Explainable Machine Learning for Crop Yield Forecasting within an Agentic AI Decision Support System

Abstract

Accurate crop-yield forecasting is essential for agricultural planning, resource allocation, and risk management under increasingly variable climatic and soil conditions. However, many machine-learning-based forecasting approaches emphasize predictive accuracy without adequately explaining the factors responsible for individual predictions or translating model outputs into actionable agricultural recommendations. This study presents an explainable machine-learning framework integrated with an Agentic Artificial Intelligence decision-support architecture for transparent crop-yield forecasting. A generated agricultural benchmark dataset containing 6,000 observations of rice, wheat, maize, and soybean was developed using climatic, soil, nutrient, irrigation, vegetation, management, and historical-yield attributes. Random Forest, XGBoost, LightGBM, and Artificial Neural Network models were comparatively evaluated using the coefficient of determination, root mean square error, mean absolute error, and mean absolute percentage error. The results demonstrated strong forecasting capability across the investigated models, with boosting-based algorithms exhibiting particularly favourable predictive performance. SHapley Additive exPlanations were incorporated to determine both global feature importance and prediction-specific contributions of agricultural variables. The explainability analysis indicated that vegetation condition, historical productivity, rainfall, nutrient availability, irrigation, and temperature were among the important factors influencing yield forecasts. The proposed Agentic AI architecture integrates Data, Prediction, Explainability, Recommendation, and Monitoring Agents to transform model predictions into context-dependent agricultural guidance. Consequently, the framework progresses beyond conventional yield estimation by providing the expected yield, identifying the factors responsible for the forecast, and generating management-oriented recommendations. The principal novelty lies in integrating comparative crop-yield forecasting, explainable machine learning, and coordinated agent-based reasoning within a unified explain–reason–recommend decision-support pipeline. The developed framework demonstrates the potential of transparent and interpretable artificial intelligence for more informed and accountable agricultural decision support.