Repository logo

Infoscience

  • English
  • French
Log In
Logo EPFL, École polytechnique fédérale de Lausanne

Infoscience

  • English
  • French
Log In
  1. Home
  2. Academic and Research Output
  3. Conferences, Workshops, Symposiums, and Seminars
  4. GALET: A deep learning image segmentation model for drone-based grain size analysis of gravel bars
 
conference paper

GALET: A deep learning image segmentation model for drone-based grain size analysis of gravel bars

Mörtl, Christian  
•
Baratier, Alexandre
•
Berthet, Johan
Show more
Ortega-Sánchez, Miguel
2022
Proceedings of the 39th IAHR World Congress
39th IAHR World Congress

In this study, we present the deep learning image segmentation model for drone-based grain size analysis of gravel bars called GALET. The objectives are to quantify the performance of the code and to test its applicability in river research and management. GALET is built with the Mask R CNN convolution neural network code. It performs instance segmentation based on a combination of the two pre-existing models Faster R CNN, which is used for object detection and Feature Pyramid Network (FCN), which classify images at pixel level. GALET is based on the open source Matterport implementation (standard for 3D space capture) and the pre-trained convolutional neural network ResNet101 and is implemented in the QGIS environment. To be effective, deep learning models need a large amount of training data. In the case of segmentation models, sample images must be annotated individually, which is particularly difficult for individual pebble detection. Our approach is based on the automation of image creation. River bars from three different gravel rivers in the French and Swiss Alpine region, with different morphological and flow regime characteristics, were sampled by line count and individual grain size measurements. At the Sarine residual flow reach, a subsurface sample from a natural gravel bar was also analysed in the laboratory by sieve analysis. Orthographic photos at all sites were taken by drone under varying environmental conditions and at different flight heights. In individual pebble detection, the model GALET detects the B-axis length with a relative absolute error of 0.29. In the comparison of grain size distributions between GALET and analog measurements, the median difference of relative occurrences ranges from -13% to 9%, depending on the size class and resolution of the ortho image. In application, GALET provided important insights in the investigations of a sediment augmentation measure and the spatial distribution of the armor layer on a natural gravel bar in a residual flow reach. The lack of technology and skill for high resolution mapping of spatial and temporal granulometric distributions of gravel bars has left a missing link for better understanding of sediment transport processes. GALET presents a promising tool to fill this gap and contribute to a better understanding of morphological and water processes from snow to the sea.

  • Files
  • Details
  • Metrics
Type
conference paper
DOI
10.3850/IAHR-39WC252171192022895
Web of Science ID

WOS:001070410605092

Author(s)
Mörtl, Christian  
Baratier, Alexandre
Berthet, Johan
Duvillard, Pierre-Allain
De Cesare, Giovanni  
Editors
Ortega-Sánchez, Miguel
Date Issued

2022

Publisher

International Association for Hydro-Environment Engineering and Research (IAHR)

Publisher place

Granada, Spain

Published in
Proceedings of the 39th IAHR World Congress
ISBN of the book

978-90-832612-1-8

Total of pages

10

Start page

5326

End page

5335

Subjects

Technology

•

Physical Sciences

•

GALET

•

Drone

•

Grain size analysis

•

Deep learning

•

Image segmentation

Note

[1398-3]

Editorial or Peer reviewed

REVIEWED

Written at

EPFL

EPFL units
PL-LCH  
Event nameEvent placeEvent date
39th IAHR World Congress

Granada, Spain

June 19-22, 2022

FunderGrant Number

Swiss Federal Office of the Environment (FOEN)

Haut-Jura National Park

Use this identifier to reference this record
https://infoscience.epfl.ch/handle/20.500.14299/202705
Logo EPFL, École polytechnique fédérale de Lausanne
  • Contact
  • infoscience@epfl.ch

  • Follow us on Facebook
  • Follow us on Instagram
  • Follow us on LinkedIn
  • Follow us on X
  • Follow us on Youtube
AccessibilityLegal noticePrivacy policyCookie settingsEnd User AgreementGet helpFeedback

Infoscience is a service managed and provided by the Library and IT Services of EPFL. © EPFL, tous droits réservés