Estimation of reference evapotranspiration using data driven techniques under limited data conditions

Abstract

Estimation of reference evapotranspiration $(E{T_0})$ is a major task in irrigation and water resources development and management. In this context, the FAO56 Penman–Montieth (FAO56PM) equation, one of the most accurate equation is used, which however requires a high number of climatic parameters that are not always available for most meteorological stations. The present study measures the effectiveness of selected machine learning techniques [evolutionary regression (ER), artificial neural network (ANN), multi nonlinear regression (MLNR), and support vector machines (SVMs)] for the estimation of $E{T_0}$ under limited data conditions. Different models were developed for estimating $E{T_0}$, including various sets of daily climatic variables, namely maximum and minimum air temperature, extraterrestrial radiation, sunshine hours, wind speed and relative humidity Statistical evaluation of different results obtained from selected techniques showed that the best performance was obtained when six meteorological variables, i.e., (${T_{max}},\;{T_{min}},\;{R_a},\;{S_h},\;{W_s},\;R{H_{min}}$) were used as input, with highest ${R^2}=0.9931 \,$ by ANN4 followed by ER4 with ${R^2}=0.9886$. The ER and ANN performed similarly though ANN performed slightly better against FAO56PM model. The ER was recommended over ANN because ER gives a physical algebraic expression, which can be further replicated without modelling tools and computers. Further, its gives the intermittent calculation visibility and understanding, also that ER technique is easy to use and more practical in application. However, ANN models need modelling tools and computers with the very skilled user and inside calculations like a black box. ANN also require simulation process to get any output.