Evaluation of an AVHRR Cloud Detection and Classification Method over the Central Arctic Ocean

Abstract

A cloud classification method that uses both multispectral and textural features with a maximum likelihood discriminator is applied to full-resolution AVHRR (Advanced Very High Resolution Radiometer) data from 100 NOAA polar-orbiter overpasses tracked from an icebreaker during the 1994 Arctic Ocean Section. The cloud classification method is applied to the 32 ? 32 pixel cell centered about the ship?s position during each overpass. These overpasses have matching surface weather observations in the form of all-sky photographs or, during a period of heavy weather, an objective record that the sky was overcast with low water clouds. The cloud classifications from the maximum likelihood method are compared with the surface weather observations to determine if the automated satellite cloud classifier actually produces realistic descriptions of the scene. These comparisons are favorable in most cases, with the exception of a frequent error in which the classifier confuses Ci/Cc/Ac with extensive low water clouds over sea ice. This overall evaluation does not change appreciably if global area coverage resolution is used instead of full resolution or if the authors attempt to recalibrate the data to the NOAA-7 data for which the algorithm was originally developed. The authors find that the Ci/Cc/Ac cloud error can usually be avoided by 1) modifying the textural feature values for some cloud-over-ice categories and 2) applying a threshold value of 30% to the AVHRR channel 2 albedo averaged over the cell (and normalized by the cosine of the solar zenith angle). For a cell that the classifier identifies as containing Ci/Cc/Ac over sea ice, a cell-average channel 2 albedo greater than 30% usually indicates that the cell instead contains extensive low water clouds. When compared to the surface weather observations, the skill score of the satellite cloud classifier thus modified is 81%, which is very close to that claimed by its original author. This study suggests that satellite cloud detection and classification schemes based on both spectral signatures and texture recognition may indeed yield realistic results.