Evolution of S2GM’s short-term composite algorithm

  • Posted on: 26 March 2019
  • By: S2GM

Evolution of S2GM’s short-term composite algorithm

The current algorithm used for short-term compositing (STC) is a decision tree approach evaluating the Sentinel-2 Level 2A scene classification. As shown in the previous blog post, however, the scene classification may contain errors, which eventually result in undesired artefacts in the S2GM products. The S2GM therefore developed the STC algorithm further to make it independent from the Level 2A scene classification and thus more robust. 

 therefore a decision was made to make the STC algorithm independent from the Level 2A scene classification. 

Current approach (Version 1.0.9)

The STC approach has been motivated by the Web-enabled Landsat Data –WELD method [WELD ATBD, 2011] method and is, like WELD, based on a decision tree regarding surface reflectance values, scene classification, and several indices. The compositing approach has been designed to preferentially select valid land surface observations with minimum cloud, snow, and atmospheric contamination. Therefore, the composited mosaics are not suitable for studies of cloud, snow or the atmosphere. Different to the original WELD, the S2GM STC has not available the thermal bands of Landsat 8. It is instead adapted to the spectral characteristics, as well as the Scene Classification available in the Sentinel-2 L2A product.  

New approach (Version 1.1.0)

The new STC approach ignores the Sentinel-2 Scene Classification and is solely based on spectral tests, using multiple indices like normalized difference vegetation index (NDVI), modified normalized difference water index (MNDWI) and tasseled cap brightness (TC4). For two or three valid observations, the following decision logic is applied: 

(1) with the maximum NDVI if it has MNDWI < -0.55 and maxNDVI – meanNDVI < 0.05 else 

(2) with the maximum MNDWI if it has NDVI < 0.3 and meanMNDWI – minNDVI < 0.05 else  

(3) with the maximum NDVI if it has meanNDVI > 0.6 and meanTC4 < 0.45 else  

(4) with the minimum TC4 if the cloud test is false else 

(5) with the minimum TC4 if the snow test is false and minTC4 <1.0 else 

(6) with the maximum MNDWI if the meanNDVI < -0.2 else 

(7) with the minimum NDVI if the meanTC4 > 0.45 else 

(8) with the maximum NDVI 

If only a single valid observation is available, it is provided as output. For more than three observations, the Medoid algorithm is used. 

In most cases, the output of the new algorithm is very much comparable to the results achieved with the current implementation but has the advantage of delivering consistent results independent from the quality of the Sentinel-2 Level 2A scene classifications.  

Results for both STC algorithms are very similar for the great majority of cases. The evolution, however, still delivers consistent results for erroneous scene classification, particularly for cases with multiple valid observations that are contaminated by uncorrected haze. 

The following figures illustrate typical differences between the STC based on the SCL and the improved one based on spectral properties only. 

Figure 1 shows a comparison of a quarterly S2GM product over Ireland between July and September 2018. Due to high cloud coverage during this period the number of valid inputs pixels was low and thus many pixels have been processed by the STC. Both products look very much alike with, albeit with less remaining  haze removal in the new implementation. 

Figure 2 shows the same area but a monthly product for January 2018. Here, the new algorithm reduces some of the remaining cloud contaminations. 

Figure 3 shows a 10-days composite over central Sweden. Again, the products are very similar, but the new algorithm removes cloud contaminations at lake shorelines and river banks. 

Figure 4 shows a 10-days composite over southern Spain. These products reveal slight differences only over water areas. 

The new algorithm described here, together with the cloud/snow fix introduced in the previous post, will be released together with the Global Production of S2GM in early April 2019. So, keep checking our website for the official announcement.