Drift ≠ Error: Reliability Analysis of Agricultural Foundation Models Under Distribution Shift

Recent geospatial foundation models (GFMs) enable accurate crop yield prediction across disparate regions and scales. Yet real deployments must remain reliable in the face of distribution drift, which can arise from temporal, label, concept, domain, or representation shifts. This paper focuses on reliability: we build upon our previously published Fine-Tuning Agricultural Regression Models (FARM) framework—which adapts the Prithvi-EO-2.0-600M Vision Transformer for dense canola yield estimation over the Canadian Prairies—and develop a systematic drift monitoring and analysis pipeline. Our framework measures Mahalanobis, cosine, and Euclidean distances in both input and latent spaces to characterize distributional shifts and study their relationship with prediction error. Experiments on county-level and 30~m precision-agriculture datasets across normal (2022) and drift (2021) growing seasons show that both input and latent-space distances provide early warning of drift, but that large drift scores do not necessarily correspond to large prediction errors: some out-of-distribution samples are predicted accurately, while some in-distribution samples incur high error. These results reveal a critical gap between drift detection and uncertainty estimation, and highlight the need for reliability-aware monitoring strategies when deploying agricultural foundation models in non-stationary environments.