Intercomparison of VIIRS Neural Network Cloud Detection and Current Operational Methods

Abstract:

Clouds are a critical component of the earth’s weather and climate system, and in conjunction with aerosols, are one of the largest sources of uncertainty in future climate scenarios. Accurately determining the presence of clouds in observational records is a fundamental step in understanding their variability and impact. In this work, we develop a neural network cloud detection algorithm for the Visible Infrared Imaging Radiometer Suite (VIIRS) and evaluate its performance against two operational cloud masks: the NOAA Enterprise Cloud Mask (ECM), and the Continuity MODIS-VIIRS Cloud Mask (MVCM). The neural network utilizes all 16 moderate-resolution channels from VIIRS, takes into account local spatial variability, and does not use any ancillary datasets. Using a globally distributed set of collocations with CALIOP spanning 2016-2019, we find significant improvement in cloud detection performance under a wide range of scenarios by using the neural network. In particular, we report increases in class-balanced accuracy of roughly 16% in nighttime scenes in the arctic (87% compared to 71%), 10% over nighttime land (92% compared to 82%), and 5% over daytime land (97% compared to 92%). We demonstrate that these differences in performance are not solely due to differing optical-depth based definitions of clouds. Additionally, we analyze cloud detection inconsistency across different solar illumination and surface conditions, which we argue is a contributor to spatial and temporal artifacts in cloud fraction analyses from satellite imagers. We find that all masks struggle with consistency in detecting optically thin clouds (optical depths less than 0.1) under varying conditions, but of the three masks, the neural network is likely the most reliable for characterizing spatial and temporal variability of cloud fraction in most regions.