Forecast of Precipitation Probability Based on a Pattern Recognition Algorithm

Abstract

A nonparametric statistical technique is developed which is capable of interfacing with dynamic atmospheric models to produce probability of precipitation forecasts. The technique is implicitly nonlinear, utilizing an efficient algorithm to represent joint probability densities as N-dimensional histograms. Operationally, the method can be updated on a daily basis. The technique is tested in an atmospheric data sample. Initial predictor selection is based on a general nonlinear approach applying fields of the sample correlation ratio. An iterative algorithm optimization technique is described and utilized. The technique is tested and shown to have significant skill relative to climatology.