New approach to estimate daily reference evapotranspiration based on hourly temperature and relative humidity using machine learning and deep learning

Computation of reference evapotranspiration (ETo) poses a challenge under limited meteorological data availability. However, even in this case, hourly data may be available since low-cost sensors can report hourly measurements. This study evaluates, for the first time, in regional and local scenarios, the use of limited hourly meteorological data (temperature and relative humidity or only temperature) to estimate daily ETo directly and by summing hourly ETo values, employing RF, XGBoost, ANN and CNN. The following options were evaluated: (i) use of daily input data (conventional approach); (ii) use of hourly data measured during a 24 h period + hourly extraterrestrial radiation (Ra) to estimate daily ETo directly; (iii) the same configuration of the last option, but with daily Ra instead of hourly Ra; and (iv) use of hourly data to estimate hourly ETo and then to estimate daily ETo by summing hourly ETo. All options used Ra. To develop and evaluate the models, two daily ETo targets were considered: ETod (computed using the daily version of the ASCE-PM equation) and ETosoh (computed by summing hourly ETo obtained with the ASCE-PM equation). Data from 53 weather stations located in the state of Minas Gerais, Brazil, were used. For all models, the best results were found using hourly data to estimate daily ETo directly. CNN models developed with 24 h hourly data + hourly Ra offered the best performance in all cases. In relation to the best models developed with daily data, RMSE reduced by up to 28.2 % (0.71 to 0.51) and NSE and R2 increased by up to 21.7 (0.69 to 0.84) and 11.4 % (0.79 to 0.88), respectively, in regional scenario. In local scenario, RMSE reduced by up to 22.4 % (0.58 to 0.45) and NSE and R2 increased by up to 10.1 (0.79 to 0.87) and 11.3 % (0.80 to 0.89), respectively.