bar fining from crests to tails and a reach-scale morphology control on sorting: bend-associated bars are moderately to well sorted, while straight reaches are more poorly sorted. Grain-size distributions converge to self-similar forms across all bars, with analysis of ca. 1.86 million grains providing unprecedented empirical validation of scale-invariant sorting, an improvement by orders of magnitude over conventional pebble counts. To understand the detailed hydraulic controls during the moderately large flood captured between surveys (ca. 180 m³/s; ca. 2–10 years recurrence), we performed detailed hydraulic modelling for one bar, estimating spatial fields of shear stress, Shields parameter, and stability conditions during the flood. We also differentiated the topography between the two surveys to map the relative elevation change. The crest and margin armour remained largely stable, whereas the tail part was extensively reworked. A hydraulically driven mobilisation model reproduced observed mobility with ca. 65 % overall accuracy (up to 82 % in tails) but under-predicted movement on crests. We also show that where floods were large enough to mobilise the grains, coarse patches were rapidly buried or completely replaced, demonstrating that local hydraulic geometry can override patch stability. Overall, bar adjustment was deposition-dominated for that bar, consistent with the waning stage of the flood, during which reduced shear stresses promoted net deposition. Our data indicates that flows <200 m³/s can remobilise large bar areas, and analysis of gauging data for the Sense River since 1928 shows that such events are becoming more frequent. Our results highlight the important geomorphic role of moderate to moderately large floods in such rivers and demonstrate high-resolution, hydraulically informed grain mapping as a robust tool for predicting gravel-bed river response under changing flood regimes.