January 2027 – December 2028
Humanitarian crises generate evidence across multiple, disconnected data sources. Humans integrate these sources intuitively — but automated systems do not: current systems treat data streams separately, with Earth observation focusing on physical destruction and LLM-based models extracting structured event information from reports.
The CARE project (Cross modal AI for Relief Efforts) aims at creating a system that reasons like an analyst — across text and imagery, jointly. Linking data science and remote sensing laboratories at EPFL and ETH with the ICRC, CARE targets four gaps: the lack of methods to align noisy text reports with imagery, the absence of vision-language models adapted to conflict’s sparse data, untested generalization across diverse conflict settings, and limited validation with humanitarian analysts.
