Machine learning-based sorghum yield prediction using multisource data in challenging agricultural environments of Ethiopia

Sorghum is a major staple crop that contributes substantially to food security and rural livelihoods in Ethiopia. However, its production is increasingly threatened by climatic variability, soil and crop-management constraints, and conflict-related disruptions. This study developed an integrated framework for sorghum-yield classification, continuous-yield prediction, and conflict-scenario forecasting in the West Gondar Zone of Ethiopia. The analysis used 10,473 observations collected from 2012 to 2023, integrating climatic, soil, vegetation, agronomic, and management variables. Conventional machine-learning, ensemble-learning, and deep-learning models were evaluated, with SMOTE used to address class imbalance, one-vs-rest classification applied as a robustness check, and SHAP used to interpret model predictions. For yield classification, XGBoost achieved the highest direct multiclass accuracy of 99.94%, while the one-vs-rest analysis confirmed the consistency of the classification results. SMOTE improved the recognition of less-represented yield classes, particularly in terms of recall and F1-score. For continuous-yield prediction, XGBoost achieved the best performance (R 2 = 0.928), followed closely by Random Forest (R 2 = 0.926) and Gradient Boosting (R 2 = 0.923). Five-fold cross-validation confirmed the stability and generalization ability of the ensemble models. For the 2024–2025 forecasts, XGBoost produced the most consistent performance, with a mean R 2 above 0.910 and comparatively low prediction errors. Scenario analysis indicated potential yield reductions of 10%, 25%, 40%, and 60% under low, medium, high, and severe conflict-exposure conditions, respectively. These values represent scenario-based estimates rather than directly observed yield losses. Overall, the framework demonstrated strong predictive accuracy, model stability, and interpretability. It provides a practical decision-support approach for sorghum-yield assessment, production-risk analysis, resource allocation, and climate- and conflict-resilient agricultural planning in vulnerable farming systems.