Assessment of a 4-input artificial neural network for ETo estimation through data set scanning procedures

Evapotranspiration is a complex and non-linear phenomenon that depends on the interaction of several climatic parameters. As an alternative to traditional techniques, artificial neural networks (ANNs) are highly appropriate for the modeling of non-linear processes. In general, in the most common ANN applications, the available climatic series are usually split up into 3 data sets: one for training, one for cross-validating, and one for testing. Up to now, the studies regarding ANN-models for reference evapotranspiration estimation and forecasting consider usually only a single chronological assignment of data for the definition of these 3 data sets. In these cases, the ANN performance can only be referred to this specific data set assignment. This paper analyzes the performance of a simple ANN model, a temperature-based 4-input ANN, taking into consideration a complete scan of the possible training, cross-validation, and test set configurations using ‘leave one out’ procedures. The results of a comparative analysis between both methodologies show that the performance results achieved with the traditional methodology can be misleading when evaluating the real ability of a model, as they are referred to the single specific data set assignment assumed.