RandomGeoSampler does not work with flipped geotifs

[tanpinquan]

Description

Certain geotifs have their axes flipped such as this one, which has the following transform flipping the direction of the x-axis:
Affine(-0.346241067933152, 0.2915401535597796, 366317.96341952495, 0.2915401535597782, 0.3462410679331555, 130406.04105973356).

Depending on the transform, I believe this affects how random bounding boxes bounds should be ordered . For the example geotif above with x-axis flipped, this line maybe should be bounds = (query.maxx, query.miny, query.minx, query.maxy) instead of bounds = (query.minx, query.miny, query.maxx, query.maxy). I have not tested it myself.

RandomGeoSampler and RasterDataset do not check for this, resulting in a Bounds and transform are inconsistent error

Steps to reproduce

1. Download the sample data from here
2. Run the following code to reproduce

from torch.utils.data import DataLoader
from torchgeo.datasets import RasterDataset
from torchgeo.samplers import RandomGeoSampler
   
   
class UmbraDataset(RasterDataset):

   # umbra tifs are named like: 2023-08-02-17-08-40_UMBRA-05_GEC.tif
   filename_glob = '*.tif'
   filename_regex = r"^(?P<date>\d{4}-\d{2}-\d{2}-\d{2}-\d{2}-\d{2})"
   date_format = "%Y-%m-%d-%H-%M-%S"
   is_image = True
   separate_files = False 
   

dataset = UmbraDataset('/path/to/data/')

sampler = RandomGeoSampler(dataset, size=512, length=None)
dataloader = DataLoader(dataset, sampler=sampler, batch_size=2)

for i, batch in enumerate(dataloader):
   sample = batch

Version

0.5.1

[adamjstewart]

I'm able to reproduce this bug. @nilsleh this reminds me of a bug you had when working on rioxarray support.

I wonder if we could use something like min(query.minx, query.maxx) and max(query.minx, query.maxx) to solve this problem. But I'm not sure where all we would need to put it.

[nilsleh]

For rioxarray support I encountered .nc files where the latitude and longitude coordinates were sometimes from -90 -> 90 or vice versa etc, but I think I handled that by sorting them to a common structure when reading the xarray.