Bed material facies mapping at braided river scale and evidence for trends in fine sediment

Characterizing the spatial distribution and dynamic nature of bed facies in gravel-bed braided rivers is challenging but necessary to understand fluvial and ecological processes. Topographic point cloud and image datasets are increasingly used in fluvial geomorphology to compare riverscapes over time and classify substrate facies. However, repeatable and efficient methods that operate at large spatial scales and also resolve bimodal or fine sediments remain underdeveloped. This study collected high-resolution lidar and optical imagery over a 56-km reach of the Rangitata River, New Zealand, generating a variety of multiscale lidar-derived, optical and local morphological predictors. Ensemble machine learning methods were used to classify facies at a 1 m resolution, and a sensitivity analysis incorporating downscaled and reduced-fidelity datasets was conducted to understand the importance of data acquisition strategies. We report the predictor importance by class, finding that the key predictors for fine sediment were colour and colour complexity, while lidar predictors including reflectance were key in differentiating shallow water. The classification method was found to be robust with decreasing lidar point density but the performance was degraded if either RGB or lidar datasets were removed entirely.

The sole use of spatially local predictors allowed an analysis of trends in fine sediment facies in a large braided river subject to hydrologic pressures. The quantity and proportion of exposed fine sediment increased downstream, indicating a decrease in transport capacity associated with river widening. This new lidar – machine learning – substrate processing pathway offers a synoptic view of river form and composition that can be used to parameterize numerical models and provide distributed insights into sediment transport and sorting processes. The approach is easy to customize and can readily adapted to predict different surface classes, providing a robust basis for change detection in natural scenes.