Open nilsleh opened 2 years ago
Great questions!
I think both of these can be done at the same time
An alternative could be to plot a false color image from a user defined subset of the bands in the image (basically a user defines indices like [0, 1, 2]) as an argument to plot
. This is how it's done in torchgeo.datasets.IDTReeS
when plotting the hyperspectral image.
Another issue I noticed is that So2Sat
doesn't allow you to select bands but So2SatDataModule
does. The issue with this is that the So2Sat.plot
method assumes that the sample will always have all bands. We should copy the So2SatDataModule
band selection to So2Sat
and update So2Sat.plot
to handle this.
Relating to band selection: as more datasets arise in torchgeo, could band selection become standardized to the STAC specification common_names? This could also become a standard way of doing band selection from STAC items when some datasets point to STAC items.
FYI, I'm currently trying this "band selection interface" out with a dataset implemented torchgeo's way. I think it has some potential.
Related to, and partially addressed by #687.
As I mentioned in #394, I'm starting to think that it doesn't make much sense to allow band selection for curated benchmark datasets, only for uncurated raster datasets. What do you think?
Why not? E.g. experiments that compare how multispectral information compare to only rgb information on benchmark datasets are quite interesting, I think.
Hmm, that's fair, someone might want to do a more in-depth ablation study on which bands actually matter. I guess I'm not opposed if someone wants it, but I wouldn't go changing everything if we don't need it, especially if it doesn't speed up data loading times.
I think I want to rename the So2Sat bands anyway, will submit a PR to do that.
I am currently working on adding plotting methods based on #253. For datasets that contain Sentinel multispectral imagery, I have two questions.