Estimation of Monthly Reference Evapotranspiration with Scarce Information Using Machine Learning in Southwestern Colombia

Abstract

This research aimed to identify an alternative method to estimate reference evapotranspiration (ETo) with scarce climatological information in southwestern Colombia between 1983-2017 by evaluating and comparing different machine learning techniques. The FAO Penman-Monteith (FAO-PM56) was used as the reference method and four empirical methods (Hargreaves, Thornthwaite, Cenicafé, and Turc) were assessed with five metrics to evaluate the method of best fit to FAO-PM56, root mean square error (RMSE), mean absolute error (MAE), mean bias error (MBE), Nash-Sutcliffe model efficiency coefficient (NSE), and Pearson correlation coefficient (R). Three models were designed using machine learning techniques to estimate ETo, multiple linear regression (MLR), artificial neural networks (ANN), and autoregressive integrated moving average model (ARIMA). The results showed that the ARIMA-M3 model reported the best performance metrics (RMSE = 4.13 mm month-1, MAE = 3.15 mm month-1, MBE = -0.08 mm month-1, NSE = 0.96 and r = 0.98). However, it restricts in that it can only be used locally and cannot be extrapolated to other climatological stations,because it was calibrated with specific conditions (exogenous variables) and stations,unlike the ANN-M1 model, which only requires training the network for its application. This method will allow estimating ETo in places with scarce information, as vital for water management in places with much uncertainty regarding accessibility and availability.