Automated Classification Of Oceanic Cloud Patterns With Applications To Cloud Type Dependent Retrievals Of Meteorological Parameters

Abstract

Oceanic cloud patterns are classified in twenty classes from visible and infrared imagery available from a geostationary satellite. A vector of features representing height, albedo, shape and multilayering characteristics of the cloud fields permits an objective classification. An original aspect of the scheme is its capability to recognize direc-tional patterns such as cloud 'rolls' or 'streets', and doughnut-shaped open cells as well, from features derived from the power spectrum of the visible image. The classifier was trained using 2000 samples of size 128 x 128km extracted from February 1984 images over the Northwestern Atlantic. Expert nephanalysts suggest strict accuracy in 79% of the cases while the machine gives at least the second best choice among twenty classes 89% of the time. The McIDAS system is used to process the imagery. The grid of analysis is super-imposed on the satellite image and as the program runs, the class number appears in the middle of each box at the rate of one every 2.5 seconds while all the information retrieved is stored in a file. Applications of the scheme are suggested for meteorological para-meters such as the probability of precipitation and the surface air and dew point temperature.