En poursuivant votre navigation sur ce site, vous acceptez le dépôt de cookies dans votre navigateur. (En savoir plus)

PhD student (M/F): Artificial intelligence for monitoring and forecasting fire risks and their ecological impact

This offer is available in the following languages:
- Français-- Anglais

Date Limite Candidature : jeudi 29 février 2024

Assurez-vous que votre profil candidat soit correctement renseigné avant de postuler

Informations générales

Intitulé de l'offre : PhD student (M/F): Artificial intelligence for monitoring and forecasting fire risks and their ecological impact (H/F)
Référence : UMR3589-JEACAL-003
Nombre de Postes : 1
Lieu de travail : TOULOUSE
Date de publication : jeudi 8 février 2024
Type de contrat : CDD Doctorant/Contrat doctoral
Durée du contrat : 36 mois
Date de début de la thèse : 1 octobre 2024
Quotité de travail : Temps complet
Rémunération : 2 135,00 € gross monthly
Section(s) CN : Earth System: superficial envelopes

Description du sujet de thèse

In the context of accelerating global warming, extreme events are increasing. Changes in temperature and drought conditions favor the development of forest and vegetation fires around the Mediterranean basin and, more generally, in France. Current meteorological fire hazard models do not take into account the human hazard and the explanatory factors of vegetation fires are currently only empirically modeled. The aim of this thesis is to introduce new, more accurate forecasting methods by integrating earth observation and machine learning techniques. The contribution of machine learning and its complementarity with traditional modeling approaches will be evaluated. We will evaluate which observations are most relevant to improve risk monitoring and forecasting for different types of landscapes (forests, crops, protected natural areas). Applied to Occitanie and mainland France, the methods will also be validated on a global scale. First, the data deemed necessary (in situ, satellite, atmospheric, etc.) will be collected and pre-processed. Models will be built to answer the research questions of the thesis. They will be spatialized and validated over Occitanie, metropolitan France and the global scale. The thesis work will involve the analysis of observational data (in situ and satellite) and model outputs. Good knowledge of data processing and analysis techniques, machine learning and computer coding (Python, Fortran) is required, as well as knowledge of land surface modeling. Written and oral communication skills are also essential.
# Objectives #
As part of its mission to ensure the safety of people and property, Meteo-France provides operational support to fire-fighters and to the French directorate general for civil protection and crisis management. These actions are the subject of ongoing improvements to the diagnostics and modelling used to assess the meteorological danger of vegetation fires. This project will encourage the development of a research component associated with this service, by introducing new methods for forecasting the risk of wildfire, as well as greater geographical precision. In addition to Occitanie and mainland France, the project will include a global component, to validate the methods in contrasting climatic and geographical situations.
# Method #
A wide range of external data can be integrated into fire risk monitoring models, such as Earth observation data. Deep learning approaches enable these data to be processed more efficiently. Today, these approaches based on satellite data remain disconnected from other models using weather forecasts. Another disadvantage of these two types of approach is that they do not take any data from numerical models of continental surfaces. Yet several variables provided by these models (soil water content at various depths, surface dry biomass, etc.) have a direct link with the factors explaining wildfire, which are currently only modelled empirically. These modelled surface data form a rich set to be exploited in such a context. Combining them with Earth observation data could improve fire risk monitoring models. Meteo-France is developing a land surface model, ISBA (Interactions Soil, Biosphere, Atmosphere), used in various applications (surface conditions for numerical weather forecasting and future climate modeling, simulation of water resources in France). A satellite data assimilation system, LDAS-Monde (Land Data Assimilation System which can be used over any region of the world), enables ISBA model simulations to be corrected by integrating satellite data linked to the variables simulated by ISBA (e.g. LAI, Leaf Area Index). Studies have shown the value of the LDAS approach for monitoring vegetation, droughts and their forecasting. See for example Albergel et al. 2019 for the case of the 2018 heat wave. Work has shown that AI makes it possible to build observation operators for assimilating new satellite data into the ISBA model (Corchia et al. 2023). The LDAS system contributes to a better understanding of the evolution of surface conditions, which are decisive for the risks not only of forest fires but also of agricultural wildfires. This system and the associated model outputs are not currently used in such contexts.
# Expected results #
- Demonstrator of a system for monitoring and forecasting the risk of wildfire in Occitania as a priority, with metropolitan France to follow in a second phase.
- Development of the use of AI: observation operators in a data assimilation system, transfer function between data and products usable in an operational context of fire risk monitoring and forecasting.
- An inventory of relevant data sources for fire risk forecasting and a new annotated dataset.
- Large-scale validation of the methodology and measurement of the contribution of satellite data.

Contexte de travail

The thesis will be carried out in the VEGEO team of the CNRM's Groupe de Météorologie de Moyenne Echelle, in Toulouse, in collaboration with IRIT (Toulouse).

Contraintes et risques


Informations complémentaires

The thesis will be co-supervised by Josiane Mothe (IRIT) and Bertrand Bonan (CNRM).