Thèse Multifloodimpact Jumeaux Numériques Probabilistes Intégrés des Inondations Fondés sur l'Observation de la Terre pour l'Évaluation des Risques et des Impacts Socio-Économiques dans un Context H/F Doctorat.Gouv.Fr

Établissement : Université de Toulouse École doctorale : SDU2E - Sciences de l'Univers, de l'Environnement et de l'Espace Laboratoire de recherche : CECI - Climat, Environnement, Couplages et Incertitudes / CERFACS Direction de la thèse : Sophie RICCI ORCID 0000000242325626 Début de la thèse : 2026-09-01 Date limite de candidature : 2026-11-23T23:59:59 Les inondations figurent parmi les risques naturels aux conséquences humaines et socio-économiques les plus importantes, appelées à s'accentuer dans un contexte de changement climatique. Parallèlement, l'essor de l'observation de la Terre (EO) et des nouvelles missions satellitaires offre des capacités inédites pour observer les surfaces inondées, les niveaux d'eau et la topographie. Le projet de thèse MultiFloodImpact vise à exploiter conjointement ces observations et la modélisation hydrodynamique pour développer une nouvelle génération de jumeaux numériques probabilistes des inondations, capables de quantifier les aléas, leurs incertitudes et leurs conséquences socio-économiques.
La recherche reposera sur quatre axes scientifiques étroitement intégrés. Les observations multi-missions - notamment Sentinel-1/2/6, SWOT, Pléiades et CO3D - seront d'abord exploitées pour reconstruire la topographie et la géométrie des plaines inondables, caractériser les états hydrauliques et construire puis évaluer les modèles hydrodynamiques. Une hiérarchie de modèles multi-fidélité, allant de simulations 2D haute résolution à des représentations hydrauliques simplifiées, sera ensuite développée afin de concilier réalisme physique et coût de calcul. Des méthodes avancées de quantification des incertitudes et de réduction de variance, en particulier les Approximate Control Variates et l'Importance Sampling (ACV-IS), exploiteront les corrélations entre ces différents modèles afin d'optimiser le budget de calcul et d'estimer efficacement les probabilités d'événements rares.
L'originalité de MultiFloodImpact réside également dans l'intégration, au sein d'un même cadre probabiliste, des processus physiques et de leurs conséquences pour les sociétés. Les situations critiques ne seront donc pas uniquement définies à partir de variables hydrauliques telles que la hauteur ou la vitesse de l'eau, l'étendue et la durée de l'inondation ou le dépassement de digues. Elles intégreront également la vulnérabilité et l'exposition des populations, des infrastructures critiques, de l'agriculture et des activités économiques. Le jumeau numérique permettra ainsi de passer de la probabilité d'un aléa à la probabilité de ses impacts, afin de produire une information directement pertinente pour l'évaluation des risques et l'adaptation au changement climatique.
La méthodologie sera développée et évaluée sur quatre sites aux caractéristiques hydrologiques, environnementales et socio-économiques contrastées : la Garonne près de Marmande, l'Alzette au Luxembourg, l'Ohio aux États-Unis et la plaine inondable du Mékong autour de Phnom Penh au Cambodge. Ces démonstrateurs permettront d'étudier la robustesse et la transférabilité du cadre proposé, depuis des bassins européens fortement instrumentés jusqu'à des régions où l'observation satellitaire constitue une source d'information essentielle.
La thèse sera menée en codirection internationale entre le CECI/CERFACS à Toulouse et le Luxembourg Institute of Science and Technology (LIST). Le doctorant ou la doctorante bénéficiera d'un environnement fortement interdisciplinaire associant hydrodynamique numérique, télédétection satellitaire, intelligence artificielle, calcul haute performance, quantification des incertitudes et analyse des impacts socio-économiques. Les séjours de recherche entre les deux institutions favoriseront l'acquisition de ce profil multidisciplinaire. À terme, MultiFloodImpact vise à transformer les observations satellitaires en informations probabilistes directement mobilisables pour la gestion des risques, la résilience des infrastructures, l'adaptation au changement climatique et le développement de futurs services climatiques.
