Evaluation of Tree-Based Machine Learning and Deep Learning Techniques in Temperature-Based Potential Evapotranspiration Prediction

In this study, Extreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), Random Forest (RF), Bagged Trees (BT), and Custom Deep Learning methods were used to estimate the potential evapotranspiration (PET) values at Diyarbakir airport station in the Tigris basin. In establishing the models, the average temperature, maximum temperature, minimum temperature, maximum wind speed, relative humidity, average wind speed, and total precipitation values in the monthly time period were chosen as inputs, and PET values were used as output. The data set is divided into 70% training and 30% testing. 10-fold cross-validation to avoid overfitting problems. Training and test data were randomly selected. The prediction performances of the models were evaluated according to the statistical criteria of determination coefficient (R2), root mean square error (RMSE), mean absolute error (MAE), and rank analysis. The best PET estimates were obtained using the inputs of mean, min, maximum temperature, relative humidity, total precipitation, average, and maximum wind speed. It was also concluded that XGBoost was the highest performance. When the R2 values were examined, it was seen that the Deep Learning model had higher performance. But for RMSE and MAE, XGBoost did better. As a result of the rank analysis, it was seen that XGBoost got a higher score. © 2023, HARD Publishing Company. All rights reserved.