Significant research has been done on estimating reference evapotranspiration (ET0) from limited climatic measurements using machine learning (ML) to facilitate the acquirement of ET0 values in areas with limited access to weather stations. However, the spatial generalizability of ET0 estimating ML models is still questionable, especially in regions with significant climatic variation like Turkey. Aiming at exploring this generalizability, this study compares two ET0 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, ET0 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 ET0 estimating models to maintain their generalizability.