Key points are not available for this paper at this time.
During dry seasons, accurate predictions of the reference evapotranspiration (ET0) are crucial for effective water management and irrigation. Machine learning (ML) models rely on existing data to make predictions; however, they struggle to perform in new locations where data are insufficient. Methods: This study improved ET0 prediction in diverse locations with limited data by proposing regional scenarios that utilize datasets from a wider region for training. Historical weather data from four California weather stations were used to evaluate classical ML models: linear regression (LR), ridge regression (RR), multilayer perceptron (MLP), and support vector regression (SVR), along with ensemble methods, such as: random forest (RF), extra trees (ETs), extreme gradient boosting (XGB), and gradient boosting regression (GBR). The performance was assessed using the mean absolute error (MAE) and root mean square error (RMSE) in both the local and regional scenarios. Results: ET and GBR showed significant improvements in regional scenarios. After validation at two new stations, ET consistently outperformed GBR as a robust global model for ET0 prediction in new California locations with minimal data. The performance remained near the minimum error, with MAE values of 0.1001 (RMSE: 0.1582) in Ferndale, 0.1494 (RMSE: 0.2279) in Linden, and 0.0974 (RMSE: 0.1495) in Smith River. Conclusion: A regional approach enhanced ML-based ET0 predictions, particularly in data-scarce areas. These findings support the adoption of smart farming and sustainable water resource management.
Patel et al. (Tue,) studied this question.