Super-Resolution for Enhanced Fluvial Sediment Measurement in UAV Images

Drone imagery can efficiently perform large-scale riverbed grain size measurements. However, its applicability is significantly constrained by image resolution limitations. This issue is especially critical in mountainous areas, where sediments exhibit a wide range of grain sizes and spatial heterogeneity. To address this issue, this paper develops a new fluvial sediment measurement technique for UAV images using a deep learning technique super-resolution (SR). We first used RTK-based UAV technology to collect high-resolution riverbed grain orthophotos of different types of mountain rivers, with the collected UAV images having a resolution between 3~5 mm/pixel. Four types of super-resolution models Nearest Neighbor, Lanczos filter, SRCNN and SRGAN were trained to restore the high-resolution images from low-resolution riverbed images. Three automated grain sizing methods BASEGRAIN, GrainID and ImageGrains were applied to the images restored by SR models, and 113,456 manual grain labels are created as grain size baseline for model evaluation. The efficacy of all three models diminishes with decreasing resolution, with BASEGRAIN being the most robust and GrainID the most sensitive. Application of all four SR models model significantly increase the efficacy of grain size measurement, and SRGAN models with upscaling factor of 4 (SRGAN×4) outperform other models. Further analysis shows the minimum detectable sediment particle size of SRGAN×4 is 1 pixel, which exceed the minimal human vision limitation for detecting grain size. The SR technology proposed in this paper makes it more feasible to rapidly obtain the riverbed grain size over a wide range in mountainous rivers.