FARM-HPC

FARM-HPC develops a multimodal geospatial–temporal AI foundation model to detect crop emergence failure (percentage and area hectare damage in field, non-emerged) within the first 20–25 days after planting, enabling rapid, objective replanting insurance decisions. The model fuses 1 meter super-resolved Sentinel-2 from ESA (2.5-day revisit), Sentinel-1 SAR (radar data), ERA5-Land sequences and European soil datasets into a harmonised 30-year data-cube for any given agricultural field. 

Using a transformer architecture, the system learns how early-season weather, soil and canopy signals translate into emergence success or failure, producing a calibrated emergence loss ratio suitable for replanting decisions and actuarial modelling (providing a materially lower-basis-risk index for parametric crop insurance). The project requires EuroHPC AI Factory resources due to PB-scale data volumes, long temporal sequences and the need for multi-node distributed GPU training. FARM-HPC establishes a foundation model for early-season crop risk that can be extended to additional European crops and climate-sensitive perils.