The knowledge of the placement and size of clouds in the atmosphere has many applications in Atmospheric Science. In particular, clouds play a major role in the earth's radiation budget. Given the large amounts of image data and the relative short time this information is relevant requires an automated detection algorithm that runs in real-time on current workstations. We present such an algorithm in this paper.

We show that given daytime images, human analysts agree amongst themselves at most 94% on the placement of clouds within the image. The subjectivity of the problem motivates our approach. The algorithm uses a combination of neural networks to compute the probability of a pixel being clear or cloud. Images are first so\rted into neighborhoods via a Self-Organizing map based on the image's gray scale levels. After sorting, each pixel is processed using a Mixture of Experts neural network uniqu\e to the neighborhood. The calculation is of the probability that a pixel represents clear or cloudy. A pixel level cloud mask is then constructed by assigning the most likely designation to each pixel.

The algorithm reliably identifies a variety of clouds from images. Relative agreements with human analysts are typically greater than 90%. In addition, the pixel level cloud masks show high visual correlations with unprocessed images as well as very good temporal correlations.