Validation of the LaSRC Cloud Detection Algorithm for Landsat 8 Images

This study aims at validating the cloud mask produced by the land surface reflectance code (LaSRC) for Landsat 8 data. To detect clouds in optical satellite imagery, LaSRC uses quality assurance (QA) layers, which are produced during the atmospheric correction process. The QA layers include a “cloud mask,” which is based on the estimation of a residual metric showing the quality of aerosol inversion, and “high aerosol,” which shows the impact of aerosols on the derived surface reflectance. Validation is performed using the “L8 Biome” cloud validation dataset, which is produced by the US Geological Survey, and consists of 96 Landsat 8 scenes distributed globally over 12 different biomes. We show that the LaSRC cloud detection algorithm reliably identifies thick clouds with commission and omission errors less than 4%. Large cloud overdetection errors occur for thin clouds, which is due to the subjectivity of defining and extracting thin clouds in the reference dataset. We conclude this paper with recommendations on using the LaSRC QA layers, and give suggestions on reducing subjectivity, when generating cloud validation datasets.