Evaluation of Visible Infrared Imaging Radiometer Suite (VIIRS) neural network cloud detection against current operational cloud masks

Abstract

Cloud properties are critical to our understanding of weather and climate variability, but their estimation from satellite imagers is a nontrivial task. In this work, we aim to improve cloud detection, which is the most fundamental cloud property. We use a neural network applied to Visible Infrared Imaging Radiometer Suite (VIIRS) measurements to determine whether an imager pixel is cloudy or cloud-free. The neural network is trained and evaluated using 4 years (2016–2019) of coincident measurements between VIIRS and the Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP). We successfully address the lack of sun glint in the collocation dataset with a simple semi-supervised learning approach. The results of the neural network are then compared with two operational cloud masks: the Continuity MODIS-VIIRS Cloud Mask (MVCM) and the NOAA Enterprise Cloud Mask (ECM). We find that the neural network outperforms both operational cloud masks in most conditions examined with a few exceptions. The largest improvements we observe occur during the night over snow- or ice-covered surfaces in the high latitudes. In our analysis, we show that this improvement is not solely due to differences in optical-depth-based definitions of a cloud between each mask. We also analyze the differences in true-positive rate between day–night and land–water scenes as a function of optical depth. Such differences are a contributor to spatial artifacts in cloud masking, and we find that the neural network is the most consistent in cloud detection with respect to optical depth across these conditions. A regional analysis over Greenland illustrates the impact of such differences and shows that they can result in mean cloud fractions with very different spatial and temporal characteristics.