GRAINet: Mapping grain size distributions in river beds from UAV images with convolutional neural networks

Abstract. Grain size analysis is the key to understand the sediment dynamics of river systems. We propose GRAINet, a data-driven approach to analyze grain size distributions of entire gravel bars based on georeferenced UAV images. A convolutional neural network is trained to regress grain size distributions as well as the characteristic mean diameter from raw images. GRAINet allows the holistic analysis of entire gravel bars, resulting in (i) high-resolution maps of the spatial grain size distribution at large scale, and (ii) robust grading curves for entire gravel bars. To collect a large training dataset of 1,491 samples, we introduce digital line sampling as a new annotation strategy, following the widely applied line sampling analysis of Fehr (1987). Our evaluation on 25 gravel bars along 6 different rivers in Switzerland yields high accuracy: The resulting maps of mean diameters have a mean absolute error of 1.1 cm, with no bias. Robust grading curves for entire gravel bars can be extracted if representative training data is available. At the gravel bar level the MAE error of the predicted mean diameter is even reduced to 0.3 cm. Extensive experiments were carried out to study the quality of the digital line samples, the generalization capability of GRAINet to new locations, the model performance w.r.t. human labeling noise, the limitations of the current model, and the potential of GRAINet to analyze images with low resolutions.