Non-continous GeoDataModules/GeoDatasets

[MathiasBaumgartinger]

I am trying to set up a GeoDataModule where the images/labels are scattered and are not spatially continous.

I defined __get_item__ using rasterio by querying the individual images manually depending on the query BoundingBox. However, when doing a test run, the queries go beyond the the individual tiles and I end up getting

warnings.warn('shapes are outside bounds of raster. '
               'Are they in different coordinate reference systems?')
ValueError: Input shapes do not overlap raster.

Am I correct with the conclusion that GeoDataModules need to be continous? So in case I have a patched dataset I would define it as NonGeoDataset although the individual images do have spatial information?

[adamjstewart]

I defined __getitem__ using rasterio by querying the individual images manually depending on the query BoundingBox.

You should not need to define your own __getitem__, the RasterDataset.__getitem__ should handle everything you need. There is likely a bug in your __getitem__.

the queries go beyond the the individual tiles

This is actually expected behavior for GridGeoSampler. We've gone back and forth on this a couple of times. In the original implementation, we simply skipped patches that didn't fit an entire patch inside. In #630, we changed this to get full tile coverage. #1329 changed this again so that we actually sample outside the bounding box of the tile. This was done so that we get full coverage of the tile, but don't overlap and compute metrics twice.

Am I correct with the conclusion that GeoDataModules need to be continous? So in case I have a patched dataset I would define it as NonGeoDataset although the individual images do have spatial information?

GeoDatasets do not need to be continuous. You can use either GeoDataset or NonGeoDataset depending on whether you need to combine your dataset with another and depending on how large the patches are.

[MathiasBaumgartinger]

Thank you so much for your timely answer, @adamjstewart!

Definitely, RasterDataset.__get_item__() does return the image as expected. However, as I am working on a SemanticSegmentationTask, I do need an image as well as a mask. The documentation clearly states:

For datasets with both model inputs and outputs, a custom getitem() method must be implemented.

Am I overthinking things? I would gladly accept a simpler solution to the problem 😄.

After writing the previous post I closely followed the implementation of the existing Chesapeake Land Cove dataset/module which seems to prevent the error, by taking into account the dataset projection. They first increase the patch_size and in turn crop each sample back to the original patch_size .

[adamjstewart]

Ah, we should update those docs. Can you open an issue for this (or even a PR to fix this)?

The current recommended strategy is to use 2 RasterDatasets (one for the image and one for the mask) and use IntersectionDataset to combine them. See L7 Irish, L8 Biome, and I/O Bench for examples of this in torchgeo/datasets.

[MathiasBaumgartinger]

FYI: #2293