Exploring Pattern Recognition Enhancements to ACSPO Clear-Sky Mask for VIIRS: Potential and Limitations

摘要

Discriminating clear-ocean from cloud in the thermal IR imagery is challenging, especially at night. Thresholds in automated cloud detection algorithms are often set conservatively leading to underestimation of the Sea Surface Temperature (SST) domain. Yet an expert user can visually distinguish the cloud patterns from SST. In this study, available pattern recognition methodologies are discussed and an automated algorithm formulated. Analyses are performed with the SSTs retrieved from the VIIRS sensor onboard S-NPP using the NOAA ACSPO system. Based on the analyses of global data, we have identified low-level spectral and spatial features potentially useful for discriminating cloud from clear-ocean. The algorithm attempts to mimic the visual perception by a human operator such as gradient information, spatial connectivity, and high/low frequency discrimination. It first identifies contiguous areas with similar features, and then makes decision based on the statistics of the whole region, rather then on a per pixel basis. Our initial objective was to automatically identify clear sky regions misclassified by ACSPO as cloud, and improve coverage of dynamic areas of the ocean and coastal zones.