ANN-RBF Hybrid Model for Spatiotemporal Estimation of Monthly Precipitation Case Study: Ardabil Plain

1. Introduction

Precipitation can be considered as the most important component of the hydrological cycle. The value of precipitation on the earth ground is subjected to vast spatiotemporal changes. Striking changes in spatiotemporal precipitation on one hand and lack of usual rain gauge stations to record the rainfall depths on the other hand make it inevitable to develop promising estimation models.

Inherent stochastic property, nonlinear behavior, essential need to chronological data and complexity of distributed physical models have urged researchers to utilize nonlinear black box models of time series such as Artificial Neural Networks (ANNs) (Nourani et al., 2009). According to published article by ASCE (2000), ANN could find several successful applications to different fields of hydrology, especially for precipitation modeling. For instance French et al. (1992) used ANN to estimate spatiotemporal precipitation and showed the ability of ANN against other classic methods. Vankasetan et al. (1997) predicted the monsoon rainfalls in India by using back propagation neural networks and compared the results with results obtained from other statistical methods. Luk et al. (2001) modeled rainfall in the Parramatta River basin in the Sydney by using feed forward ANN considering several lags in inputs. Ramirez et al. (2005) investigated feed forward ANN and reactionary dissemination learning algorithm. The main purpose was to model precipitation over six regions of Sao Paulo state performed for a period of 6 years. Coulibaly and Evora (2007) and Nourani et al. (2012) considered the usage of ANN to complete missing data in rain gauge stations. Chattopadhay and Chattopadhay (2008) in a study predicted the seasonal rains in India by neural networks. Dahamsheh and Aksoy (2009) considered the forecasting of monthly precipitation in arid regions using ANN and Multiple Linear Regression (MLR). The obtained results represented the vast ability of ANN in the forecasting of monthly rainfall. Nourani et al. (2009) used a hybrid Wavelet-ANN model to forecast the monthly precipitation of the Lighvanchibasin in Iran. In this research, a hybrid model was developed between the wavelet analyze and neural network to predict the precipitations. Despite numerous applications of ANN in prediction of various temporal quantities such as monthly precipitation, this tool is not noticed for spatial predictions and usually classic methods of hydrologic engineering are used to extend the value of monthly rainfall in a station to others in its neighborhood. But to predict spatial quantities, some classic prediction methods such as Arithmetic Average Method, Inverse Distance Method (IDM) and Inverse Distance Squared Method (IDSM) or geostatistics estimators can be employed (Nourani et al., 2011). Table 1 shows a brief overlook of performed studies of other researchers.

¹Correlation Coefficient, ²Determination Coefficient, ³Multiple Linear Regression, ⁴Feed-Forward Back Propagation