i.spec.unmix: endmember pixels do not sum up to 100%

[neteler]

While the endmember maps of i.pysptools.unmix sum up to 100%, those of i.spec.unmix apparently don't.

GRASS 7.7.dev (nc_spm_08_grass7):~ >

g.region raster=lsat7_2002_10 -p

# requires https://github.com/OSGeo/grass/pull/39
g.extension i.pysptools.unmix
g.extension i.spec.unmix

# auto-generate endmember matrix and endmember maps
i.pysptools.unmix input=lsat_2002 endmembers=endmembers endmember_n=5 \
   output=spectrum.txt prefix=lsat_spectra --v

# check data range (here: not reflectance), e.g.
r.info -r lsat7_2002_10
min=42
max=255

r.info -r lsat7_2002_40
min=1
max=255

# check created spectrum.txt
cat spectrum.txt 
# Channels: lsat7_2002_10 lsat7_2002_20 lsat7_2002_30 lsat7_2002_40 lsat7_2002_50 lsat7_2002_70 lsat7_2002_80
# Wrote 5 spectra line wise.
#
Matrix: 5 by 7
row0: 68.0	44.0	30.0	2.0	1.0	4.0	13.0
row1: 73.0	64.0	37.0	182.0	92.0	34.0	117.0
row2: 133.0	123.0	140.0	79.0	250.0	255.0	161.0
row3: 255.0	239.0	255.0	141.0	156.0	98.0	250.0
row4: 255.0	255.0	255.0	221.0	255.0	214.0	22.0

# derive endmember maps from spectrum.txt
# hypothesis: endmember maps should be similar to those generated by i.pysptools.unmix
i.spec.unmix group=lsat_2002 matrix=spectrum.txt result=lsat7_2002_unmix \
   error=lsat7_2002_unmix_err iter=lsat7_2002_unmix_iterations

# hypothesis: sum of all endmember maps should sum up to 100 (%):
r.mapcalc "all_unmix = lsat7_2002_unmix.1 + lsat7_2002_unmix.2 + lsat7_2002_unmix.3 + lsat7_2002_unmix.4 + lsat7_2002_unmix.5 - lsat7_2002_unmix_err"

r.univar -e -g all_unmix
...
min=-100
max=71
range=171
mean=21.3971956456606
...
first_quartile=18
median=22
third_quartile=25
percentile_90=28

... indeed, the sum is far from the expectation of SUM(endmembers) = 100%.

The addon i.spec.unmix is relevant because it expects known endmembers as input which is important for a defined analysis (while i.pysptools.unmix apparently always auto-generates the endmembers which are not necessarily the best or those the user wants).

I suspect a bug in i.spec.unmix (maybe just a scaling issue internally?).