Decoding the climate–potato link: six decades of teleconnection-driven European potato yield variability revealed by explainable machine learning

Abstract

Understanding how large-scale climate teleconnections affect staple crop yield is essential for anticipating production risks and strengthening food security in a warming world. This study examines the influence of major teleconnection patterns on potato yield across Europe, where potatoes remain a key component of the food supply. To account for differences in cropping calendars, Europe was divided into three regions (North, Center, South) based on potato growing season timing. Nine machine learning techniques were applied to model regional yield responses using a dataset spanning 1961–2023, incorporating multiple teleconnection indices with lags of up to 36 months. Among all models, the Light Gradient Boosting Machine (LightGBM) performed best in every region, with the highest accuracy in the South ( R ² = 0.79) and lower but still robust skill in the North ( R ² = 0.62). Explainable artificial intelligence using SHapley Additive exPlanations (SHAP) showed that the East Atlantic (EA) and East Atlantic/West Russia (EA/WR) patterns were consistently the most influential predictors, although their importance varied regionally. Lagged EA indices emerged as particularly significant, indicating a strong climate-memory effect in potato production. Despite the highest prediction accuracy in the South, this region exhibited the lowest historical yield levels, likely reflecting combined heat and water stress. By demonstrating that regional potato yield is strongly modulated by specific teleconnection patterns, these findings highlight the value of teleconnection-aware prediction systems for climate-risk management and underscore the need to integrate climate indices into long-term agricultural and food security planning in Europe.