PREdiction de la Vulnérabilité par Observation de la terre et Intelligence aRtificielle explicable multimodale

Vulnerability is defined as the conditions determined by physical, social, economic and environmental factors or processes which increase the susceptibility of an individual, a community, assets or systems to the impacts of hazards. In the context of climate change and the increase of natural disasters, assessing and forecasting vulnerabilities becomes a critical goal. Remote sensing has emerged as a valuable technology for assessing and monitoring the vulnerability of people and infrastructure in the context of disaster risk management, enabling a more proactive approach to mitigating potential impacts.

Indeed, several studies have demonstrated the relevance of remote sensing data to identify high-risk regions and assess the structural integrity of buildings and critical infrastructure before and after disasters strike, to map socio-economic vulnerabilities by cross-referencing physical data with demographic information, or to assess resilience-centered development interventions in local administrative units where socioeconomic survey data are available. The existing studies remain limited in terms of spatial or temporal extent. Artificial Intelligence appears as a promising solution to assess vulnerability at large scale by turning existing massive remote sensing datasets into meaningful insights.

This research project aims to position AI at the forefront of disaster risk management, providing a scalable, automated, and responsive tool for vulnerability assessment using EO data. Scientific and technical challenges include the desynchronization and heterogeneity of data modalities, the robustness to missing modalities during inference, the high level of performance needed to meet end-user requirements (e.g., 80% accuracy), and the low quantity and quality of labels available to train AI models. Moreover, learning the nuanced concept of vulnerability from available data presents its own set of difficulties.

The work program will include the following steps: 1) conceptual understanding and indicator review (study the theoretical underpinnings of vulnerability in the context of climate change and disaster risk reduction, conduct a comprehensive review of relevant indicators that capture vulnerability effectively); 2) dataset compilation (focus on European datasets due to data availability and relevance to EU decision-makers, collect and compile EO data, including satellite imagery and nightlight imagery, along with textual and tabular datasets for model training); 3) baseline evaluation (establish a baseline performance using existing models and analyze their explainability, develop a data processing and analysis pipeline that integrates AI with EO datasets); 4) development of explainable multimodal AI models (create AI models capable of processing multimodal data (EO, textual, tabular) and providing explainable outputs, perform operational validation of the models to ensure they meet business requirements), 5) integration into a decision-making tool (translate the validated models into a component of a decision-making tool for EU policymakers and responders, contribute to implement a system that can monitor changes in vulnerability indicators, especially following disaster events, to provide real-time insights).

This PhD project is part of the collaborative doctoral programme of JRC (https://joint-research-centre.ec.europa.eu/working-us/collaborative-doc…) for which UBS has been selected as one of the 2 international academic partners on the AI4EO topic. Agreement between JRC and UBS under signature. As such, the candidate will spend 12-18 months in Vannes (IRISA, OBELIX) and 18-24 months in Ispra (JRC, Disaster Risk Management Unit). Grantholders must be nationals of a Member State of the European Union or an Associated Country or must have resided in a Member State for at least five years prior to the start of the contract.