Floods are among the natural hazards whose frequency, intensity and socio-economic consequences are expected to increase under climate change. Anticipating these events requires not only accurate simulations of flood dynamics but also robust estimates of the probability of hazardous situations and their associated impacts. Decision-makers increasingly need probabilistic information rather than deterministic scenarios to support risk management, adaptation planning and operational crisis response.
In recent years, EO has been undergoing a profound transformation. The Copernicus Sentinel missions, SWOT [1], Sentinel-6, Pléiades Neo and CO3D now provide unprecedented observations of flood extent, water surface elevation and topography, at different spatial and temporal scales [2]. These observations create unique opportunities to develop the next generation of flood digital twins capable of continuously integrating observations and numerical simulations.
Despite these advances, current flood risk assessment frameworks still face several scientific and operational limitations. High-resolution two-dimensional (2D) hydrodynamic simulations remain computationally demanding for ensemble approaches for uncertainty propagation and rare-event probability estimation. Conversely, simplified models allow large ensembles but often fail to represent the hydraulic complexity governing flood propagation. Moreover, although multi-mission EO datasets are increasingly available, they are generally used either for calibration or validation, rather than being fully integrated into probabilistic modelling frameworks. Finally, physical flood hazards and socio-economic impacts are still too often investigated separately, limiting the development of integrated risk assessment frameworks.
Recent research has addressed complementary aspects of this challenge. Multi-level statistical methods and variance-reduction techniques have demonstrated their ability to drastically decrease the computational cost of rare-event probability estimation while preserving estimator accuracy. Parallel developments in hydrodynamic modelling have enabled increasingly realistic flood simulations, whereas advances in EO and photogrammetry now facilitate the 3D reconstruction of high-resolution terrain models and flood monitoring from satellite imagery. Other studies have investigated uncertainty propagation in levee failures and flood hazard assessment or the estimation of socio-economic flood damages. However, these advances remain largely disconnected, and no integrated EO-informed probabilistic flood digital twin framework currently combines EO, multi-fidelity hydrodynamic modelling, uncertainty quantification and probabilistic socio-economic impact assessment.
Several complementary studies have contributed to advancing probabilistic flood hazard assessment, although they generally address only isolated components of the overall problem. Maurin et al. [3] developed one of the first comprehensive probabilistic frameworks for assessing the reliability of Loire River levees (France). Their work combines hydraulic loading, geotechnical characteristics and structural inspection data to estimate the probability of several failure mechanisms, highlighting the importance of probabilistic approaches for flood risk management. However, the hydraulic forcing remains based on predefined model configurations and does not account for uncertainties related to floodplain geometry or EO-derived information.
Building upon this perspective, Goeury et al. [4] investigated uncertainty propagation in 2D hydrodynamic simulations by considering uncertainties in levee breach characteristics and roughness parameters within a Monte Carlo framework. Their work demonstrated the strong influence of these uncertainties on predicted flood depths and inundation extents, particularly for extreme events. Importantly, the authors identified the representation of floodplain geometry as one of the main remaining sources of uncertainty and highlighted the need to better account for topographic uncertainties in future studies.
More recently, Espoeys et al. [5] introduced an innovative multi-fidelity [6] statistical framework combining Approximate Control Variates and Importance Sampling (ACV-IS) for efficient rare-event probability estimation in hydrodynamic flood simulations. By exploiting correlations between numerical models of different fidelity, this approach dramatically reduces computational cost while maintaining the accuracy of probabilistic estimators. The methodology was successfully demonstrated using two levels of TELEMAC-2D (www.opentelemac.org, [7]) models differing in their representation of floodplain geometry, opening new perspectives for uncertainty quantification in computational hydraulics.
MultiFloodImpact will build directly upon these pioneering developments while considerably extending their scope. Rather than relying on only two levels of numerical fidelity, the project will investigate a hierarchy of hydrodynamic models generated from progressively simplified representations of floodplain geometry reconstructed from EO. Multi-mission EO will be exploited throughout the modelling chain for terrain reconstruction, calibration and validation, while probabilistic definitions of failure will be extended beyond hydraulic thresholds to include structural vulnerability and socio-economic impacts.
