Identification of climate data in ANN applications for estimation of evapotranspiration

Evapotranspiration (ET) is an essential parameter for estimation of irrigation water requirements. Climatic variables (CV) along with soil and plant factors are known to influence ET exhibiting completely different patterns at some locations because of the effect of multicollinearity (MC). These factors, coupled with the complex phenomenon of transpiration, have been largely responsible for the development of a large number of ET models. However, none of the models has been found to be satisfactory for all locations. In the recent past applications of Artificial Neural Network (ANN) have been reported with remarkable success over popular models. But the problem with ANN is two-fold: several trials are involved in training with different combinations of variables, and requirements of large timeseries data. The goal of this paper is to identify minimum number of variables which must be considered for use in the ANN applications. The experiments conducted in this study use reliable climatic data (CD) of four locations of a region whose data have also featured in ANN applications reported elsewhere. It is shown that much of the labor and cost involved in selecting the best combination of variables in ANN can be drastically reduced by using a procedure described in this paper. Importantly, the procedure answers the question of why some models perform poorly and some others do well. The tools employed for development of the procedure on-line are: singular value decomposition (SVD) from linear algebra; and Radial Basis Function Networks from ANN, using a new technique of error analysis on-line.