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

Abstract

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 for the holistic analysis of entire gravel bars, resulting in (i) high-resolution estimates and maps of the spatial grain size distribution at large scale and (ii) robust grading curves for entire gravel bars. To collect an extensive training dataset of 1491 samples, we introduce digital line sampling as a new annotation strategy. Our evaluation on 25 gravel bars along six different rivers in Switzerland yields high accuracy: the resulting maps of mean diameters have a mean absolute error (MAE) of 1.1 cm, with no bias. Robust grading curves for entire gravel bars can be extracted if representative training data are available. At the gravel bar level the MAE of the predicted mean diameter is even reduced to 0.3 cm, for bars with mean diameters ranging from 1.3 to 29.3 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 with respect to human labeling noise, the limitations of the current model, and the potential of GRAINet to analyze images with low resolutions.

Measuring the grain‐size distributions of mass movement deposits

“…In this study, photos were taken with a 50 cm  50 cm frame for reference with resolution >0.12 mm pi À1 . et al, 2018) segmentation techniques isolate and measure the visible axes of individual grains (e.g Graham et al, 2005;Purinton & Bookhagen, 2019), whereas texture-based techniques are statistical approaches which produce GSDs using information about how intensity and colour vary within 2D and 3D images (Buscombe, 2013;Lang et al, 2021), for example a high-resolution digital elevation model (DEM).…”

Comparison of software accuracy to estimate the bed grain size distribution from digital images: A test performed along the Rhine River

“…Morphological approaches use thresholding and segmentation processing to define the outline of each visible particle, while statistical approaches tend to estimate the grain size through image texture analysis from the semi-variance approach (Carbonneau, Lane, & Bergeron, 2004), autocorrelation approach (Warrick et al, 2009), the wavelength approach (Buscombe, 2013), and recently, following the k-means approach (Purinton & Bookhagen, 2019). Deep learning methods were also developed to automatically estimate the GSD using digital images such as SediNet (Buscombe, 2019) and GRAINet (Lang et al, 2021). Although Basegrain (morphological approach) and DGS (statistical approach) programs are frequently used to estimate the GSD from digital images, a question still remains:…”

CNN for image-based sediment detection applied to a large terrestrial and airborne dataset

“…Generally, image-based automated grain sizing methods can be classified from percentile-based to object-based methods (Buscombe, 2020). Percentile-based methods (Carbonneau et al, 2004;Rubin, 2004;Buscombe, 2020;Buscombe et al, 2010) estimate grain size distribution based on statistical analysis of image intensity and texture through pixel-wise simple autocorrelation algorithms (Rubin, 2004), grain size prediction as a function of both local image texture and semi variance (Carbonneau et al, 2004), spectral decomposition of an image (Buscombe et al, 2010) and convolutional neural networks (CNN) (Buscombe, 2020;Mueller, 2019;Lang et al, 2021).…”