[GRASS-dev] [GRASS GIS] #3469: i.atcorr: Sentinel-2 support broken on some systems

[GRASS GIS]

#3469: i.atcorr: Sentinel-2 support broken on some systems
--------------------------+-------------------------
  Reporter:  sbl          |      Owner:  grass-dev@…
      Type:  defect       |     Status:  new
  Priority:  normal       |  Milestone:  7.4.0
 Component:  Packaging    |    Version:  unspecified
Resolution:               |   Keywords:  i.atcorr
       CPU:  Unspecified  |   Platform:  Unspecified
--------------------------+-------------------------

Comment (by mmetz):

 Replying to [comment:7 mlennert]:
 > [...] This means that it is something systematic which we should be able
 to find. Stefan, you tried different versions. Are there any older
 versions which work for you ?
 >

 It seems that the 6s code in i.atcorr is numerically unstable: some
 internal variables can get very close to zero. If these close-to-zero
 values are snapped to zero, nan and inf results can appear. This is why
 the result can be all NULL, or not.

 The symptom can be cured by using double precision floating point
 variables throughout. Currently, i.atcorr uses single precision floating
 point variables. When using double precision floating point variables,
 Debian testing produces reasonable results. However, these results differ
 from Fedora 27, also using double precision floating point variables
 throughout. I would say that both results (Debian testing and Fedora 27)
 are slightly wrong because of the numerical instability of the 6s code in
 combination with the sentinel-2 parameters.

 I suggest to review the parameters for sentinel-2 as used by i.atcorr.

 I also suggest to review i.atcorr. The code has been translated from
 Fortran to C++ and might need some manual adjustments.

--
Ticket URL: <https://trac.osgeo.org/grass/ticket/3469#comment:8>
GRASS GIS <https://grass.osgeo.org>