Joint training of hydrologic and hydraulic models using Deep Learning and remote sensing data for the Torne River

Floods are among the most devastating natural disasters, affecting both developed and developing regions. However, developing countries often lack sufficient monitoring and early warning systems, making them more vulnerable. The ESA EO4FLOOD project aims to enhance flood forecasting by integrating satellite data with hydrologic and hydraulic models. Within this effort, we introduce a novel joint modelling framework that couples hydrologic and hydraulic models using differentiable programming.

Hydraulic and hydrologic models are constrained by data and traditionally rely on in situ measurements, which are expensive to obtain, may be access-limited, and can be dangerous to collect in remote terrain or during crises. Remotely sensed data from satellites or airborne campaigns offer a potent and low-cost alternative, with satellites providing data irrespective of national or geographic borders. Hydraulic-geometric parameters, water surface elevations (WSE), and slopes (WSS), as well as inputs to the hydrologic model, can be resolved through remote sensing.

With the launch of the Surface Water and Ocean Topography (SWOT) satellite mission, high-accuracy spatially distributed (2D) WSE and WSS observations have become available at a global scale. The primary instrument is a Ka-band radar interferometer that observes two, 50km wide swaths on each side of the ground track of the satellite, with a science requirement to detect rivers larger than 100m in width; however, even smaller rivers can be measured. The ICESat-2 satellite enables accurate global WSS and river topography observations, which can be locally substituted by national topographic LIDAR missions.

We present a differentiable hydraulic-hydrologic framework integrating large-scale Earth observation (EO) data while maintaining physical consistency. Both models are jointly trained using SWOT data, with the output of the hydrologic model serving as input to the hydraulic model. Joint training enables both models to benefit from the information contained in the SWOT data, as well as potentially satellite earth observations of additional state variables (e.g., soil moisture, evapotranspiration, terrestrial water storage). Additionally, the coupled approach allows independence from rating-curve-based discharge, marking a significant leap forward in the global applicability of hydraulic models.

We demonstrate this approach on the Torne River, located between northern Sweden and Finland. With extensive in-situ data, Torne provides an ideal case for validation. Our joint model supports accurate water level and discharge forecasting, aiding flood preparedness, informing local adaptation strategies, and enhancing climate resilience. This proof of concept highlights the method’s global potential under the EO4FLOOD initiative.