Solving the scale problem in multimodal learning in AgriAI

Geographical scale is intrinsic to geospatial data, especially in agriculture, which relies on heterogeneous datasets for yield estimation, soil moisture prediction, and crop mapping. Multimodal foundation models offer transferability but face scale-related challenges when fusing data with diverse spatial, temporal, and semantic resolutions. We argue that the next generation of Agricultural Foundation Models must leverage existing scientific knowledge from biology, physics, and agriscience, thereby embedding crop, soil, and hydrometeorological process knowledge into scale-aware architectures. Such models can generalize better across crops, climates, and regions while preserving interpretability. We outline a vision for integrating multi-scale learning with mechanistic constraints to enable robust AgriAI under data scarcity and geographic variability.