TerraNova: A Foundation Model Coupling Earth and Societal Data

TerraNova, a newly developed foundation model, seeks to represent the Earth and human societies as an interconnected system, tackling a significant issue of the Anthropocene. Detailed in a paper available on arXiv (2607.29527), the model utilizes 1,024 records from both physical and societal domains, comprising 512 Earth-system fields and 512 national indicators, all in their original geometries. A primary advancement is its ability to resolve the geometric discrepancies between continuous Earth-system data and societal information based on administrative units, which previously necessitated lossy averaging. TerraNova employs specialized encoders for location, country, time, and task, integrating them through cross-modal transformers into a unified spatiotemporal framework. Additionally, a hypernetwork creates a per-query decoder with an evidential head to provide a predictive distribution, while two contrastive objectives link the representations, allowing for accurate modeling across borders without information loss.