r.kappa lacks:
- User's and Producer's accuracies
- Output in CSV
- sort output differently:
- arrange user and producer accuracies and omission and commission around the error matrix instead of the summary columns
- add notes to "map1" and "map2" to explain what that means (more user friendly)
For a working solution, see https://github.com/OSGeo/grass-addons/blob/master/grass7/raster/r.confusionmatrix/r.confusionmatrix.py