I am facing the problem that in certain Sentinel 2 L2A (also with L1C, I checked) scenes, there is a generally high amount of pixels that show the value 1 (minimum). This occurs mainly over forest areas or large water bodies - of course these are areas of low reflection, but interpreting e.g. NDVI images derived from these scenes, this “problem” seems to be following a certain pattern (see last image below).
The images showing the Spectral Plot using ENVI also illustrates my concern that in my opinion the amount is unusual in the second picture.
Furthermore, scenes acquired at a different time do not show this issue (first picture of spectral plot). My observations are with regards to scenes in southern Greece, as well as in Burkina Faso.
Scenes here:
S2A_MSIL2A_20170824T091021_N0205_R050_T34SFF_20170824T091748 (“problematic one”)
S2A_MSIL2A_20170725T090601_N0205_R050_T34SFF_20170725T091952
Besides the Vegetation Indices with values being too high values (third picture), it became problematic for the classification of these scenes as now completely different land cover classes share similar values.
Thank you for providing ideas - help - similar issues!
it is correct, at least for the next time.
It occurs, that Sen2Cor overestimates the atmospheric influence which results in negative reflectance values. Negative values are not allowed in JPEG2000 format and therefore these pixels are clipped with value 1. We don’t use DN values of 0 because S2 Toolbox shows NaN for these.
However, the amount of pixels providing negative reflectance should not be larger than 1%. What you can do at present time to get more useful information from those pixels is to reprocess the granule without visibility update setting the visibility in the configuration to a higher value than the final visibility, which you find in the processing report file. Note, that this also changes the reflectance for all other pixels.
It is planned to allow negative reflectance values in future to keep the information for the shape of the spectra. This will be done by introducing an offset to the DN and avoid a large number of pixels with the same reflectance value. However, I cannot say you when this will be realized.
