Evaluation of Six Neural Networks and Two Geostatistical Methods for Generating the Missing Precipitation Data

Improving the accuracy of the missing precipitation data, particularly in large watershed with low-density precipitation network, is one of the challenges of the hydrologists. This study investigated six different types of artificial neural networks, namely: the MLP, the TLFN, the RBF, the RNN, the TDRNN, the CFNN along with different optimization methods, and geostatistical methods namely the Kriging and the Cokriging models for infilling the missing daily precipitation. Daily precipitation records from 15 rain gaging stations located within the Karkheh Watershed in the southwest Iran, were used to evaluate the accuracy of different models for infilling data gaps of daily precipitation. The results suggest that the MLP, the TLFN the CFNN and the Cokriging can provide more accurate estimates of the missing precipitation values than the other ones. However, the MLP overall appears to be the most effective method for infilling the missing daily precipitation values. Moreover, the results show that the dynamically driven networks (RNN and TDRNN) are less suitable for infilling the daily precipitation records whereas the RBF are appeared to be fairly suitable. Also, the kriging model is less effective than the MLP, the Cokriging, the TLFN and the CFNN models, but shows better results than the RNN, TDRNN and RBF networks.