Random forest techniques for spatial interpolation of evapotranspiration data from Brazilian’s Northeast

Abstract

Irrigated agriculture is the human activity responsible for the highest consumption of water from the environment. The evapotranspiration represents, in practice, the consumption of water by a culture and the quantitative information of this parameter assists in a large number of water management problems. By employing spatial interpolation methods, one can determine evapotranspiration in places where there is no information of this parameter. This work evaluates algorithms for spatial interpolation of evapotranspiration data in terms of precision and performance. It compares conventional strategies as the Inverse Distance Weighting (IDW) and Ordinary Kriging (OK), and machine learning strategies, represented by the Random Forest (RF) and a Random Forest variation for spatial predictions (RFsp). The evaluation uses data of evapotranspiration from automatic meteorological stations located in the northeast region of Brazil, in January 2017. The leave-one-out cross-validation method was used to compare the precision of each interpolation algorithm. The results showed that RF obtained better results for the estimation of the reference evapotranspiration than conventional approaches, coming to reduce to roughly half the error obtained. Despite to resolve blocky artifacts present in RF spatial distribution, RFsp did not present better results than RF, obtaining a very similar performance to IDW and OK. Concerning computational performance, IDW obtained the shortest time to create the interpolation model. The IDW also obtained the smallest prediction times.