A Bayesian cloud mask for sea surface temperature retrieval

Bayesian methods are used to develop a cloud mask classification algorithm for use in an operational sea surface temperature (SST) retrieval processin g system for Advanced Very High Resolution Radiometer (AVHRR) area coverage (LAC) resolution data. Both radiative and spatial features are incorporate d in the resulting discriminant functions, which are determined from a larg e training sample of cloudy and clear observations. This approach obviates the need to specify the arbitrary thresholds used by hierarchical cloud-cle aring methods, provides an estimate of the probability that an instantaneou s field of view is cloudy (clear), and allows the skill of different cloud discriminant models to be objectively analyzed. Results show that spatial information is of particular importance in reduci ng the false alarm rate of the cloudy class. However, while the use of comp lex textural measures such as gray-level difference statistics-as opposed t o simple statistics such as the standard deviation-improves the skill of ni ghttime cloud-masking algorithms, they are of little advantage during dayti me hours. Cloud mask discriminant models having similar high Kuipers' performance ind ex scores (i.e., 0.935) are developed for both day and night satellite data from the Southern Hemisphere midlatitudes. Applied to LAC orbital (i.e., o perational) data, the characteristics of the cloud masks appear to be simil ar to those derived from analysis of the training sample data. However, in this case. to enhance processing performance, a hybrid algorithm is employe d-obviously cloudy instantaneous fields of view (IFOVs) are first removed v ia a gross Ihreshold check and the Bayesian method applied only to the rema ining IFOVs. This same (hybrid) algorithm is also applied to an ensemble of 30 days of AVHRR LAC data from the New Zealand region. Analysis of the res ulting time-composited SST data (means and standard deviations) shows there is little evidence of a day-night bias in the performance of the Bayesian cloud-masking algorithm and that the resulting SST data may be used to dete rmine the variability of oceanographic features. Although this paper uses AVHRR data to demonstrate the principles of the Ba yesian cloud-masking algorithm, there is no reason why the approach could n ot be used with other instruments.