Add support for GCPS

[vincentsarago]

Some GeoTIFF don't use Affine transform for geospatial coordinates but use a set of GroundControlPoint to do the transformation between the pixel space to the spatial space.

ref:

• https://gdal.org/en/stable/user/raster_data_model.html#gcps
• https://gdal.org/en/stable/development/rfc/rfc4_geolocate.html#changes-to-warp-api-and-gdalwarp

Dataset: the Sentinel-1 dataset on AWS is a good example

ref:

• Add image with GCPS or RPCS geotiff-test-data#21
• https://github.com/cogeotiff/rio-tiler/blob/main/tests/fixtures/cog_gcps_ovr.tif

[kylebarron]

Can you elaborate on what support you want?

• Expose GCP metadata? So it exposes something like a sequence of mapped points between spatial point and pixel point?
• Right now when you call read it returns a RasterArray, which contains a transform. I assume this would have to change? A user could want an affine fit to the locally-read portion of the image, but that wouldn't be necessarily desired out of the box.
• I think fitting the GCPs onto a transform is the part that's non-trivial. We have to do some sort of polynomial or spline interpolation. Perhaps our first task should be to just expose GCPs.
• I'm curious with the typing whether it would be better to split multiple classes of GeoTIFF objects? Like GeoTIFF will always have a non-null transform while GCPGeoTIFF could be a separate class without a transform attribute, but which exposes gcps as an attribute

[kylebarron]

Assuming we do want to add support for GCP-based GeoTIFF files, I think we'll need a bit of a refactor here. In particular, I don't want to change transform: Affine to transform: Affine | None. That is, 99% of the time users will have Affine-based GeoTIFFs, and I don't want type checkers to alert every time that "transform could be None".

So I'm considering:

• Changing GeoTIFF.open to a top-level async_geotiff.open
• async_geotiff.open can then return a number of concrete classes depending on the type of GeoTIFF it finds
• So the user will only have to do one isinstance check, and then all accesses of transform will be validated by the type checker

This means we'd have to switch to multiple physical classes: AffineGeoTIFF, GCPGeoTIFF, RCPGeoTIFF

Thoughts?

[kylebarron]

A good idea from @geospatial-jeff: we can make GeoTIFF just mean a geotiff with an affine georeferencing. This is non-breaking and 99% of people only care about affine-based GeoTIFFs.

Then we can support separate GCPGeoTIFF and, in the future maybe, RPCGeoTIFF classes. And add a top-level open method that returns one of the three classes.