Federated learning based reference evapotranspiration estimation for distributed crop fields.

Water resource management and sustainable agriculture rely heavily on accurate Reference Evapotranspiration (ETo). Efforts have been made to simplify the (ETo) estimation using machine learning models. The existing approaches are limited to a single specific area. There is a need for ETo estimations of multiple locations with diverse weather conditions. The study intends to propose ETo estimation of multiple locations with distinct weather conditions using a federated learning approach. Traditional centralized approaches require aggregating all data in one place, which can be problematic due to privacy concerns and data transfer limitations. However, federated learning trains models locally and combines the knowledge, resulting in more generalized ETo estimates across different regions. The three geographical locations of Pakistan, each with diverse weather conditions, are selected to implement the proposed model using the weather data from 2012 to 2022 of the selected three locations. At each selected location, three machine learning models named Random Forest Regressor (RFR), Support Vector Regressor (SVR), and Decision Tree Regressor (DTR), are evaluated for local Evapotranspiration (ET) estimation and the federated global model. The feature importance-based analysis is also performed to assess the impacts of weather parameters on machine learning performance at each selected local location. The evaluation reveals that Random Forest Regressor (RFR) based federated learning outperformed other models with coefficient of determination (R2) = 0.97%, Root Mean Squared Error (RMSE) = 0.44, Mean Absolute Error (MAE) = 0.33 mm day-1, and Mean Absolute Percentage Error (MAPE) = 8.18%. The Random Forest Regressor (RFR) performance yields the local machine learning models against each selected site. The analysis results suggest that maximum temperature and wind speed are the most influential factors in Evapotranspiration (ET) predictions.