Day 4

Detailed paper information

Back to list

Paper title Estimating reservoir bathymetry from DEM with deep learning
  1. Christophe Fatras Collecte Localisation Satellites (CLS) Speaker
  2. Jérémy Augot Collecte Localisation Satellites (CLS)
  3. Lionel Zawadzki CNES
  4. Santiago Peña Luque Centre National d’Etudes Spatiales (CNES)
Form of presentation Poster
  • A7. Hydrology and Water Cycle
    • A7.01 Inland Water Storage and Runoff: Modeling, In Situ Data and Remote Sensing
Abstract text Water volumes available in natural and artificial lakes are of prime interest, either for water management purposes or water cycle understanding. However, less than 1% of global lakes are monitored. To such an end, remote sensing has been a useful tool providing continuous and global information for more than 30 years.

Combining information of water elevation with altimetry along with water surface from optical and SAR images can lead to the relative volume variation through the creation of a height-surface-volume relationship (hypsometric curve). This method is currently limited by the altimetry data coverage which is non global (less than 3% worldwide).

Even if the future wide swath altimeter, SWOT, will provide the first global survey of water bodies, the estimation of bathymetry and corresponding hypsometric curve remains a challenge to estimate water volume. Contextual approach can be considered and even trained to approximate a reservoir bathymetry from a “filled” DEM. We used such a contextual approach to develop an algorithm using deep learning to recreate the reservoir’s bathymetry.

The first step consisted in using Digital Elevation Models (DEM) cropped to the sub-basin provided by the Hydrobasin shapefile database. This led to the creation of an artificial database of DEM patches with their associated pseudo-water basins and associated 20m high reservoir. This approach was applied to relatively “dry” but mountainous or hilly countries like Chile, Turkey, Morocco, among others. A specific attention has been given to avoid planar DEM area from already existing water dams with varying water heights from 5 to 20meters. We also checked for each sub-basin that the created virtual reservoir was realistic, for instance in terms of dam length or related water surface. Thereby, we created around 9000 DEM/water surfaces patches to train a Unet deep learning algorithm. We also used data augmentation, data refining and cross-validation over the simulated reservoirs to get a realistic model. The recreated bathymetry led to an error lower than 10% on volume estimation and still improving at this time.

Further advances could be applied not only to reservoirs, but also to lakes and rivers. This would improve the global water volume estimations, but also discharges, thanks to the ever-improving DEM datasets produced in terms of precision and resolution, like with the future CO3D mission (CNES).