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

Abstract

The quantification of the bed grain size distribution (GSD) of river surfaces is primarily conducted through manual approaches in the field. These methods are time consuming and not able to accurately represent the spatial diversity of the grain size distribution of rivers. Recently, several software programs and procedures have been developed using semi-automatic and automatic methods to estimate bed GSD from digital imagery. The purpose of this study is to compare softwares accuracy between reference GSDs and estimated GSDs using geometric approaches (Basegrain software and a procedure developed on ImageJ), statistical approaches (digital grain size[DGS] and PebbleCounts softwares), and a machine learning framework (SediNet).This study evaluates ten digital images recorded along the Rhine River downstream of the city of Basel. The results showed that all software programs considerably underestimated the manually measured GSDs. Nevertheless, it is possible to significantly improve the estimation of bed GSD by applying calibration laws. Both DGS and Basegrain softwares are reliable to estimate the GSD, while the three others softwares are accurate for percentiles equal and higher than the D 50 . After linear regression correction, the mean normalized root mean square error of percentile errors did not exceed 13% for DGS and Basegrain software, while the others did not exceed 22% for percentiles coarser than the D 50 .

Comparison of three grain size measuring methods applied to coarse-grained gravel deposits

“…In the past years, applications of semi-automatic grain size measuring methods, where algorithms model ellipsoids around single grains, have gained an increasing popularity (Detert and Weitbrecht, 2012;Purinton and Bookhagen, 2019). Despite improvements in such techniques, measuring sizes of fluvial gravels in an accurate and reproducible way still bears challenges (e.g., Chardon et al, 2021;Purinton and Bookhagen, 2021;Mair et al, 2022).…”

Grain size of fluvial gravel bars from close-range UAV imagery – uncertainty in segmentation-based data

“…We base this inference on the results of other analyses, which were accomplished with the same segmentation software and which documented a systematic underestimation of related percentile values, thus hinting at an effect related to over-segmentation (Chardon et al, 2022). This issue might be addressed if (i) images are segmented semi-automatically where manual measurements are accomplished occasionally to set a benchmark (Purinton and Bookhagen, 2021), (ii) reference measurements are conducted for calibration purposes (Chardon et al, 2022), or if (iii) the automated segmentation is improved. However, more research is needed to improve our understanding of systematic traits of segmentation-based grain sizes and the related dependency on survey-specific characteristics.…”

“…Second, other segmentation-based approaches, e.g., Basegrain (Detert and Weitbrecht, 2012) or manual segmentation (Sulaiman et al, 2014), require manual processing of each image and are therefore not suitable for the large number of processed images as is the case in this study. We acknowledge that there are known shortcomings of Pebble-Counts, and we refer to Chardon et al (2022) for a comparison with other software results and to Purinton and Bookhagen (2021) for mitigation strategies of some shortcomings.…”

“…However, current segmentation techniques are prone to biases that result from underor over-segmentation and 2D projection effects of 3D structures. Therefore, in such cases, reduction in inaccuracies might be achieved through manual filtering of grains during segmentation (e.g., Purinton and Bookhagen, 2019;Detert and Weitbrecht, 2012) and/or through a calibration of the measurements with a reference dataset (e.g., Chardon et al, 2022), where data were collected in the field, as exemplified in this work. Such a strategy is likely to improve the accuracy of grain size data and yields in an estimate of the related uncertainty.…”

Cobble edge detection algorithm based on digital image processing

“…Furthermore, semiautomated or fully automated particle size calculation software based on digital images has been developed. 22 Schmitt et al 23 compared various grain size estimation software and found that all of them underestimated the actual grain size. In addition, the cobbles often cover each other on the bed surface, resulting in incomplete edges after image segmentation.…”