Mosaic artifacts caused by erroneous input from L2A scene classification
The objective of the S2GM service is the provision of mosaic image products derived from the Sentinel-2A and -2B platforms. The data products shall serve as input to further thematic processing, generation of higher-level products, and analysis. The overarching requirement is the selection of the most representative spectrum for a given pixel, selected from the set of observations made during the temporal compositing period. This shall help to condense the overall information content; i.e. reduction of the data volume to a selection of most-appropriate observations within a given time period in order to lower the processing burden of users. Input to the mosaicking process are surface reflectance values (bottom-of-atmosphere or so-called Level 2A product). The Copernicus (ESA) ground segment operationally produces the level 2A (L2A) product. ESA uses the Sen2cor atmospheric correction processor (ESA Sen2Cor) for the generation of the L2A products.
The purpose of this blog entry is to described observed issues within the output products of the S2GM service. The issues described herein have been detected during routine quality control, for which a set of pre-defined Regions of Interests (ROI) are checked on a regular basis in order (1) to detect artefacts after system updates, and (2) to identify general product issues.
Figure 1 shows one of the ROIs used for validation purpose. It is a region in Norway covering four neighboring Sentinel-2 tiles (S-2 MGRS tile grid shown on the left-hand side) and a ten-day mosaic for the first decade of October 2019 as produced with the S2GM service (image on the right). The white areas in the ten-day mosaic are locations for which no cloud-free, cloud shadow-free, or snow-free observations could be found within the entire month (except for the image boarders).
Figure 1: Region of interest for the underlying analysis situated in Norway covering the Sentinel-2 MGRS grid tiles (grid over land areas shown in left image together with Google baselayer) 33WVN (upper left), 33WVM (lower left), 33WWN (upper right, detailed analysis), and 33WWM (lower right). The figure on the right-hand side shows a ten-day mosaic as produced with the S2GM service for the first decade of October 2019. The white areas in the resulting mosaic are locations with no valid observations (cloud and cloud-shadow free) within the entire month. Striking features in the mosaic are the strip-like patterns particularly obvious in the tiles 33WWN (upper right) and 33WWM (lower right).
A striking feature in the mosaic is the stripe-like patterns particularly obvious in the tiles 33WWN (upper right) and 33WWM (lower right). Since these stripes have been observed for the first time, we were interested in finding the source of these artefacts in the product. The S2GM service products provide the necessary details and information to trace the data origin of each individual pixel, therefore allowing for a detailed analysis to trace back the error source. In order to do so, a user needs to activate the “advanced settings” tab when selecting the desired spectral bands during the ordering process to get access to the selection of particular QA/QC and metadata files. One of the additional bands called “Validation source index” (VSI) is a raster image which provides the link between each data pixel and its original Sentinel-2A or -2B L2A input value. The VSI image contains the image index as integer numbers as listed under the Source_product_list tag of the metadata_*.json file (cf. file excerpt in Figure 2) provided with each data delivery. Figure 3 displays part of the ten-day mosaic (Sentinel-22 MGRS tile 33WWN) as introduced in Figure 1. The left part of the figure shows the mosaic of band 8A with the obvious striping pattern (here shown as black-striped pattern), the image on the right provides the respective VSI image presenting the source index to the original input images.
Figure 2: Excerpt of the metadata_*.json file showing the source product list which contains the file names of all Sentinel-2A and -2B input files considered for a particular mosaic linking to the validation source index raster image (cf. Figure 3, right).
Figure 3: Part of the mosaic (S2 MGRS tile 33WWN) shown in Figure 1 extracted for the first decade of October 2019. The left image shows band 8A of the resulting product, the figure on the right-hand side the “Validation source index” linking each pixel to its original Sentinel-2A or -2B L2A image (more details see text). Black-striped regions contain no-data pixels.
From the VSI raster image (Figure 3, right) it is obvious that large parts of the mosaic are composed of only two input images, namely a Sentinel-2B observation from 5 October 2019 and a Sentinel-2A observation from 7 October 2019. These images are shown in Figure 4 and Figure 5. Both figures display the original spectral band 8A and the Sen2Cor Scene Classification (SCL). The SCL is created with the Sen2Cor processor and is also used within the S2GM algorithm to select appropriate input pixels; i.e. information from the SCL is used for the pre-selection of appropriate image pixels. The SCLs for both images show that the input images contain a mixture of cloud-free, cloudy and cloud-shadow affected areas with the mosaicking algorithm selecting the best representative pixels from both input images.
Figure 4: Sentine-2B input image S2B_MSIL2A_20191005T104029_N0213_R008_T33WWN_20191005T131310 taken on 5 October 2019. The left image shows band 8A, the right image the SCL as produced by the Sen2Cor processor.
Figure 5: Sentine-2A input image S2A_MSIL2A_20191007T103021_N0213_R108_T33WWN_20191007T131042.SAFE taken on 7 October 2019. The left image shows band 8A, the right image the SCL as produced by the Sen2Cor processor.
However, zooming into areas which are affected by the striping pattern (cf. Figure 6) reveals some misclassifications of the Sen2Cor processor. Figure 6a shows the resulting mosaic (band 8A) and its corresponding VSI image (Figure 6b). Figure 6c and 6d are the Sentinel-2B input image and respective SCL for 5 October 2019, Figure 5e and 5f are the Sentinel-2A input image and respective SCL for 7 October 2019. Focusing on the artefacts shown in Figure 5a (black-striped patterns) and comparing them with the original input data, one can see that in both cases the SCL contains misclassified cloud shadows (brown color) in regions where obviously no clouds are apparent (left part of Figure 6c, lower right part of Figure 6e). Since pixels identified as “cloud shadow” (SCL) are not considered as valid observations in the mosaicking algorithm and since no other valid observations could be found for the first decade of October 2019 for the pixels which are part of the stripe-like pattern, the resulting ten-day mosaic contains these artefacts as no-data regions. Obviously, the Sen2Cor processor misclassified some of the cloud-shadow regions affecting the quality of the resulting mosaic products. Processing of additional mosaic products for the same region but for other time periods considering data from summer 2018 and 2019 did not reveal similar patterns. Therefore, we think that the observed issues should not be a systematic problem of the S2GM products, however, further checks will be done to verify this.
Figure 6: Zoom into tile 33WWN showing some of the observed artefacts (black-striped pattern). The ten-day mosaic of band 8A (a) and respective validation source index image (b) are shown in the left column. The column in the middle shows the original Sentinel-2B image taken on 5 October 2019 with the spectral band 8A (c) and the respective Sen2Cor Scene Classification (SCL)(d). The column on the right shows the original Sentinel-2A image taken on 7 October 2019 with the spectral band 8A (e) and the respective Sen2Cor SCL(e). The SCL legend is shown in Figures 4 and 5.
With this blog post we tried to provide an explanation for some of the artefacts which can be observed in the S2GM service products and which can be traced back to the L2A output of the Sen2cor processor.
In addition, we also demonstrated that beside the spectral bands provided through the service, the additional QA/QC layers and metadata provide useful information to trace back the original input imagery for each individual pixel allowing to get a deeper understanding how the products are composed. This kind of information can also be useful for the development of downstream products and services.