Automated mapping of grain size distributions from UAV imagery using the CNN-based GRAInet model

Abstract

The grain size distribution of gravel riverbed material is an essential parameter to estimate the sediment transportation, groundwater-river flow interaction, river ecosystem, and fluvial geomorphology. Conventional and present methods of obtaining grain size distribution analysis of more extensive areas are time-consuming and remain challenging in effectively modeling sediment load. On this account, this paper appraised the role of employing the end-to-end data-driven GRAINet approach, a convolutional neural network (CNN) application, to predict and map the grain size distribution at particular locations over an entire gravel bar based on georeferenced drone-based orthoimagery. We conducted multiple drone surveys after post-flood events in the Javoří Brook Šumava National Park (Šumava NP) in Czechia over a small unregulated montane stream with an exposed gravel bar and frequently changed fluvial dynamics. The GRAINet model performances between the predicted mean diameter (dm) and ground truth diameter dm (human performance) produce the result of different loss functions, i.e., the mean absolute errors (MAEs), the mean squared errors (MSEs), and the root-mean-square errors (RMSEs). Corresponding averages of MAEs varied between 3 cm to  4.8 cm with standard deviations (STDs) of 1.7 cm and 3.8 cm, respectively. The averages of MSE ranged between 13 cm to 14.5 cm with  STDs of 12.7 cm and 12.8 cm, and RMSE of 3.2 cm to  5.6 cm with STDs of 1.6 cm and 4.6 cm, respectively. With high to moderate accuracies and lower computational costs than other deep learning approaches, the tested ensemble model shows that the integration of UAV remote sensing and machine learning (ML) provides a promising tool to help make decisions using timely mapped high-resolution grain size maps without access to direct object counts or locations.