Phenology-Aligned Representation and Normalization for Cross-Country Crop Yield Prediction

Abstract

Accurate crop yield forecasting across heterogeneous geographic regions is significantly hindered by domain shift, driven by regional variations in crop calendars, season lengths, and climatic conditions. Traditional calendar-based time series indexing fails to account for these differences, misaligning physiological growth stages and degrading the cross-country generalization of machine learning models. To overcome this limitation, this thesis introduces PhenoNorm, a phenology-aligned preprocessing pipeline designed to mitigate temporal misalignment. PhenoNorm re-indexes multi-source environmental features, including weather, vegetation indices, and soil moisture, along a unified physiological coordinate system using normalized cumulative Growing Degree Days (nGDD). This transformation enables observations from diverse agro-ecological regions across different countries to be compared at equivalent developmental stages. This study systematically evaluates the interaction of this biological alignment with various feature normalization strategies (global versus country-wise) across multiple model architectures, including Random Forest, XGBoost, LSTM, PatchTST, and a pretrained time series foundation model, Moirai. Empirical results demonstrate that global normalization outperforms country-wise normalization by preserving critical scale differences and informative spatial variations required for cross-domain transfer. Sequence-based models outperformed tabular baselines, with the transformer-based PatchTST achieving the highest overall predictive performance by effectively capturing stage-dependent temporal interactions and early-season vegetation signals. Furthermore, while the Moirai foundation model failed to generalize under domain shift in a zero-shot setting, it demonstrated strong adaptation efficiency and became highly competitive after task-specific fine-tuning. Overall, this research establishes that temporal representation and normalization can strongly affect cross-country generalization.