Preprints
https://doi.org/10.5194/esurf-2021-67
https://doi.org/10.5194/esurf-2021-67

  15 Sep 2021

15 Sep 2021

Review status: this preprint is currently under review for the journal ESurf.

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

Xingyu Chen1,2, Marwan A. Hassan2, and Xudong Fu1 Xingyu Chen et al.
  • 1Department of Hydraulic Engineering, State Key Laboratory of Hydroscience and Engineering, Tsinghua University, Beijing, China
  • 2Department of Geography, The University of British Columbia, Vancouver, BC, Canada

Abstract. Image-based grain sizing has been used to measure grain size more efficiently compared to traditional methods (e.g. sieving and Wolman pebble count). However, current methods (e.g. BASEGRAIN) are largely based on detecting grain interstices from image intensity which not only require a significant level of expertise for parameter tuning but also underperform when they are applied to sub-optimal environments (e.g. dense organic debris, various sediment lithology). We proposed a model (GrainID) based on convolutional neural networks to measure grain size in a diverse range of fluvial environments. A data set of more than 125,000 grains from flume and field measurements were compiled to develop GrainID. Tests were performed to compare the predictive ability of GrainID with sieving, manual labeling, Wolman pebble counts and BASEGRAIN. When compared with the sieving results for a sandy-gravel bed, GrainID yielded high predictive accuracy (comparable to the performance of manual labeling) and outperformed BASEGRAIN and Wolman Pebble counts (especially for small grains). For the entire evaluation dataset, GrainID once again showed fewer predictive errors and significantly lower variation in results in comparison to BASEGRAIN and Wolman pebble counts and maintained this advantage even in uncalibrated rivers with drone images. Moreover, the existence of vegetation and noise have little influence on the performance of GrainID. Analysis indicated that GrainID performed optimally when the image resolution is higher than 1.8 mm/pixel, the image tile size is 512*512 pixel*pixel and the grain area truncation values (the area of smallest detectable grains) were equal to 18–25 pixels.

Xingyu Chen et al.

Status: open (until 27 Oct 2021)

Comment types: AC – author | RC – referee | CC – community | EC – editor | CEC – chief editor | : Report abuse

Xingyu Chen et al.

Data sets

Data sets for "CNN for image-based sediment detection applied to a large terrestrial and airborne dataset" Xingyu Chen, Marwan Hassan, Xudong Fu https://zenodo.org/record/5240906

Model code and software

GrainID model code for "CNN for image-based sediment detection applied to a large terrestrial and airborne dataset" Xingyu Chen, Marwan Hassan, Xudong Fu https://zenodo.org/record/5240906

Xingyu Chen et al.

Viewed

Total article views: 221 (including HTML, PDF, and XML)
HTML PDF XML Total BibTeX EndNote
175 42 4 221 0 0
  • HTML: 175
  • PDF: 42
  • XML: 4
  • Total: 221
  • BibTeX: 0
  • EndNote: 0
Views and downloads (calculated since 15 Sep 2021)
Cumulative views and downloads (calculated since 15 Sep 2021)

Viewed (geographical distribution)

Total article views: 169 (including HTML, PDF, and XML) Thereof 169 with geography defined and 0 with unknown origin.
Country # Views %
  • 1
1
 
 
 
 
Latest update: 27 Sep 2021
Download
Short summary
We compiled a large image dataset containing more than 125,000 sediments and developed a model (GrainID) based on convolutional neural networks to measure individual grain size from images. The model was calibrated on flume and natural streams images covering wide range of fluvial environments. The model showed high performance in comparison to other methods. Our model showed great potential for grain size measurements from small patch of sediment in a flume to watershed-scale drone survey.