Seasonal and annual regional drought prediction by using data-mining approach

This study examines the seasonal regional drought analysis based on the standardized precipitation index (SPI) method and the decision tree technique which is a data-mining approach. The cumulative rainfall series for five reference periods (four seasonal and one annual series) were constituted by using monthly rainfalls from 17 stations in Cekerek Watershed, Turkey, which has an area of 1165 440 ha. Regional analysis was performed by forming the stations initially as homogeneous group(s) according to the discordancy criteria considering by l-moment ratios. There was no discordant station according to discordancy measure of site characteristics except for the first reference period. The heterogeneity measures showed that the selected groups were homogeneous. Based on the goodness of fit criteria {norm of matrix}Z DIST{norm of matrix} the candidate regional distributions having the minimum Z DIST for k-reference periods were the Generalized Pareto (GPA), Generalized Extreme Values (GEV), Generalized Logistic (GLO), Pearson Type III (PE3), GEV and 3-parameter Log Normal (LN3), respectively. The drought categories for each region were predicted by applying the decision tree rules obtained from the training phase of the k-reference periods. The results revealed that there was no significant difference between drought categories calculated from the conventional SPI algorithm and decision tree approaches. Moreover, the accuracy of prediction for k-reference periods was greater than 94%, except for k3 (81.2) and k5 (86.4%) reference periods