Rain Estimation from Infrared and Visible GOES Satellite Data

Abstract

An automated statistical pattern recognition technique is presented that uses visible and IR satellite imagery to estimate instantaneous surface rainfall rates. The technique uses both brightness and textural statistics to estimate rainfall in 10 ? 10 pixel arrays of satellite data. Each array is centered over one of 137 Service A weather stations scattered over southeastern United States. Surface reports from these stations obtained during a 30 day period in August of 1979 are used to ground truth the technique. The technique classifies each 10 ? 10 array into one of three categories: no rain, light rain, moderate/heavy rain. Cross-validation is used to estimate classification errors; results of these estimates yielded an overall error rate of 35% when both visible and IR data are used. When only visible or IR data are used the overall error rates are 39% and 42%, respectively. In addition to the three class problem, the two class problem of classifying rain / no rain is studied. Overall error rates of 18% are achieved using a technique with 16 image statistics and both visible and IR data. A simpler technique that uses only the mean and standard deviation statistics, derived from the visible and IR data, achieved an overall error rate of 20%. We conclude that the visible and IR pattern recognition technique could be used successfully to estimate instantaneous rainfall in three classes: no rain, light rain, moderate/heavy rain. During the night and during hours of low sun attitude, IR data could be used but with a slight decrease in accuracy. We also conclude that a simpler pattern recognition technique, based upon the mean and standard deviation statistics, could be used to distinguish between rain and no rain classes.