Estimation Of Average Annual Precipitation Of Khorasan Razavi Province Using Spatial Coordinates

Prediction of annual precipitation is of prime importance for water resources management and agricultural crop pattern planning. Different methods are used such as classical estimation method, geostatistical methods, multiple linear regression, artificial neural networks and Kriging combined with neural networks for prediction of precipitation at unsampled locations.In this study, non-climatic inputs, were used such as geographic coordinates and altitude, and average annual precipitation of 47 stations (corresponding to 25 years period of 1886 to 2011) at province of Khorasan Razavi were adopted. The highest correlation was found to be between average annual precipitation (as dependent variable) and three coordinate variables of altitude and geographical coordinates of the stations (as independent variables). Based on the results, however, IDW method was not sensitive to altitude. Effective range was 30.2 kilometers under SK and OK methods, while it was increased to 40 kilometers under ordinary cokriging which supports for hight dependency of precipitation to altitude in spatial context. The semivariogram parameters of sill, nugget effect, effective range and relative nugget effect were quite different for different years which is an indication of spatio-temporal pattern of precipitation at the study area. The best results were attributed to RK and SKV, however. The structure of the adopted neural network in this study was the multilayer perceptron (MLP) with sigmoid tangent and linear functions. Different structures of neural networks were tested with different inputs, the optimum network for prediction of average annual precipitation was attributed to three inputs (longitude, latitude and altitude) corresponding to 3-6-1 structure with Levenberg-Marquardt algorithm. The highest correlation coefficient and the lowest error were due to artificial neural network method, so it was the best method in this study.