New machine learning approaches to improve reference evapotranspiration estimates using intra-daily temperature-based variables in a semi-arid region of Spain

The estimation of Reference Evapotranspiration (ET0) is crucial to estimate crop water requirements, especially in developing countries and areas with scarce water resources. In these regions, the impossibility of collecting all the required data to compute FAO-56 Penman-Monteith equation (FA056-PM) makes scientists search new methodologies to accurately estimate ET0 with the minimum number of climatic parameters. In this work, several neural network approaches have been evaluated for estimating ET0 using datasets from five weather stations located in Southern Spain (semiarid region of Andalusia). The assessment of statistical performance (Root Mean Square Error -RMSE-, Mean Bias Error -MBE-, coefficient of determination -R-2- and Nash-Sutcliffe model efficiency coefficient -NSE-) of models namely Multilayer perceptmn (MLP), Generalized Regression Neural Network (GRNN), Extreme Learning Machine (ELM), Support Vector Machines (SVM), Random Forest (RF) and XGBoost were carried out using different input variables configurations. Only temperature-based data were used as inputs; the calculation of new variables called EnergyT (the integral of the half hourly temperature values of a day) and Hourmin (the difference in hours between time sunset and the time when the maximum temperature occurs) had promising results for the most humid stations. The good results obtained with EnergyT when it is used as an input of the system demonstrated that the information contained on it gives detailed characterization of the daily thermic behavior at each location, resulting in a more efficient model than those using only daily maximum, minimum temperature and extraterrestrial radiation values. In general, the modeling results showed that no model firmly outperformed the others, although MLP and ELM were commonly the models that gave the best performances for all sites: mean values of R-2 > 0.89, mean values of NSE > 0.88, mean values of RMSE < 0.67 mm/day and mean values of MBE ranging from -0.17 to 0.30 mm/day. Therefore, EnergyT and Hourmin can be used to estimate ET0 more accurately in stations where data acquisition is limited, like in developing countries or at low-cost weather stations that cannot collect all the required meteorological variables used in FA056-PM. Overall, the use of ELM is recommended due to its high performance in terms of efficiency (NSE) for all the configurations and for all locations, especially using EnergyT as an input variable.

Automated summary

The study evaluates various neural network approaches for estimating Reference Evapotranspiration (ET0) and finds that the use of new variables such as EnergyT and Hourmin can improve accuracy, especially in humid regions. Among the models tested, Multilayer Perceptron (MLP) and Extreme Learning Machine (ELM) stand out as reliable options for accurately estimating ET0 in locations with limited data acquisition.