Automatic Cloud Tracking Applied to GOES and METEOSAT Observations

Abstract

Improvements to the SRI automatic cloud-tracking system are described that enable it to operate on multilayer clouds associated with severe storms. The improved method has been tested using rapid-scan observations of Hurricane Eloise obtained by the GOES satellite on 22 September 1975. We performed cloud tracking using target selection (clustering) based either on visible or infrared data and tracked the targets using a pattern recognition technique. The technique matches targets with their best likeness (in terms of size, brightness and shape) at successive times in a manner analogous to human pattern recognition, and also rejects vectors in disagreement with the predominant motion in their height (infrared) category. For data of 4 and 8 km resolution, the automatic system gave results very comparable in accuracy and coverage to those obtained by NASA analysts using the Atmospheric and Oceanographic Information Processing System (AOIPS). We also tracked the same targets using a cross-correlation technique, but without internal editing. For targets that were tracked properly by both methods, the rms differences in displacements were only fractions of a pixel. To learn whether the automatic system can track the motions of water vapor patterns we applied it to METEOSAT 6.7 μm water vapor measurements. For experimentation, typical data for the midlatitudes, subtropics and tropics were chosen from a sequence of METEOSAT pictures for 25 April 1978. In flat (low contrast) water vapor fields, the automatic motion computations were not reliable, but in areas where the water vapor fields contained small-scale structure (i.e., in the vicinity of active weather phenomena) the computations were successful. The tracking results appear to be similar to those obtained by visual analysis of these same data. For the same cases (including tropical convective systems and midlatitude jet stream cirrus) we computed cloud motions using METEOSAT infrared observations and obtained excellent results. The cloud-motion vectors computed automatically appear to be competitive in accuracy and coverage with motions determined by human analysts working within reasonable time limits.