Variable selection and prediction of rainfall from WSR-88D radar using support vector regression

This research utilizes linear programming support vector regression to perform variable selection and rainfall estimation. Variables selected from applying linear programming support vector regression are used to perform rainfall prediction tasks using standard support vector regression and a Bayesian neural network. Ground truth rainfall data are necessary to apply intelligent systems techniques. A unique source of such data is the Oklahoma Mesonet. With the advent of a national network of advanced radars (i.e., WSR-88D), massive archived data sets are available for data mining. The reflectivity measurements from the radar are used as inputs for the learning techniques tested. The application of linear programming support vector regression for variable selection is new for the estimation of rainfall by radar. Results show that by selecting subsets of pertinent variables, standard support vector regression is more accurate in terms of generalization error than application of either traditional regression or a rain rate formula used in meteorology. Moreover, support vector regression shows a better prediction than standard linear programming support vector regression, traditional linear regression and Bayesian neural network for rainfall estimation.