Federated Tree-Ensemble Learning for Crop Yield Prediction in Low-Connectivity Environments

This work addresses the emerging demands for intelligent, privacy-preserving, and decentralized decision-making in rural and low-connectivity agricultural scenarios. It proposes a federated tree ensemble learning framework for accurate crop yield prediction over large-scale, non-IID agricultural datasets. The approach leverages the robustness of LightGBM within a federated setting that maintains the localization of raw data while allowing models to collaborate on global learning. The dataset used in this paper was based on district-level agricultural statistics of India, aggregated to a total of more than 345,000 records, and was divided into ten regional clients to simulate realistic heterogeneity. Experimental results show that the proposed Fed-LGB ensemble reaches performance close to centralized benchmarks, with an R2value of 0.8508, while reducing communication overhead by 56% (from 7.03 MB to 3.10 MB). In addition to accuracy, communication costs, convergence under packet loss, mitigation of client drift, and personalization effects are also studied. The manuscript further provides concise algorithmic descriptions to support reproducibility at critical stages of the federated learning pipeline.