A Systematic Evaluation of Out-of-Distribution Generalization in Climate-Aware Crop Yield Prediction

TMLR Paper5960 Authors

Abstract: Accurate crop yield forecasting under shifting climatic conditions is essential for food security and agricultural resilience. While recent deep learning models achieve strong performance in in-domain settings, their ability to generalize across space and time—critical for real-world deployment—remains poorly understood. In this work, we present the first systematic evaluation of temporally-aware crop yield prediction models under spatio-temporal out-of-distribution (OOD) conditions, using corn and soybean data across more than 1,200 U.S. counties. We benchmark two representative architectures, GNN-RNN and MMST-ViT, using rigorous evaluation strategies including year-ahead forecasting, leave-one-region-out validation, and stratified OOD scenarios of varying difficulty based on USDA Farm Resource Regions. Our comprehensive analysis reveals significant performance gaps across agro-ecological zones, with some models showing negative R² values under distribution shift. We uncover asymmetric transferability patterns and identify the Prairie Gateway region as consistently challenging for generalization. These findings challenge prior generalizability claims and provide practical insights for deploying agricultural AI systems under climate variability.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Jacek_Cyranka1

Submission Number: 5960