A Multi-crop Ensemble Framework for Yield Prediction

Accurate global crop yield predictions are crucial for ensuring food security, effective agricultural planning and climate adaptation. However, existing machine learning and deep learning approaches lack crop-specific feature learning, uncertainty quantification and multiscale spatial contexts, which limits their application in precision agriculture.

This study presents a hybrid ensemble deep learning framework integrated with a crop-aware transformer encoder with heteroscedastic uncertainty for global multi-crop yield prediction (Maize, rice, wheat, soybean) at a 5-arcminute (~9 km) global resolution. The architecture integrates three key innovations: (1) conservative feature engineering using historical yields (4-year sequences with a 3-year temporal gap), geographic coordinates, climate zone indicators alongside temporal trends and stress indicators; (2) crop-aware multi-head attention different mechanisms with crop-specific Query/Key/Value projections enabling differential pattern learning per crop and (3) heteroscedastic output heads predicting both mean yield (μ) and uncertainty (σ) via negative log-likelihood optimization. We implemented rigorous validation using temporal splitting (training: years ≤2013; test: years >2013) and geographic blocking (5-fold GroupKFold with 0.5° spatial blocks) to prevent both spatial and temporal data leakage.

Evaluated on the GlobalCropYield5min dataset (1982-2015) across four crops with 60,000 samples in 34 years. The proposed model achieved an overall test R² of 0.9281, RMSE of 0.585 t/ha and MAE of 0.379 t/ha, with statistical significance confirmed by a paired t-test (p=0.011). Five-fold geographic cross-validation yielded R² of 0.9337±0.0074 and rice R²=0.9462 (best), maize R²=0.9268, wheat R²=0.9201 and soybean R²=0.8298. Uncertainty quantification achieved excellent calibration (Expected calibration error = 0.024), with empirical coverage matching theoretical values (68% intervals: 69.1% coverage; 95% intervals: 94.8% coverage). Regional analysis showed consistent performance across continents (R²=0.871-0.941), with data-scarce regions showing the expected performance reduction. Ablation studies confirmed that crop-aware attention contributed +3.4% to R², multiscale spatial features contributed 5.8% and temporal sequence features contributed +3.66%.