implement "sample size" option in "Create unsupervised dataset (from feature raster)" algorithm

[Andreas Janz]

‌

[Andreas Janz]

Issue #1064 was marked as a duplicate of this issue.

[Agustin Lobo]

Any estimate of when this issue will be fixed?

[Andreas Janz]

It’s one of the next things I will do.

[Andreas Janz]

Note that the given sample size is only an approximate upper limit. No data pixel will further decrease sample size. I will update the description to reflect that.

[Agustin Lobo]

Yes, good idea to make the remark to the user. It would be nice getting the actual nb of samples really used in the result.

[Andreas Janz]

The actual sample size/shape [nsamples, nfeatures] should be printed in the log and is also available here:

‌

[Agustin Lobo]

When I do this in R, having set n samples, I generate 1.5*n random points and discard those with NA values. If the new number is <n, then I generate 2*n points and filter out again (this never occurs as I never have so many NA values). As soon as I get the filtered nb of points >=n,

I take the first n points.

But this could be unnecessary: it the user sets n and gets a much lower final number, he/she can just re-run with a higher n.

[Andreas Janz]

Yeah, I also thought about beeing a little bit more clever to get an exact number of samples. But haven’t done it so far.

For now, I use the exact same logic as QGIS, when it calculates BandStatistics with a given SampleSize:

‌

[Andreas Janz]

resolves #1062

→ <<cset f20a91e7e7cd>>