Application of Artificial Neural Network in Estimating SPI and PDSI Drought indexes in Mashhad Synoptic Station

Drought, among the natural calamities, has the first rank in the view of severity, duration, and damage. To estimate the drought, it is necessary to use dynamic models in which the dominant processes on drought phenomenon are considered. Artificial neural networks models are dynamic models which are able to determine the relation of the inputs and outputs of a physical system joining together through nods. This research was performed to predict drought phenomenon in Mashhad synoptic station using perceptron and gradient descent with momentum training algorithm and gradient descent. The model input variables for SPI index consist of precipitation, climate large-scale index of SOI and NAO during 1951- 2007, and network output is SPI index in time delays; and input variables for PDSI index include precipitation, SOI, NAO and temperature during 1951- 2007, and the network output is PDSA index in different time delays. For modeling the SPI index, 57-year rainfall data of Mashhad synoptic station was used, of which 46 years were used for the network training and the rest (11 years) for the network testing. To model PDSI, 52-years statistical data was utilized; 33years for the network training and 9 years for validation and the rest 10 years for the network testing, by using Nero Solutions5 software. Also, for investigating the relation of ENSO with the precipitation of Mashhad synoptic station, regression method was used. Results showed that the southern fluctuations have small effect on this station precipitation and the northern atlas fluctuations have no effect on them. Occurrence of strong ENSO phenomenon in the world has affected the precipitations of Mashhad synoptic station, and the year after that, la Nina caused drought and low rainfall in this station. Occurrence of ENSO warm phase leads to increase in winter rainfall and lower rainfall in other seasons. Because of wide range of rainfall changes in time spans of 1, 2, and 3 months, the accuracy of predicted results by these time series is very low and by increasing of index time scale, the range of precipitation changes decreases while the results accuracy increases. The models of perceptron and feed forward operate well in predicting drought. Operation criteria R2= 0.78 for predicting 18-month SPI, and R2= 0.76 for predicting PSDI reflect this fact.