Hydraulic models are an essential tool for water management and associated risks. These dynamic models are used, for example, to represent flood propagation across river networks, flood management (forecasting, prevention, protection), the reconstruction of flood discharge from water levels or flooded areas, sediment transport modeling, etc. In particular, 1-D (one-dimensional) hydraulic models are widely used in practice because they offer many advantages without compromising the physical foundation of the underlying equations: they are suited to numerous operational objectives, directly linked to field data, and are computationally efficient. These models will be the main focus of my thesis.
The implementation of a hydraulic model requires information on the topography of the study area, particularly the bathymetry of the river channel, the floodplain, and hydraulic structures, as well as the basal friction, upstream and lateral boundary conditions (typically hydrographs), and downstream conditions (such as limnigrams or tide gauges). Not all of this information is systematically measured, and some data are not even measurable (e.g., friction coefficients, which, in practice, compensate for other modeling errors), so model calibration is generally performed.
This calibration is often done manually in operational practice, making it difficult to reproduce. Additionally, the expertise involved in manual calibration is not explicitly formalized, which hinders its transmission and critical review.

While there are general methodological frameworks for model calibration (such as data assimilation, Bayesian estimation, and others), adapting them to the complex situations encountered in practice requires non-trivial specifications, such as the characterization of uncertainties in data and the model, and the spatialization of parameters.
The availability of data such as altimetric measurements from the SWOT satellite or data collected during floods (e.g., drone overflights or flood marks) presents a significant opportunity to improve the estimation of hydraulic model parameters and enable their deployment "everywhere." However, using multi-source data raises methodological questions about how to weight heterogeneous data that can differ greatly in volume, spatio-temporal resolution, sparsity, quality, and even nature.
Quantifying the predictive uncertainty affecting model simulations (water levels, velocities, and discharges), and decomposing it into different sources (data, model, estimated parameters), remains largely a research domain. Yet, this is a crucial step for informed decision-making.

The objective of this thesis will be to advance towards a generic tool for estimating the parameters and uncertainties of a hydraulic model based on data from various sources. Specifically, the goal will be to propose simplifications or parametrizations that are compatible with operational applications but sufficiently general to be not specific to any site or modeling code.

Funding

AQUA, INRAE, and CNES