Estimativa da evapotranspiração real diária na caatinga usando algoritmos de aprendizado de máquina e dados meteorológicos.

Abstract

In this study, regression models were developed and evaluated to predict actual daily evapotranspiration (ET) in the semiarid region of Northeast Brazil (NEB) using data from meteorological stations. Six machine learning algorithms were implemented: Naïve Bayes (NB), Random Forest (RF), CN2, k-Nearest Neighbor (kNN), Support Vector Machine (SVM), and Artificial Neural Network (ANN). The input variables used in the models were total daily precipitation (PPT), mean daily air temperature (Ta), mean daily relative humidity (RH), mean daily vapor pressure deficit (VPD), mean daily horizontal wind speed (WND), mean daily solar radiation (Rs), and mean daily soil temperature (Ts). The analysis covered the period from 2013 to 2016 at two experimental sites with different vegetation cover conditions: Restored Caatinga (CR) and Degraded Caatinga (CD), based on micrometeorological data. Model performance was assessed using several statistical metrics: root mean square error (RMSE), mean absolute error (MAE), symmetric mean absolute percentage error (SMAPE), relative root mean square error (rRMSE), bias, and coefficient of determination (R²). The results showed that the RF model performed best at both sites, with RMSE values of 0.80 mm/day for CR and 1.56 mm/day for CD, corresponding to rRMSEs of only 4% and 5%, respectively. These findings highlight that both vegetation conservation status and the appropriate selection of predictive models are determining factors for the regulation and reliable estimation of ET in the Brazilian semiarid region. In the joint analysis of the data, disregarding vegetation conservation status, the RF model again achieved the best performance, with an RMSE of 1.27 mm/day and an rRMSE of 3%. To reduce model complexity, broaden its applicability, and shorten training time, additional simulations were conducted with the RF model using a reduced number of input variables. Results indicated that, with six variables (PPT, RH, VPD, WND, Rs, and Ts), the model maintained—and even slightly improved—its predictive efficiency, achieving an RMSE of 1.26 mm/day and an rRMSE of 3%. The variable importance analysis, based on performance deterioration, revealed that PPT and RH are the most relevant predictors of ET, while Ta and WND had comparatively less influence on the RF model’s predictive ability in the Brazilian semiarid region.