Phenological dynamics of irrigated wheat yield prediction using multisensor data sources and machine learning in the Central Highlands of Ethiopia

Predicting wheat yield is crucial for monitoring food security, managing agriculture, and promoting sustainable development, particularly in data-deficient areas, by establishing data infrastructure to improve food security oversight. This study proposed a phenology-based framework for yield prediction utilizing multisensor predictors and highlighted the dynamic changes in predictor dominance across various growth stages. Using Sentinel-2 optical imagery, Sentinel-1 SAR backscatter, and SRTM-derived topographic attributes, we extracted 21 predictor variables across three distinct phenological phases: vegetative, reproductive, and ripening. Field measurements from 170 systematically sampled elementary sampling units (20 m × 20 m, representing 0.04 ha each) provided ground-truth yield values (range: 2.0–7.8 t/ha; mean: 4.49 t/ha) at the ripening growth stage. Three machine learning algorithms, random forest (RF), XGBoost, and LightGBM, were calibrated and validated using an 80/20 training-test split with randomized search hyperparameter tuning. The results revealed a higher observed predictive performance toward later phenological stages, culminating in the ripening phase, where LightGBM achieved the highest observed test-set performance (R 2 = 0.89; RMSE = 0.41 t/ha; MAE = 0.30 t/ha), followed closely by XGBoost (R 2 = 0.88) and RF (R 2 = 0.87). Variable importance analysis was conducted using built-in tree-based importance scores, which represent predictive associations rather than causal effects, revealing a dynamic shift in predictor dominance: water-related indices (MNDWI1 and MNDWI2) and red-edge NDVI combinations were critical during the vegetative and early stages, the senescence-related PSRI became paramount during reproduction, and the NDWI overwhelmingly dominated at ripening, underscoring the predictive association of canopy water content with final yield determination. Topographic variables contributed less to the predictive performance than spectral variables did, particularly during the reproductive and ripening stages, which is consistent with the relatively flat irrigated terrain of the study area. The predicted yield maps captured increasing spatial variability as the crop matured, with standard deviations rising from approximately 0.85 t/ha (vegetative) to 1.18 t/ha (ripening). Notably, these standard deviations reflect model-predicted variability rather than direct measurements of actual crop variability. While constrained to a single agroecology and a single year (one growing season observed at three phenological stages) without crop model integration/deep learning approaches, climate variables, and soil properties and without using SHAP or permutation methods to analyze the contribution of individual predictors, the results demonstrate the potential of strategically timed, providing a strong basis for crop monitoring under data-limited conditions and reconstruction yield prediction for the future.