Improving cloud detection in S2GM

  • Posted on: 26 March 2019
  • By: S2GM

Improving cloud detection in S2GM

Remaining cloudy pixels are a common feature of S2GM products. This has also been pointed out recently by JRC’s Cloud Free S2 Platform, which provides Sentinel-2 L1C cloud-free composites from 2015 to 2018. They compared a L1C cloud-free composite over a region of Ireland in the time period from June to September 2018 with the corresponding S2GM product, which contained a severe contamination of remaining cloud pixels.  

The striking differences obviously raise the question why the L1C Cloud-Free project is doing a good cloud screening while the S2GM product struggles with it? 

Besides the used compositing algorithms (median vs. medoid) and the input products (L1C vs. L2A) the major difference is the different cloud screening applied to the two composites. While the L1C composite uses its own cloud screening approach, the S2GM product relies entirely on the cloud classification included in the official Sentinel-2 L2A products, which is generated by sen2cor. 

The S2GM service task is to generate best possible composites from the given L2A inputs and distribute them to users. Implementing an additional S2GM cloud screening is not sensible, read scientifically sound, because the atmospheric correction that has been applied to the L2A product has altered the original measurements.   

The mediocre results nevertheless motivated the S2GM team to investigate the root cause of this issue. Fortunately, the S2GM products come along with multiple auxiliary data, one of them being the scene classification (quality_scene_classification). This ancillary band provides the L2A scene classification for each pixel of the product. The figure below shows the S2GM quarterly product over Ireland with the corresponding scene classification.  

It clearly reveals that the remaining cloud pixels are classified as snow. As snow is a clear sky observation, it is kept in the S2GM processing and may eventually lead to the observed artifacts. 

Further analysis revealed that this misclassification is systematic and considerably impairing the quality of the S2GM composites. To get rid of these clouds classified as snow, a post processing step, which only affects pixels classified as snow, has been implemented by the S2GM team. To this end, the snow test of Sentinel-2 L1C IdePix has been used and adapt to L2A. The test is based on Normalized Difference Snow Index - NDSI and tasseled cap brightness (Kauth & Thomas, 1976). If a L2A snow pixel doesn’t fulfill this post-processing test, it is reclassified as cloud.  

This solution is not ideal, as it is operating on atmospherically corrected data, but is able to remarkably improve the product quality in many cases. As illustrated by Figure 3, most but not all remaining clouds have been removed in the example scene. The clouds still remaining after the snow test, are misclassified as “not vegetated”. This problem will be the topic of another investigation and hopefully fixed with another update of the S2GM Mosaic Hub. 

It is planned to release the cloud/snow fix with the next update in early April 2019. 



Kauth, R. J., & Thomas, G. S. (1976). The Tasselled-Cap—A Graphic Description of the Spectral-Temporal Development of Agricultural Crops as Seen by Landsat. Proceedings, Symposium on Machine Processing of Remotely Sensed Data, Purdue University, West Lafayette, IN, 29 June-1 July 1976, 41-51.