General and regional cross-station assessment of machine learning models for estimating reference evapotranspiration

Abstract

Significant research has been done on estimating reference evapotranspiration ($E{T}_{0})$ from limited climatic measurements using machine learning (ML) to facilitate the acquirement of $E{T}_{0}$ values in areas with limited access to weather stations. However, the spatial generalizability of $E{T}_{0}$ estimating ML models is still questionable, especially in regions with significant climatic variation like Turkey. Aiming at exploring this generalizability, this study compares two $E{T}_{0}$ modeling approaches: (1) one general model covering all of Turkey, (2) seven regional models, one model for each of Turkey’s seven regions. In both approaches, $E{T}_{0}$ was predicted using 16 input combinations and 3 ML methods: support vector regression (SVR), Gaussian process regression (GPR), and random forest (RF). A cross-station evaluation was used to evaluate the models. Results showed that the use of regional models created using SVR and GPR methods resulted in a reduction in root mean squared error (RMSE) in comparison with the general model approach. Models created using the RF method suffered from overfitting in the regional models’ approach. Furthermore, a randomization test showed that the reduction in RMSE when using these regional models was statistically significant. These results emphasize the importance of defining the spatial extent of $E{T}_{0}$ estimating models to maintain their generalizability.

Appendices

Appendix 1: Variable means and standard deviations

This appendix contains the mean and standard deviations of the variables used in this study (Table 4).

Table 4 Mean (standard deviation) of the training and test data

4).

Appendix 2: Model performance metrics

This appendix contains the model performance metrics (Table 5).

Table 5 Model performance metrics

5).