Prediction of reference evapotranspiration in northwestern Africa with limited data using factorial and SVM regressions

Abstract

Accurate estimating of reference evapotranspiration (ETo) is of great importance for achieving irrigation scheduling and water management studies. In the absence of lysimeters measurements, which is common in many developing countries, ETo estimating represents a real challenge for the water planners community, they often have recourse to less required-data empirical equations that are frequently associated with low accuracies, or to the well-known Food and Agriculture Organization Penman–Monteith (FAO–PM) equation that necessitates a large number of climatic parameters. The present study aims to develop regional data-driven models for ETo estimating that require a limited number of measured climatic inputs (MCI) coupled with geographic coordinates as auxiliary variables. It explores two modeling methods, namely factorial (FR) and support vector machines (SVMR) regressions. The used data concerned 45 meteorological stations, situated in different climatic zones in northwestern Africa, gathered from FAO databases. Pearson matrix of correlation coefficients was used to explore the most important input combinations. The obtained FR and SVMR models were evaluated relative to FAO–PM equation estimates using the root mean square error (RMSE), the correlation coefficient (R), the RMSE–observations standard deviation ratio (RSR), and the mean absolute error (MAE). The results showed that both explored methods gave satisfactory results with a slight superiority of the SVMR that gave more accurate models. Four models (two FRs and two SVMRs) were pointed out depending on the number of MCI. The best models including two MCI were, FR-BS11 and SVMR11, with RMSEs values of 0.28 and 0.29 mm day⁻¹, respectively; those including three MCI were, SVMR8 and FR-BS8, with RMSEs values of 0.19 and 0.20 mm day⁻¹, respectively. The overall results were useful when dealing with limited MCIs in the study area.