Performance Evaluation of Regression-Based Machine Learning Models for Modeling Reference Evapotranspiration with Temperature Data

Abstract

In this study, due to their flexibility in forecasting, the capabilities of three regression-based machine learning models were explored, specifically random forest regression (RFr), generalized regression neural network (GRNN), and support vector regression (SVR). The above models were assessed for their suitability in modeling daily reference evapotranspiration (ETo), based only on temperature data (Tmin, Tmax, Tmean), by comparing their daily ETo results with those estimated by the conventional FAO 56 PM model, which requires a broad range of data that may not be available or may not be of reasonable quality. The RFr, GRNN, and SVR models were subjected to performance evaluation by using statistical criteria and scatter plots. Following the implementation of the ETo models’ comparisons, it was observed that all regression-based machine learning models possess the capability to accurately estimate daily ETo based only on temperature data requirements. In particular, the RFr model outperformed the others, achieving the highest R value of 0.9924, while the SVR and GRNN models had R values of 0.9598 and 0.9576, respectively. Additionally, the RFr model recorded the lowest values in all error metrics. Once these regression-based machine learning models have been successfully developed, they will have the potential to serve as effective alternatives for estimating daily ETo, under current and climate change conditions, when temperature data are available. This information is crucial for effective water resources management and especially for predicting agricultural production in the context of climate change.

Keywords

• daily reference evapotranspiration
• random forest regression
• generalized regression neural network
• support vector regression