Potato Crop Yield Prediction: A Data-Driven Federated Learning Approach

Accurate prediction of potato yield is crucial for optimizing agricultural productivity and resource management in Bangladesh, where potatoes play a key role in the economy. This study explores the application of Federated Learning (FL) to predict potato yields across six major agricultural districts in Bangladesh: Chittagong, Dhaka, Dinajpur, Jessore, Mymensingh, and Rajshahi. Unlike traditional centralized models, FL enables decentralized training on local data, ensuring data privacy while capturing region-specific agricultural patterns. A diverse set of machine learning models was evaluated, including Linear Regression, Random Forest, Gradient Boosting, SVR, XGBoost, Extra Trees, AdaBoost, K-Nearest Neighbors, Decision Tree, and LightGBM, with extensive hyperparameter tuning conducted for each district to optimize performance. The local models were aggregated using the Federated Averaging (FedAvg) algorithm, enhancing predictive accuracy through knowledge integration from diverse districts. The evaluation was based on metrics such as Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>). The findings demonstrated that the feder-ated global model achieved an overall MSE of 0.0078 for the validation set and an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.8794, with the lowest district-level MSE observed at 0.0052. The results demonstrate that Federated Learning significantly improves yield prediction while preserving data privacy. This approach provides a scalable, privacy-preserving solution for integrating advanced analytics into agricultural planning, benefiting policymakers and farmers alike. The study’s findings also have broader implications for applying FL in precision agriculture globally.