MultiFloodImpact therefore proposes an integrated probabilistic flood digital twin framework bringing together scientific disciplines that are currently addressed separately. Instead of considering Earth Observation, hydrodynamic modelling, uncertainty quantification and socio-economic impact assessment as independent research areas, the project will integrate them into a unified probabilistic flood digital twin framework capable of propagating uncertainties from multi-mission Earth Observations to decision-oriented impact indicators.
Beyond methodological advances, the MultiFloodImpact framework will establish a direct connection between physical flood processes and societal consequences through probabilistic definitions of failure ranging from hydraulic variables and structural vulnerability to socio-economic losses. It will contribute to the next generation of climate services, probabilistic flood forecasting and operational disaster-risk management. More broadly, the project will demonstrate how digital twins can transform multi-mission EO into decision-ready information supporting climate adaptation and resilience.
Floods are among the most devastating climate-related hazards, with increasing frequency and socio-economic consequences under global change. Although high-resolution hydrodynamic models and the rapid expansion of Earth Observation (EO) missions now provide unprecedented opportunities to monitor and simulate flood events, the computational cost of probabilistic simulations and the fragmented exploitation of multi-source observations still limit the development of operational climate services and next-generation flood digital twins.
The objective of the MultiFloodImpact project is to train a PhD candidate at the interface of hydrodynamic modelling, remote sensing EO, Artificial Intelligence (AI), uncertainty quantification and socio-economic impact assessment to develop an integrated EO-informed probabilistic flood digital twin framework. The overall architecture of the proposed MultiFloodImpact framework is illustrated in Figure 1, which summarises the four scientific pillars and the progressive transformation of multi-mission EO into decision-ready probabilistic flood-risk information.
The originality of MultiFloodImpact lies in its truly interdisciplinary approach. The PhD candidate will investigate the physical processes governing flood dynamics and bridge these physical sciences with socio-economic impact assessment by translating probabilistic flood hazards into indicators of potential consequences for populations, critical infrastructures, agriculture and ecosystems through exposure and damage assessment methodologies. This interdisciplinary framework will contribute to climate resilience by supporting risk-informed decision making and adaptation strategies.
This PhD research project addresses a fundamental scientific challenge in uncertainty quantification for flood modelling: how to accurately estimate rare-event failure probabilities from computationally expensive hydrodynamic simulations. Rather than restricting the notion of failure to a single hydraulic criterion, the project will investigate a generalized probabilistic framework in which failure may be defined either by physical hazard indicators (e.g., water depth, flow velocity or dike overtopping) or by impact-based criteria (including vulnerability and exposure) describing potential socio-economic consequences. To address this challenge, the project will develop a novel framework combining EO-derived multi-fidelity hydrodynamic models with Approximate Control Variate (ACV) estimators. The research will investigate the construction of optimal model hierarchies, the propagation of epistemic uncertainties, and the statistical coupling between low- and high-fidelity models required to achieve reliable probability estimates at affordable computational cost.
Leveraging Copernicus and next-generation satellite missions (Sentinel-1, Sentinel-2, Sentinel-6, SWOT, Pléiades and CO3D), the MultiFloodImpact framework will provide an end-to-end methodology, from satellite data exploitation and hydrodynamically consistent terrain reconstruction to multi-fidelity hydrodynamic simulation, uncertainty propagation and probabilistic socio-economic impact analysis. Beyond methodological advances, the integrated MultiFloodImpact framework aims to deliver scalable and transferable probabilistic digital twins supporting future climate services, operational flood-risk management and evidence-based climate adaptation.
The PhD candidate will benefit from the complementary expertise of CECI (France) and the Luxembourg Institute of Science and Technology (LIST, Luxembourg) through a joint international supervision. CECI will contribute its expertise in physics-based hydrodynamic modelling, uncertainty quantification, high-performance computing and operational flood forecasting, while LIST will provide expertise in EO for disaster risk monitoring and modelling, as well as Machine Learning and AI. The mandatory mobility between both institutions will allow the candidate to acquire a unique multidisciplinary profile spanning numerical modelling, satellite remote sensing, AI for EO and probabilistic impact assessment.
