i.landsat.acca - cloud/shadow detection (vs. fmask)

[martinl]

Recently I was working with Landsat data and performed cloud detection using G72:i.landsat.acca. The module produced quite reasonable results at least for clouds (see attachment:landsat_acca.png​, yellow color), but much more worst results for shadows (orange color, see -s flag (1)), also -2f flags have been used. Please compare with RGB composition (attachment:landsat_rgb.png​ and attachment:landsat_acca_comp.png​).

The module uses basically map algebra for cloud detection (2) and for shadows (3). Shadows detection seems to be somehow unfinished (or very simple).

I tested fmask (4) which gives much more better results (attachment:landsat_fmask.png​, attachment:landsat_fmask_comp.png​). I didn't studied the algorithm, but it seems to me that it's based on object segmentation approach.

Do you think that new better implementation of G72:i.landsat.acca would be topic for GSOC or student project? The fmask algorithm seems to be well-known, I also found Python library (5).

For now I would suggest to disable shadow detection or at least to add big fat warning about results. What do you think?

(1) https://grass.osgeo.org/grass72/manuals/i.landsat.acca.html

(2) ​https://trac.osgeo.org/grass/browser/grass/trunk/imagery/i.landsat.acca/algorithm.c#L76

(3) ​https://trac.osgeo.org/grass/browser/grass/trunk/imagery/i.landsat.acca/algorithm.c#L465

(4) ​https://github.com/prs021/fmask/blob/master/README.md

(5) ​http://pythonfmask.org/en/latest/