Evolution of SLSTR Bayesian cloud detection in coastal zones
Updates to extend the Bayesian cloud-detection capability for Sentinel-3 SLSTR for coastal zones have now been developed.
24 March 2022
23 March 2022
By Claire Bulgin and Chris Merchant (University of Reading) and Gary Corlett (EUMETSAT)
Compared to open oceans, coastal zones are characterised by more turbid water and greater temperature contrasts from upwelling, river outflows, etc. These differences between coastal waters and open oceans suggest adaptations to the cloud detection algorithm, compared to open-ocean configurations, could be beneficial. Cloud detection is an essential step prior to many remote sensing tasks, including surface temperature estimation. This project has developed updates to extend the Bayesian cloud-detection capability for SLSTR for coastal zones.
A key development is a pre-processor that improves the mapping of higher-resolution reflectance channel observations to the infrared observations. Figure 1 shows a comparison between reflectance from the operational scheme (a) and the new pre-processor (b), compared to the infrared image (c). The representation of the coastline in the reflectance channel observations now more closely matches the infrared view, which is beneficial to using all the SLSTR channels together to detect clouds.
This new pre-processor also enables calculation of the standard deviation of the S3 reflectance channel observations mapped to an infrared pixel (SSD3), which we have shown is a useful additional metric in the daytime cloud detection.
The cloud detection challenges addressed by this project are as follows:
Co-registration of infrared and reflectance channel observations.
‘Blockiness’ in cloud features along coastlines.
Identification of scattered cloud.
Over-flagging of optically bright waters.
Sun glint regions.
Over-flagging of thermal fronts.
Over-flagging of cold coastal upwelling waters.
Below are two examples of improvements in these areas, comparing the current operational configuration with a prototype of the updated one.
The first example is of optically bright waters. Figure 2 shows the Bohai Sea, China, where the coastal waters are bright in the S2 reflectance channel (top left). The operational cloud detection screens these optically bright waters as cloud (top right). The yellow colours on this plot denote land and/or cloud, and the black colour clear-sky over water. The blue colours identify places where the Bayesian cloud detection returns intermediate probabilities of cloud (usually also screened).
The SSD3 metric described above is a beneficial alternative to the S2 reflectance channel. SSD3 is shown in Figure 2 (bottom left) and shows little sensitivity to most of the optically bright waters, except over the most turbid regions. The resulting cloud detection from the updated configuration (bottom right) shows less false flagging of bright waters.
The second example shows reduced false flagging of thermal fronts around the South Korean coastline. The reason for the extensive false flagging of these features is the open-ocean thermal texture measure, which is sensitive to the strong thermal gradients across these features. Figure 3 shows the difference between the infrared and SSD3 sensitivity to ocean fronts, which drives the improvement.