Modeling VIIRS Brightness Temperature for Improving Nighttime Cloud Detection

Abstract:

Nighttime observations acquired by the Visible Infrared Imaging Radiometer Suite (VIIRS) form an important repository for monitoring Earth system components and derived higher-level products from these datasets find use in various inferences and analyses. These observations are affected by cloud contamination and accurate cloud detection is a vital step for ensuring the use of high-quality data in several VIIRS-derived products. However, cloud detection at night remains a challenge due to the presence of fewer active and cloud-sensitive VIIRS channels, leading to increased missed detections. The VIIRS Day/Night Band (DNB), which is an input dataset for a range of study areas such as aurorae, city lights, fires, aerosols, opaque nocturnal cloud optical depth, particulate matter, is very sensitive to cloud cover and missed detections greatly affect DNB and derived product quality.

We explore the potential of machine learning to extract cloud signature from VIIRS Brightness Temperature (BT) to improve nighttime cloud detection and flag contaminated DNB pixels to inform the DNB data analysis pipeline and derived products. We form regional BT models to extract cloud signature and utilize this model for detecting the presence or absence of cloud in VIIRS nighttime observations. This is achieved by learning a model that transforms the VIIRS BT bands of a region into a lower dimensional representation. Cloud free pixels with higher BT are more similar and lie closer in this lower dimension than cloudy pixels with lower BT and distance measures are used to determine the dissimilarity of cloud free and cloudy areas. Subsequent observations from the region are projected using the regional model. The dissimilarity measures in the latent space are then clustered to determine pixel-level probabilities of cloud cover in the region and multiclass labels are assigned to each BT cluster. Unseen test set observations from the same region are transformed to the lower dimension and pixels are assigned the label of the closest or most similar cluster, which is shown to improve cloud detection performance compared to the VIIRS cloud mask. We also extend this technique across seasons and varied geographic regions to study the generalizability of this approach.