Insights
In recent years, events such as the floods in Emilia-Romagna in 2023 and Valencia in 2024 have made it clear that flood risk is now a concrete, recurring and systemic issue. It concerns not only emergency response, but also how territories, infrastructure and economic activities can prepare for increasingly unstable climate conditions. Understanding flood hazard therefore does not simply mean describing what happens after an event: it means measuring, modelling and comparing complex physical phenomena to enable those who manage assets and territories to prevent, plan for and reduce the impact of future damage. How far can the water reach? At what depth? With what probability? And how might these conditions change under future climate scenarios? Answering these questions today requires the integration of hydrology, physical modelling, geospatial data, satellite observations and artificial intelligence. In this article, we explore this technical chain in detail: how flood hazard is estimated, which approaches are available, what their strengths are and which limitations must be understood in order to interpret the results correctly.
Estimating flood risk requires three distinct elements, which experts in the field identify as hazard, exposure and vulnerability. Hazard refers to the dangerousness of the physical phenomenon and estimates, for example, how far water may extend in the event of a river overflowing, what level it may reach and how frequently the phenomenon may recur. Exposure defines what is located within that area, such as buildings, networks and people. Vulnerability indicates how fragile these elements are in relation to the intensity of the event. While estimating these three elements is already inherently difficult, climate change introduces an additional layer of complexity by altering the hazard component in the future. Projecting flood risk into the future requires an understanding of what climate conditions may be like over the coming years. To do so, internationally shared climate scenarios are used, known as Shared Socioeconomic Pathways, or SSPs: plausible representations of future conditions based on assumptions regarding emissions, population growth and economic development.
A flood risk assessment today follows a logical chain that starts with assets and produces physical and economic indicators. The first step is to locate the elements being assessed through geocoding: a plant, a section of a network or a public building. This information is then overlaid with hazard maps, which describe where water may spread, how deep it may become and with what probability. This is the technical and scientific core of the analysis. Today, hazard maps are mainly built using two prevailing methodologies: numerical simulations, rooted in the physics of the phenomenon, and artificial intelligence models, which draw on increasingly extensive satellite archives to estimate hazard across broader geographical areas.
The first approach is based on a hydrological model, often a mathematical model, which translates rainfall into discharge: the amount of water flowing through a river at a given time in response to the rain that has fallen across the upstream catchment. A hydraulic model, often configured in one dimension along the river channel or in two dimensions across the entire floodplain, then simulates where the water spreads, how deep it becomes and how fast it flows. These models incorporate real physical constraints, including topography, embankments, bridges and drainage systems, and are extremely valuable for high-resolution local analysis. Their main limitation, however, is scalability: they are slow, costly and difficult to extend across entire river basins or at continental scale.
To meet the need to scale analysis across increasingly large areas, a second approach has emerged in recent years, based on technologies that combine artificial intelligence models with increasingly rich and widespread satellite data. Unlike numerical simulations, which begin with physical equations and detailed local data, this method relies primarily on observations of the territory from above: Synthetic Aperture Radar, or SAR, imagery, which can detect the presence of water even through cloud cover; digital elevation models, which describe the morphology of the land or allow its properties to be derived; and archives of past floods, which document where water actually extended during previous events. In practice, satellites do not directly “measure” flood hazard, but provide continuous spatial information over very large areas, often at continental or global scale, that would otherwise be impossible to cover with the same speed.
This information must then be appropriately processed by an artificial intelligence model trained to recognise the patterns connecting meteorological, climatic, morphological and human-related characteristics to flood hazard. This training process is at the heart of the method: the model is provided with thousands of examples in which both the input conditions, such as rainfall, elevation, urbanisation and proximity to watercourses, and the observed outcome, such as flood extent, depth or velocity, are known. The model gradually learns to recognise which combinations of factors are associated with floods of different intensities, without the developer having to explicitly encode every physical relationship. Once trained, the model can be used for inference: when applied to a new area, it estimates flood hazard based on what it has learned from previous examples. The advantage is speed: once developed, it can produce hazard maps over vast geographical areas much faster than a complete hydrological-hydraulic modelling chain. The limitation is equally clear: the quality of the output depends on the quality and representativeness of the data used to train the model. If a model is asked to estimate flood hazard in a context that differs significantly from the examples it encountered during training, the estimate may be less reliable.
The real potential, however, does not lie in replacing numerical simulations, but in combining the two approaches. Maps based on artificial intelligence and satellite data provide coverage and speed; physical simulations provide local accuracy and interpretability. Used together, they can compensate for each other’s limitations: a satellite-based model can provide broader territorial context on the conditions associated with potential flood hazard, while numerical simulations can validate or provide training data for artificial intelligence estimates, particularly at critical points such as embankments, infrastructure or basins with complex flood-defence systems. Within this hybrid approach, artificial intelligence does not “discover” physical laws unknown to science: it recognises statistical patterns within the data. This is why, in our work at Eoliann, we consider these tools complementary: useful for achieving scale, but insufficient when used in isolation.
The previous paragraphs describe how flood hazard is measured: how far the water may reach, at what intensity and with what frequency. Hazard alone, however, is not sufficient to estimate risk. Two additional pieces of information are required: exposure, meaning what is located within the potentially flooded area, including buildings, infrastructure and people; and vulnerability, meaning how fragile these elements are when faced with a flood of a given intensity. Without exposure and vulnerability, a hazard map remains a description of the physical phenomenon rather than a basis for evaluating its consequences. This is where depth-damage curves come into play: empirical relationships connecting water depth to the expected damage to an asset or activity. By applying them to each exposed asset and appropriately combining hazard and vulnerability information, several economic indicators can be obtained, including Expected Annual Damage, or EAD, the average expected annual loss. This is one of the metrics most commonly used to compare different risks, prioritise interventions or support insurance and investment analysis. The final steps involve projection under future climate scenarios and validation: comparing estimates with observed events or official maps produced by public authorities, where available, to assess whether the model results are sufficiently robust.
Despite significant technological and methodological progress, flood risk assessment never produces an absolute truth. Every assessment depends on the quality of the available data, the resolution of the maps, the representation of embankments, walls and flood defences, the ability of climate models to describe local phenomena and the assumptions adopted for future scenarios. Uncertainty should therefore not be treated as a flaw in the model, but as a structural dimension of the scientific problem. In this article, we address it for methodological completeness: not as an operational output that is already available, but as an essential element for interpreting any risk estimate correctly. When uncertainty is not explicitly quantified, it is still important to acknowledge its existence and indicate where results may be more robust or, conversely, require greater caution. Looking ahead, the use of climate-model ensembles and comparisons between plausible scenarios can provide a better representation of this variability, avoiding the interpretation of a single estimate as a certain value.
Flood hazard assessment therefore does not depend on a single model, but on a chain of scientific choices: which data to use, how to represent the territory, how to integrate satellite observations and physical simulations, how to validate the results and how to distinguish what can be estimated with greater confidence from what requires caution. This is where the quality of the technical work is determined. At Eoliann, we are developing a hybrid approach that combines the physics of the phenomenon, geospatial data, artificial intelligence and validation against observed events, with the aim of making climate risk assessment more scalable, transparent and interpretable. Expertise does not lie in promising certainty, but in building models that can explain complexity more effectively: transforming different sources of data into a more robust understanding of flood hazard, its limitations and the conditions under which it can become useful information for prevention, planning and resilience.
This article was produced by Eoliann’s Climate Risk Models Unit and launches Eoliann Lens, a series of in-depth articles exploring the technology and methodologies behind our work.
Want to go deeper?
