2. Review of literature

This section follows the discussion of some of the significant contributions made by various researchers in the application of different ANN techniques for modeling ETo or pan evaporation (Ep). A radial basis function (RBF) neural network was developed in C language to estimate daily soil water evaporation [26]. The input layer of network consists of average relative air humidity, air temperature, wind speed (Ws) and soil water content. They compared the results of RBF networks with the multiple linear regression (MLR) techniques. A feed-forward back propagation (FFBP)-based ANN model was developed to estimate daily Ep based on measured weather variables [27]. They used different input combinations to model Ep. They compared the developed ANN models with the Priestly-Taylor & MLR models. RBF neural network model was developed to estimate the FAO Blaney-Criddle b factor [28]. The input layer to RBF model consists of minimum daily relative humidity (RHmin), day time Ws and mean ratio of actual to possible sunshine hours (n/N). The b values estimated by the RBF models were compared to the appropriate b values produced using regression equations. FFBP ANN models were implemented for the estimation of daily ETo using six basic climatic parameters as inputs [16]. They trained ANNs using three learning methods (with different learning rates and momentum coefficients), different number of processing elements in the hidden layers, and the number of hidden layers. The compared the results of developed ANN models with the Penman Monteith (PM) method and with a lysimeter measured ETo. ANN-based back propagation models for estimating Class A Ep with minimum climate data (four input combinations) were developed and compared with the existing conventional methods [22].

The potential of RBF neural network for estimating the rice daily crop ET using limited climatic data was demonstrated [23]. Six RBF networks, each using varied input combinations of climatic variables were trained and tested. The model estimates were compared with measured lysimeter ET. A sequentially adaptive RBF network was applied for forecasting of monthly ETo [29]. Sequential adaptation of parameters and structure was achieved using extended Kalman filter. Criterion for network growing was obtained from the Kalman filter's consistency test. Criteria for neuron/connections pruning were based on the statistical parameter significance test. The results showed that the developed network was learned to forecast ETo,t+1 (current or next month) based on ETo,t-11 (at a lag of 12 months) and ETo,t-23 (at a lag of 24 months) with high reliability. The study examined that whether it is possible to attain the reliable estimation of ETo only on the basis of the temperature data [24]. He developed RBF neural network for estimating ETo and compared the developed model with temperature-based empirical models.

ANN-based daily ETo models were trained based on different categories of conventional ETo estimation methods such as temperature based (FAO-24 Blaney-Criddle), radiation based (FAO-24 Radiation method for arid and Turc method for humid regions) and combinations of these (FAO-56 PM) [14]. A comparison of the Hargreaves and ANN methods for estimating monthly ETo only on the basis of temperature data was done [19]. They developed ANN models with the data from six stations and tested these developed models with the data from remaining six stations, which were not used in model development. The capability of ANN for converting Ep to ETo was studied using temperature data [18]. The conventional method that uses pan coefficient (Kp) as a factor to convert Ep to ETo was considered for this comparison. Generalized ANN (GANN)-based ETo models corresponding to FAO-56 PM, FAO-24 Radiation, Turc and FAO-24 Blaney-Criddle methods were developed [15]. These models were trained using the pooled data from four California Irrigation Management Information System (CIMIS) stations with FAO-56 PM computed values as targets. The developed GANN models were tested with different stations which were not used in training. Multilayer perceptron (MLP) neural networks for estimating the daily Ep using input data of maximum and minimum air temperature and the extraterrestrial radiation was developed [20]. The potential for the use of ANNs to estimate the ETo based on air temperature data was examined [21]. He also conducted comparison of estimates provided by the ANNs and by Hargreaves equation by using the FAO-56 PM model as reference model.
