Time, Space, and Modality: Probing Earth Foundation Models for Intra-Field Crop Yield Forecasting

Accurate spatio-temporal crop yield forecasting within fields depends on capturing real crop phenology over time while preserving fine-grained intra-field spatial variability. We evaluate four Earth foundation models (AlphaEarth, OlmoEarth, Presto, and AnySat) against a multi-view gated fusion (MVGF) baseline trained from scratch on Sentinel-2, weather, elevation, and soil data. Across 140 corn and 197 wheat field-years, OlmoEarth paired with a ConvLSTM head achieves the highest field- and pixel-level R-squared on both crops, benefiting from joint self-attention across space, time, and modality. In spatial evaluations, AnySat (pretrained on 11 sensors via scale-adaptive prediction) best matches ground-truth spatial autocorrelation (Moran's I), while OlmoEarth's monthly self-attention best recovers the intra-field yield distribution. These results show that temporal fidelity and spatial detail stem from distinct design choices, highlighting the need for future agricultural architectures to combine calendar-wide temporal attention with flexible, high-resolution tokenization.