Scale-dependent lag effects and interpretable machine learning: Advancing rubber yield prediction

The ever-increasing global demand for natural rubber in the international market has driven its cultivation area to undergo diversified and rapid expansion worldwide over the past decades. However, predicting global rubber yield has long been challenged by unclear influencing mechanisms and low precision. This study addressed this gap by developing a multi-scale, lagged-factor framework for rubber yield explanation and prediction. Focusing on Xishuangbanna, China, daily and monthly yield data from 1995 to 2004 were used to identify optimal lag periods for each factor using Random Forest. To capture complex non-linear interactions, six advanced machine learning models—Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting Machine (GBM), eXtreme Gradient Boosting (XGBoost), Artificial Neural Networks (ANN), and Bayesian Probabilistic Model (BPM)—were implemented and compared. The best-performing model was interpreted using SHapley Additive Explanations (SHAP) to quantify factor contributions and interactions. The results demonstrate that: (1) Driving factors exhibited strong scale dependency. Monthly yield was primarily driven by market price (r = 0.77), while daily yield was dominated by the non-linear effects of short-term physiological factors; (2) the influence of key factors varied with the lag period. Strategy with selecting the optimal lag period for each factor improved the accuracy of all models in predicting rubber yield, underscoring the need to assign factor-specific lag periods; (3) machine learning models exhibited varying performance, with XGBoost achieving the best performance at both the monthly (R² = 0.77, RMSE = 1.0 t, rRMSE = 26.35 %) and daily (R² = 0.49, RMSE = 0.07 t, rRMSE = 45.45 %) scales. Non-linear effects and interactions of factors on rubber yield were also found. This study improves the scientific understanding of the relationship between environmental factors and rubber yield.