1- Scientific methodology
The overall architecture of the integrated MultiFloodImpact probabilistic flood digital twin framework is illustrated in Figure 1, where the four pillars interact to progressively transform satellite observations into probabilistic flood risk assessments supporting climate resilience.
The project capitalises on the complementary expertise of CECI (France) and LIST (Luxembourg), creating a unique interdisciplinary research environment at the interface of EO, computational hydraulics, uncertainty quantification and socio-economic impact assessment. Together, the two institutions provide the scientific foundations required to develop the next generation of probabilistic flood digital twins. While the expertise of the two institutions is complementary, each pillar is jointly developed to ensure strong scientific integration throughout the project.
Pillar 1 - Earth Observation and Hydrodynamic Terrain Reconstruction
The first pillar aims at exploiting the rapidly growing capabilities of EO to generate hydrodynamically consistent representations of floodplains. High-resolution Digital Terrain Models (DTMs) will be reconstructed from Pléiades and CO3D stereoscopic imagery using the HydroDTM [8] framework, combining photogrammetry reconstruction (CARS https://cars.readthedocs.io/), land-cover classification (SLURP), DSM-to-DTM conversion and filtering (Bulldozer https://github.com/CNES/bulldozer), riverbank detection and bathymetry reconstruction (Cephee https://gitlab.com/ludoviccassanCECI/cephee). Complementary flood observations from Sentinel-1, Sentinel-2, SWOT and Sentinel-6 will provide flood extent and water surface elevation products used throughout the project for hydrodynamic model calibration and validation.
Beyond supporting hydrodynamic modelling, this pillar will generate harmonised analysis-ready EO datasets that will constitute valuable resources for future developments in AI-assisted flood mapping, surrogate modelling and physics-informed machine learning [9]. It will establish a hierarchy of geometrical representations with different levels of geometric fidelity, forming the basis of the multi-fidelity hydrodynamic MultiFloodImpact framework.
This pillar will primarily build upon LIST's expertise in EO and geospatial Artificial Intelligence, while benefiting from CECI developments on HydroDTM and hydrodynamically consistent terrain reconstruction.
Pillar 2 - Multi-Fidelity Hydrodynamic Digital Twins
The second pillar focuses on developing a hierarchy of hydrodynamic models with progressively decreasing levels of fidelity. Starting from high-resolution TELEMAC-2D models, with their topography derived from LiDAR or HydroDTM products, then to simplified 2D models, storage-cell, 1D models, and finally 0D representations (Manning equations) will be constructed while preserving the dominant hydraulic processes governing flood propagation.
Rather than defining fidelity solely through numerical simplifications, MultiFloodImpact will investigate multiple sources of fidelity arising from both physical model complexity and the quality of floodplain geometry reconstructed from EO data. This original approach will enable the generation of scalable probabilistic digital twins adapted to different levels of available data and computational resources.
CECI will lead the development of the multi-fidelity hydrodynamic MultiFloodImpact framework, while LIST will contribute to the exploitation of EO products required to generate transferable digital twins across different geographical and environmental contexts.
Pillar 3 - Probabilistic Uncertainty Quantification
The third pillar addresses one of the major scientific challenges of probabilistic flood modelling: accurately estimating rare-event probabilities under computational constraints.
Building upon the multi-fidelity ACV-IS methodology developed during the PhD of Romain Espoeys at CECI, the project will extend variance-reduction techniques to multiple fidelity levels. Correlations between numerical models will be exploited to optimise computational effort while preserving estimator accuracy. Both epistemic uncertainties (geometry, bathymetry and hydraulic parameters) and aleatory uncertainties (hydrological forcing and stochastic processes) will be propagated throughout the modelling chain to generate robust probabilistic estimates of flood occurrence and associated impacts.
This pillar directly capitalises on CECI expertise in uncertainty quantification and advanced statistical methods while extending these developments towards EO-informed probabilistic flood digital twins.
Pillar 4 - Risk, Impact Assessment and Climate Resilience
The fourth pillar extends probabilistic flood modelling beyond hydraulic hazard assessment towards integrated flood-risk analysis.
