Automatic cloud classification of whole sky images

Depending on their microphysical and macrophysical properties, clouds have different effects on the Earth's radiation and energy budget. To study these impacts on the climate system, satellite imagery as well as ground based image devices can be used. At the Leibniz Institute of Marine Sciences at the University of Kiel (IFM-GEOMAR), Germany, sky imagers have been developed to enable temporal and spatial high resolution observations of the cloudy atmosphere. Our automatic cloud classification algorithm, trained with images captured onboard the German research vessel "Polarstern" on the transect ANT XXIV/1 from Germany to South Africa in autumn 2007, will be presented in this work.

The algorithm is based on a set of statistical features describing the color as well as the texture of an image. For the classification process, the k-nearest-neighbour method is chosen due to its high performance in solving complex issues, simplicity of implementation and low computational complexity. Seven different sky conditions can be distinguished: high thin clouds, high patched cumuliform clouds, stratocumulus clouds, low cumuliform clouds, thick clouds, stratiform clouds and clear sky.

This algorithm is capable of processing the high volume of data produced by the cameras in realtime. It achieves a success rate of up to 97%, thereby outperforming previous algorithms. Only datasets including ambiguous images, e.g. images showing more than one class, or images not manually identifiable, result in a lower accuracy. Successfully classifying those images will be part of future research.