EGU24-8222, updated on 08 Mar 2024
https://doi.org/10.5194/egusphere-egu24-8222
EGU General Assembly 2024
© Author(s) 2024. This work is distributed under
the Creative Commons Attribution 4.0 License.

Automated and flexible measuring of grain size and shape in images of sediment with deep learning

David Mair1, Guillaume Witz2, Ariel Henrique Do Prado1, Philippos Garefalakis1, and Fritz Schlunegger1
David Mair et al.
  • 1University of Bern, Institute of Geological Sciences, Bern, Switzerland (david.mair@geo.unibe.ch)
  • 2University of Bern, Data Science Lab, Bern, Switzerland

The size and shape of sediment particles record crucial information on erosion, transport, and deposition mechanisms during sedimentary processes. Therefore, data on grain morphometry is a critical component in understanding sediment production and transport dynamics in various environments, such as fluvial or hillslope settings. However, traditional field methods are labor-intensive, and results may suffer from a limited number of observations. At the same time, remote measurements in images or point clouds still need improvements to counter low accuracy or the need for time-consuming manual corrections (e.g., Steer et al., 2022). These persisting challenges impede the capability of routinely obtaining size and shape information.

Here, we present a new and automated approach (Mair et al., 2023) for obtaining morphometric information on coarse sediment particles from segmented images. To do so, we tap into the capability for transfer learning of deep neural networks. In particular, we use state-of-the-art deep learning, developed to find cells in biomedical images, to segment individual grains in pictures of various sediments and image types. Our method validation includes assessing segmentation performance against ground truth from annotated images and evaluating the measurement quality by comparing results to independent measurements in the field and in images. This approach facilitates precise and rapid grain segmentation and outperforms existing methods. In addition, we observe that higher segmentation quality directly leads to improved precision and accuracy for grain size and shape data. Furthermore, any model of the used architecture can easily be re-trained for new image conditions, which we successfully did for several different settings. This highlights the potential for easy adapting to different environments and scales with comparatively small datasets.

References

Mair, D., Witz, G., Do Prado, A. H., Garefalakis, P., and Schlunegger, F.: Automated detecting, segmenting and measuring of grains in images of fluvial sediments: The potential for large and precise data from specialist deep learning models and transfer learning, Earth Surf. Process. Landforms, 1–18, https://doi.org/10.1002/esp.5755, 2023.

Steer, P., Guerit, L., Lague, D., Crave, A., and Gourdon, A.: Size, shape and orientation matter: fast and semi-automatic measurement of grain geometries from 3D point clouds, Earth Surf. Dyn., 10, 1211–1232, https://doi.org/10.5194/esurf-10-1211-2022, 2022.

How to cite: Mair, D., Witz, G., Do Prado, A. H., Garefalakis, P., and Schlunegger, F.: Automated and flexible measuring of grain size and shape in images of sediment with deep learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-8222, https://doi.org/10.5194/egusphere-egu24-8222, 2024.

Comments on the supplementary material

AC: Author Comment | CC: Community Comment | Report abuse

supplementary materials version 1 – uploaded on 15 Apr 2024, no comments