Introducing PebbleCounts: A grain-sizing tool for photo surveys of dynamic gravel-bed rivers

Abstract

Abstract. Grain-size distributions are a key geomorphic metric of gravel-bed rivers. Traditional measurement methods include manual counting or photo sieving, but these are achievable only at the 1–10 m2 scale. With the advent of unmanned aerial vehicles and increasingly high-resolution cameras, we can now generate orthoimagery over hectares at sub-cm resolution. These scales, along with the complexity of high-mountain rivers, necessitate different approaches for photo sieving. As opposed to other image segmentation methods that use a watershed approach to automatically segment entire images, our open-source algorithm, PebbleCounts, relies on k-means clustering in the spatial and spectral domain and rapid manual selection of well-delineated grains. The result is improved grain-size estimates for complex river-bed imagery, without any post processing. In a second step, we develop a fully automated method, PebbleCountsAuto, that relies on edge detection and filtering suspect grains, without the k-means clustering or manual selection steps. The algorithms are tested in controlled indoor conditions on three arrays of pebbles and then applied to 12 × 1 m2 orthomosaic clips of high-energy mountain rivers collected with a camera-on-mast setup (akin to a low-flying drone). A 20-pixel b-axis length lower truncation is necessary for attaining accurate grain-size distributions. For the k-means PebbleCounts approach, average percentile bias and precision are 0.03 and 0.09 ψ, respectively, for ~ 1.16 mm/pixel images, and 0.07 and 0.05 ψ for one 0.32 mm/pixel image. The automatic approach has higher bias and precision of 0.13 and 0.15 ψ, respectively, for ~ 1.16 mm/pixel images, but similar values of −0.06 and 0.05 ψ for one 0.32 mm/pixel image. For the automatic approach, only at best 70 % of the grains are correct identifications, and typically around 50 %. PebbleCounts operates most effectively at the 1 m2 scale, where the algorithm can be rapidly applied in ~ 5 minutes in many small areas to acquire accurate grain-size data over 10–100 m2 areas. These data can be used to validate PebbleCountsAuto applied at the scale of entire survey sites (102–104 m2). We synthesize results and recommend best practices for image collection, orthomosaic generation, and grain-size measurement using both algorithms.

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

“…Traditional image processing, also referred to as objectbased approaches (e.g., Carbonneau et al, 2018), has been applied to segment individual grains and measure their sizes, by fitting an ellipse and reporting the length of its minor axis as the grain size (Butler et al, 2001;Sime and Ferguson, 2003;Graham et al, 2005Graham et al, , 2010Detert and Weitbrecht, 2012;Purinton and Bookhagen, 2019). Detert and Weitbrecht (2012) presented BASEGRAIN, a MATLAB-based object detection software tool for granulometric analysis of ground-level top-view images of fluvial, noncohesive gravel beds.…”

“…Despite this automated image analysis, extensive manual parameter tuning is often necessary, which hinders the automatic application to large and diverse sets of images. Recently Purinton and Bookhagen (2019) introduced a python tool called PebbleCounts as a successor of BASEGRAIN, replacing the watershed approach with k-means clustering.…”

“…However, at test time the proposed approach requires no parameter tuning by the user, which is a considerable advantage for large-scale applications, where traditional image processing pipelines struggle, since they are fairly sensitive to varying imaging conditions. Semiautomatic image labeling with the support of traditional image processing tools (Detert and Weitbrecht, 2012;Purinton and Bookhagen, 2019) might be an alternative way to speed up this annotation process. However, one would have to carefully avoid systematic algorithmic biases in the semiautomatic procedure, otherwise the CNN will almost certainly learn to faithfully reproduce those biases.…”

“…Since it is nondestructive, multiple operators can label the exact same location. Much research tried to automatically estimate grain size distributions from ground-level images (Butler et al, 2001;Rubin, 2004;Graham et al, 2005;Verdú et al, 2005;Detert and Weitbrecht, 2012;Buscombe, 2013;Spada et al, 2018;Buscombe, 2019;Purinton and Bookhagen, 2019). On the contrary, relatively little research has addressed the automatic mapping of grain sizes from images at larger scale (Carbonneau et al, 2004(Carbonneau et al, , 2005Black et al, 2014;de Haas et al, 2014;Carbonneau et al, 2018;Woodget et al, 2018;Zettler-Mann and Fonstad, 2020), which is needed for prac-tical impact.…”

Answer to RC2

“…Due to the end-to-end learning, the proposed CNN is able to extract global features that are informative about grain size beyond the sensitivity of human photo-interpretation as well as traditional photosieving that relies on local image gradients to delineate individual grains. Even the latest work of Purinton and Bookhagen (2019) can only detect individual grains that have a b-axis 20x the ground sampling distance. Also previous statistical approaches (Carbonneau 2004) are limited by the input resolution and can only predict D50 down to 3 cm, so we do believe that our approach overcomes some of the limitations of prior art.…”