JWSS - Journal of Water and Soil Science

---

Estimating spatial distribution of precipitation is vital to execute water resources plans, drought, land-use plans environment, watershed management, and agricultural master plans. High variation in amount of precipitation in various parts, lack of measurement stations, and the complexity of relationship between precipitation and parameters affecting it have doubled the importance of developing efficient methods in estimating spatial distribution of precipitation. Artificial neural network has been proved to be efficient as a new way for modeling and predicting the processes for which no solution and explicit relationship has been available in accurately identifying and describing them. The purpose of this study is to investigate the efficiency of artificial neural network in estimating spatial monthly precipitation. To achieve this objective, neural network with multilayer perceptorn topology was employed for preparing model for spatial monthly precipitation in five synoptic and rain-gauge stations located in Kurdistan province. In order to design the topology of the model in each station, as the adjustable parameters (including transfer function, learning rule, amount of momentum, number of hidden layers, number of neurons of the hidden layers, and the number of epochs) changed, different neural networks were made and carried out. In each case, the topology with the minimum amount of root mean square error (RMSE) was selected as the optimal model. Owing to the fact that the selection of each of the variable parameters of neural network necessitated recurring trails and errors, and consequently teaching a large number of networks with various topologies, genetic algorithm method was utilized for finding the optimization of these parameters the efficiency of this method, too, was examined in terms of the optimization of neural network. The findings indicated that neural network enjoys a high degree of accuracy in modeling and estimating spatial distribution of monthly precipitation. In addition, combining it with genetic algorithm method was positively evaluated in optimizing the requirements for executing neural network. In most cases, mixed method proved its superiority over executing neural network without optimization. The most precise model in all of the stations under study was achieved by the use of transfer function, sigmoid, learning rule of Levenberg Marquardt in the selected models, the determination coefficient (R2) observed between the model output amounts and the data observed in station were found to be 0.86 0.89 0.94 0.77 and 0.94.