Automated Recognition of Oceanic Cloud Patterns. Part I: Methodology and Application to Cloud Climatology

Abstract

A scheme is presented for the automated classification of oceanic cloud patterns. The 20 cloud classes reflect the rich variety of morphologies that are detectable from space. A training set is defined by 2000 samples of size 128 ? 128 km taken from GOES visible and infrared images over the western Atlantic in February 1984. Class discrimination is obtained from 13 features representing height, albedo, shape and multilayering characteristics of the cloud fields. Two features derived from the two-dimensional power spectrum of the visible images proved essential for the detection of directional patterns (cloud ?streets or rolls) and open cells. Based on the assumption of multinormal distributions of the features, a simple classification algorithm is developed. The generation of artificial samples yields a theoretical separability of 97% while the actual separability obtained on the training set is 95%. From 1020 independent samples, the separate verification of three expert nephanalysts indicates strict accuracy in 79% of the cases while there is agreement with their first or second choice in 89% of the cases. The cloud climatology is compared in 20 classes for January and February 1984. In agreement with available climatology, multilayered cloud fields are observed 42% of the time. The cloud fraction maps are also compared with the observed fields from ships.