Several complementary definitions of failure will be investigated, ranging from hydraulic variables (water depth, flow velocity, inundation duration and flood extent) and structural criteria (e.g. levee overtopping) to socio-economic indicators derived from land use, exposed assets, critical infrastructures and damage functions. The resulting probabilistic products will quantify not only where flooding is likely to occur but also where significant societal impacts are expected.
This pillar capitalises on ongoing interdisciplinary collaborations developed at both institutions. At CECI, recent research on socio-economic flood impact modelling, conducted in collaboration with Université Toulouse Jean Jaurès and CNES through the PhD work of Théo Garin [10], provides the foundations for coupling probabilistic flood hazards with exposure and damage assessment. Complementarily, LIST has developed recognised expertise in risk assessment methodologies applied to complex socio-technical systems, including the evaluation of flood impacts on critical infrastructures as well as cascading risks associated with physical disruptions and cyber-attacks affecting interconnected energy networks. Bringing these complementary perspectives together will allow the PhD candidate to investigate flood risk beyond physical hazards alone, towards integrated assessments of societal vulnerability and climate resilience.
This multidisciplinary pillar therefore bridges hydrodynamic modelling, EO, environmental sciences and socio-economic analysis, contributing to operational flood-risk management and next-generation emergency response services.
2- Proposed Use Cases and Demonstration Sites
To ensure both methodological robustness and operational relevance, the MultiFloodImpact framework will be developed and evaluated over four complementary study areas representing a wide diversity of hydro-climatic, environmental and socio-economic contexts. Rather than relying on synthetic benchmark cases, the PhD will capitalize on existing hydrodynamic models and extensive multi-mission EO datasets already available within the consortium, allowing the research to focus on methodological innovations.
The first demonstration site is located in France along the Garonne River near Marmande. This catchment constitutes a particularly valuable testbed because it already benefits from high-resolution TELEMAC-2D model, detailed topographic and bathymetric datasets, extensive Sentinel-1, Sentinel-2 and SWOT observations, and numerous in-situ measurements collected during previous projects. It hs already been investigated for hydrodynamic modelling and data assimilation, providing excellent baselines for the developments proposed in MultiFloodImpact. In addition, this French case study offers access to comprehensive national geospatial and socio-economic databases (IGN, INSEE) together with operational flood damage functions developed by INRAE and distributed by the French Ministry of Ecology, enabling quantitative assessments of direct socio-economic impacts on buildings, agriculture and exposed assets.
A second demonstration site is located in Luxembourg in the Alzette catchment, a small-sized basin characterized by significant urbanized areas and rapid hydrological responses. The consortium already disposes of a LISFLOOD-FP hydrodynamic model based on a high-resolution LiDAR-derived DTM. Upstream boundary conditions can be prescribed either from observed river discharge or from simulations produced by a SUPERFLEX hydrological model. Owing to its limited spatial extent, strong urban influence and higher sensitivity to local hydraulic controls, this catchment provides a complementary testbed to the larger French river systems and offers additional scientific challenges for flood modelling, uncertainty analysis and impact assessment in urban and peri-urban environments.
A third demonstration site is located along the Ohio River (USA), where the consortium has already developed a calibrated TELEMAC-2D hydrodynamic model together with a comprehensive archive of Sentinel-1, Sentinel-2, Landsat-8/9, and SWOT observations. This case study will provide an opportunity to evaluate the transferability of the proposed methodologies in a different climate, land-use, and institutional context, where flood mapping products, topographic datasets and socio-economic information differ from the European framework.
Finally, the Mekong Basin near Phnom Penh in Cambodia will constitute a particularly challenging demonstration site representative of many data-scarce regions worldwide. In contrast with the French and American study areas, access to topography, bathymetry, and in-situ data (of water levels and discharge) are much more limited, although valuable hydrological observations are available through the Mekong River Commission. This context makes satellite observations a key source of information for terrain reconstruction, flood monitoring, and hydrodynamic model calibration and validation, providing an ideal testbed for evaluating the added value of digital twins in regions where conventional monitoring networks remain sparse. This case study therefore provides an ideal framework to demonstrate the added value of integrating multi-mission EO within probabilistic flood digital twins. Beyond flood hazard assessment, the Cambodian case also opens perspectives toward broader environmental and societal applications, including water resource management, water quality monitoring and the analysis of flood-related public health issues such as the propagation of water- and vector-borne diseases.
