Abstract's details
All-in-one SWOT lake approach: Volumes, Bathymetry and Uncertainty estimation by Implicit Neural Network Representations
Event: 2025 SWOT Science Team Meeting
Session: Hydrology: SWOT Lakes, Estuaries and Wetlands (SLEW)
Presentation type: Oral
A novel approach for deriving dynamic 3D models of lakes (dry-bathymetry) and their flood basins from Surface Water Ocean Topography (SWOT) pixel cloud data is proposed. Supported by recent advances in Implicit Neural Representations (INR) and Physics-Inspired Machine Learning (PIML), we leverage physical biases regarding the geometric and temporal behavior of water and ground surfaces to learn a continuous implicit 4D model of the lake surface water and its surrounding area. Our experimental results show that despite high levels of noise in the pixel cloud - particularly in ground areas - our proposed model is able to 1. extract a static Digital Elevation Model of the dry bathymetry, 2. estimate the evolution of water height, area, volume, as well as their relations to each other, and 3. generate 2d water masks with no prior label information. We take inspiration from Bayesian Neural Networks and implement a probabilistic framework wherein the model outputs are provided with uncertainty estimations that take into account the level of noise in the SWOT pixel cloud data.
The main advantage of the method is the independency on PIXC Classification attribute. Instead, it takes advantage of 3D information, like in the case of precise dry bathymetry geolocation of PIXC on sandbanks (frequently misclassified as water). This also allows an independent estimation of water levels and areas based only on 3D information and the calculation of dry-bathymetry through the consistency of pixels’ geolocations along all available PIXC observations.
The proposed approach has been validated on Ebro reservoir, using local water manager gauge data on water resources monitoring assessment (water surface elevation and volume stock change timeseries). Furthermore, we have validated our dry bathymetry estimate with 1-meter resolution DEM from Pleiades-stereo acquisitions during a low water period.
Current research is focused on expanding the method to a larger number of reservoirs, implying a complete automated pipeline with no human intervention or calibration. We hope that the proposed approach might be reused on the future FloodPlainDEM product generation by SWOT Algorithm Definition Team.
Back to the list of abstractThe main advantage of the method is the independency on PIXC Classification attribute. Instead, it takes advantage of 3D information, like in the case of precise dry bathymetry geolocation of PIXC on sandbanks (frequently misclassified as water). This also allows an independent estimation of water levels and areas based only on 3D information and the calculation of dry-bathymetry through the consistency of pixels’ geolocations along all available PIXC observations.
The proposed approach has been validated on Ebro reservoir, using local water manager gauge data on water resources monitoring assessment (water surface elevation and volume stock change timeseries). Furthermore, we have validated our dry bathymetry estimate with 1-meter resolution DEM from Pleiades-stereo acquisitions during a low water period.
Current research is focused on expanding the method to a larger number of reservoirs, implying a complete automated pipeline with no human intervention or calibration. We hope that the proposed approach might be reused on the future FloodPlainDEM product generation by SWOT Algorithm Definition Team.