Modelling of Evapotranspiration using Artificial Neural Networks

Artificial neural networks (ANNs) provide a quick and flexible means of creating models for many hydrological processes, and have performed well in comparison with the conventional methods in the past. The present study aims at utilisation of the input-output mapping capabilities of the ANN in daily grass reference crop evapotranspiration (ETo) estimation and other issues associated with the use of ANNs such as the effect of different normalisation schemes and length of training dataset on ETo prediction performance. The networks were trained with daily climatic data as input and the Penman-Monteith (PM) estimated ETo as target using the Cascade-Correlation (CC) and the Back-Propagation (BP) learning algorithms. The best ANN architecture was selected on the basis of weighted standard error of estimate (WSEE). The ANN architecture of 6-11-1 gave the minimum WSEE (0.29 mm d⁻¹) for the CC algorithm, whereas, 6-7-1 network, when trained with 5000 training cycles, gave the minimum WSEE (0.28 mm d⁻¹) for BP algorithm. Furthermore, when the effect of different normalisation schemes was evaluated, simple normalisation with linear scaling of data between 0.1 and 0.9 gave the minimum WSEE of 0.257 mm d⁻¹ for the BP algorithm and linear scaling between 0 and 1 gave minimum WSEE of 0.287 mm d⁻¹ for the CC algorithm. Also, the minimum length of training dataset for estimating ETo was found to be 5 years using both BP and CC.