Together, these four demonstration sites span a broad spectrum of environmental conditions, data availability and socio-economic contexts, ensuring that the proposed MultiFloodImpact framework is both scientifically robust and readily transferable to operational flood-risk assessment across Europe and beyond.
3- PhD research roadmap
The PhD research programme is organised as the progressive implementation of the four scientific pillars of the MultiFloodImpact framework. Each stage builds upon the previous one, ensuring a continuous increase in scientific complexity, from the exploitation of EO data to the development of an operational probabilistic flood digital twin supporting climate-risk assessment.
Stage 1 - Pillar 1: Earth Observation and Hydrodynamic Terrain Reconstruction
The PhD candidate will develop expertise in processing multi-mission EO datasets, including Copernicus data (Sentinel-1, Sentinel-2, Sentinel-6), Pléiades, CO3D, and SWOT observations over test catchments. The candidate will generate and work with hydrodynamically-consistent terrain models, reconstruct river bathymetry and become familiar with numerical mesh generation and hydrodynamic model preparation.
Stage 2 - Pillar 2: Multi-Fidelity Hydrodynamic Digital Twins
Leveraging these datasets, the candidate will develop a hierarchy of hydrodynamic models with multiple levels of geometric fidelity. High-resolution TELEMAC simulations will serve as references for constructing simplified 2D, storage-cell, 1D and 0D representations. Calibration and validation against EO products will ensure consistency across fidelity levels while evaluating the trade-off between computational efficiency and physical realism.
Stage 3 - Pillar 3: Probabilistic Uncertainty Quantification
The candidate will then develop upon the existing ACV-IS methodology taking into account multiple fidelity levels. Correlations between numerical models will be analysed to optimise computational budgets, while uncertainty propagation strategies will be developed to estimate rare-event probabilities under realistic computational constraints. This stage will lead to the implementation of the probabilistic core of the MultiFloodImpact digital twin.
Stage 4 - Pillar 4: Risk, Impact Assessment and Climate Resilience
Finally, the candidate will integrate multiple definitions of failure based on hydraulic variables, structural vulnerability and socio-economic indicators. The probabilistic flood digital twin will be evaluated using multi-mission EO and applied to estimate both flood hazards and their associated impacts, demonstrating its potential for disaster risk assessment and operational decision support.
Throughout the project, alternating research stays between CECI and LIST will progressively expose the PhD candidate to complementary scientific cultures in computational hydraulics, remote sensing, Artificial Intelligence and probabilistic risk assessment. This international mobility constitutes an integral component of the research strategy and will enable the development of a genuinely interdisciplinary scientific profile.
3- Integration within the MultiFloodImpact framework
As illustrated in Figure 1, the four scientific pillars are tightly interconnected within the MultiFloodImpact probabilistic flood digital twin framework. Multi-mission EO provides the geometric, topographic and hydrological information required to construct and validate a hierarchy of hydrodynamic models with progressively decreasing levels of fidelity. These models are combined with advanced uncertainty quantification techniques to estimate probabilistic flood hazards, which are subsequently translated into hydraulic, structural and socio-economic impact indicators supporting climate services, disaster-risk management and evidence-based adaptation strategies.
The resulting MultiFloodImpact framework will deliver an end-to-end methodology capable of transforming multi-mission EO data into decision-ready probabilistic flood information for probabilistic flood forecasting, impact assessment and climate resilience. Beyond methodological advances, MultiFloodImpact will establish a transferable scientific framework where EO, physics-based modelling, uncertainty quantification and AI mutually reinforce each other. By generating harmonised analysis-ready datasets, probabilistic digital twins and scalable modelling tools, the project will also provide a unique foundation for future developments in AI-assisted flood forecasting, physics-informed machine learning and next-generation climate services.