AUTOMATED RECOGNITION OF OCEANIC CLOUD PATTERNS AND ITS APPLICATION TO REMOTE SENSING OF METEOROLOGICAL PARAMETERS (SATELLITE METEOROLOGY, CLIMATOLOGY, WEATHER ANALYSIS).

摘要

A scheme is presented for the automated classification of oceanic cloud patterns in twenty classes. A training set is defined by 2000 samples of size 128 x 128 km taken in February 1984 over the Western Atlantic. The method uses visible and infrared images from a geostationary satellite. Class discrimination is obtained from thirteen features representing height, albedo, shape and multi-layering characteristics. Features derived from the two-dimensional power spectrum of the visible images proved essential for the detection of directional patterns (streets, rolls) and open cells. A simple classification algorithm is developed based on the assumption of multivariate normal distributions of the features. From 1020 independent samples, the consensus among three expert nephanalysts is an overall accuracy of 79% with the machine answer at least second best 89% of the time. The cloud climatology in twenty classes for January and February 1984 are compared.; The physical characteristics of the classes labeled by machine are investigated from collocation of 2130 cloud patterns with ship observations. It is shown that realistic estimates of the probability of precipitation can be inferred from the cloud patterns. For several meteorological parameters, multiple linear regressions involving satellite features are used to lower the variance within a class. For example, the satellite retrieved cloud base temperature is shown to be strongly related to the surface air temperature (Ta) and dew point (TD). Single retrievals of Ta and TD have rms errors less than 3.5 K for half of the classes whereas the seasonal maps over the entire domain show rms errors of 1.45 K and 1.70 K, respectively. Cloud pattern identification also leads to estimates of wind speed and sea-air temperature and humidity difference, with rms errors on seasonal retrievals of 0.92 m/s, 1.27 K and 1.36 g/kg, respectively. Resulting rms errors on the sensible and latent heat fluxes are 26 W/m('2) and 73 W/m('2), respectively. Thus, a promising method, based on the information provided by cloud patterns, is proposed for the remote sensing of meteorological parameters in cloudy atmospheres.