Lightweight Federated and Explainable AI Enhanced Crop Yield Forecasting in Heterogeneous Agricultural Systems

Crop yield estimation directly affects food security, resource utilization and agricultural decision making. The complexity of environmental interactions with soil properties and farming practices makes accurate prediction challenging by traditional methods. This paper introduces a unified lightweight framework based on advanced machine learning (ML), explainable AI (XAI) and federated learning to address these challenges. Four machine learning (ML) models (Random Forest, LightGBM, XGBoost and CatBoost) were compared across a multi-crop data. Results indicate that all the models obtained high correlations between observed and predicted yields; however, CatBoost appears to have been the best fit with an $\mathbf{R}^{\mathbf{2}}$ value of $\mathbf{0. 9 9 2}$, the lowest MSE and MAE as well as MAPE with minimal errors. Computational results indicate that CatBoost is efficient both in terms of inference speed and memory usage. SHAP feature analysis revealed relative humidity, soil moisture, sunlight hours and temperature as the most important predictors providing practical information for crops cultivation. Federated learning experiments showed that training in a decentralized fashion kept most of the data private and protected with only a marginal impact on performance, averaging $\mathbf{R}^{\mathbf{2}}$ of 0.977 across 3 rounds in comparison to centralized results. This minimal reduction emphasizes the compromise on precision and data privacy. Still, the method described above yields accurate results and allows for direct reasoning while also keeping the data private and secure, offering a new heterogenous agricultural framework for predicting crop yield that can be used broadly by farmers and policymakers for improving sustainable agriculture practices.