A day to open the digital platform to other targeted projects
The Digital Platforms targeted project of the Risks program (IRiMa) builds a distributed architecture of data and services designed to pool scientific resources produced by the entire Risks program. The Intelligent Mapping building block, led by Isabelle Manighetti and Elena Di Bernardino at Université Côte d'Azur, contributes to the Digital Platforms infrastructure and proposes a cross-cutting action to most targeted projects of the program. This action aims to develop algorithms based on artificial intelligence (AI) to identify, map and model certain hazards and risks studied in the Risks program.
After one year of activity for the Intelligent Mapping team, the day organized in Nice aimed to present advances and foster dialogue between the team and the rest of the program. Several heads of other targeted projects were present and gave presentations. A wide variety of topics was covered: morphological detections in coastal areas, human behaviors in disaster situations, object vectorization and predictions, hydroclimatic extreme events, detection of seismic faults, gravitational instabilities and submarine collapses, Mediterranean flooding, temporal monitoring of coastal turbidity, detection of extreme climate events by spectral parameters, and modeling of urban heat islands. A cross-cutting section was also devoted to issues of uncertainty and explainability of AI algorithms.
This diversity demonstrates the cross-cutting scope of this functional building block: for a common digital platform to make sense, the tools it will host must respond to the concrete needs of the targeted projects that will use it.
Increasing national capacities in automatic mapping of hazards and risks
Why develop these algorithms? Because at the time the project was launched, there were few or no AI algorithms applied to the automatic mapping of hazards and risks in remote sensing imagery. While certain tools already exist in the international community to automatically identify and map roads or buildings in satellite images, the automatic mapping of more complex objects — tectonic faults responsible for earthquakes, coastal morphologies evolving under the effect of global changes, human behaviors, etc. — remains a blind spot. At the national level, existing aerospace imaging services, such as ISDeform, employ other approaches, particularly the analysis of deformations in radar imagery, complementary to the use of artificial intelligence.
Five initial research projects
The Intelligent Mapping team brings together about twenty researchers specializing in AI, mathematics, modeling, and the various hazards and risks studied. The project aims to cover about ten themes. To date, three topics have truly started, a fourth is just entering the launch phase, and a fifth will begin in a few months.
Detecting rip currents (baïnes) before they trap swimmers:
The first topic aims to automatically map rip currents responsible for the largest number of summer drownings in France and in the world. A doctoral student has been recruited to develop an AI algorithm capable of identifying them in Earth remote sensing imagery (collaboration Inria-Montpellier, Géoazur-Université Côte d'Azur, and Université de Bordeaux). Coupled with drone or aircraft overflights, the tool would enable warning of dangerous zones in near real-time before bathing.
Recognizing distress signals in drone imagery:
When a disaster occurs — earthquake, fire, flood, etc. — drone overflights now produce continuous hours of imagery intended to assist rescue operations. A doctoral student has been recruited to develop an AI algorithm capable of identifying human distress behaviors in these images: for example, a person on a roof making large gestures (collaboration Inria and Géoazur, Université Côte d'Azur). The system can then transmit the image and location to rescue centers on the ground, significantly accelerating the response chain.
From probability to vector:
Most AI algorithms produce pixelated probability maps: such pixel has a 90% chance of being the object of interest, another 60%, etc. To scientifically exploit these results — measure a length, an angle, track an evolution over time — these probability maps must be converted into two-dimensional vectors. An operation less trivial than it appears, on which doctoral work is underway, first tackling a simple case: road vectorization (collaboration Inria and LJAD, Université Côte d'Azur).
Modeling extreme climate events in time series of images and data:
Fourth topic, in launch phase: modeling extreme hydrological events (floods, high water) from time series data. The topic mobilizes mathematical analyses, particularly spectral, in addition to artificial intelligence. A postdoctoral fellow has just been recruited to work on this challenge (collaboration Inria-Montpellier, BRGM, and LJAD-Université Côte d'Azur).
Mapping faults that produce earthquakes:
Fifth topic, which is expected to start in a few months: automatic identification in remote sensing imagery of tectonic faults responsible for major damaging earthquakes. Eventually, the algorithm could receive a satellite image and produce a map of faults likely to produce future earthquakes — a major advance in anticipating seismic risk areas. A postdoctoral fellow is in the process of being recruited for this project (collaboration Irisa-Rennes, and Géoazur-Université Côte d'Azur).
Other topics will enrich this portfolio over the course of the two or three recruitment waves planned in the coming years. Eventually, the project will mobilize about ten doctoral students and postdoctoral fellows, not counting the supervising researchers and engineers.
From fundamental research to operational use
The ultimate objective of this working group is to make the developed algorithms available to the scientific community and decision-makers, via the digital platform of the program. For this, two conditions must be met: that the algorithms are sufficiently robust, and that they are sufficiently generalizable.
The work undertaken aims precisely to design algorithms capable of transfer learning: once trained on an initial dataset, they could quickly learn from new data from other territories or other types of imagery. This capacity for adaptation would allow a risk manager to use an algorithm initially trained on a given area, following a light incremental learning phase on their own data.
The articulation with other targeted projects of the program is progressively being built. If the team is, in early 2026, still mainly in the research phase, the Nice workshop marks an important milestone: making current work known, identifying concrete needs of other projects, and preparing for the moment when the first tools can actually be used.
About the Digital Platforms project
The Digital Platforms targeted project of the Risks program (IRiMa) builds a distributed architecture of data and services to make available to the scientific community and risk actors the resources produced by the program. Intelligent Mapping constitutes one of its pillars, by developing artificial intelligence algorithms to automatically map certain hazards and risks studied in the program.