Add DeepGlobe dataset for land cover

[saumyasinha]

Here is a sample plot for the dataset. These are with alpha 0, 0.5, and 1 respectively.

[ghost]

All CLA requirements met.

[calebrob6]

We should cite https://openaccess.thecvf.com/content_cvpr_2018_workshops/papers/w4/Demir_DeepGlobe_2018_A_CVPR_2018_paper.pdf in the docstring in addition to the kaggle/codalab links

[adamjstewart]

This looks great, thanks! I'm sure @calebrob6 will have more comments to add.

To get the tests to pass, you'll need to run black . on the files and maybe correct a few things that flake8 catches. Will also need tests, see other tests/datasets/* files for guidance.

[adamjstewart]

Can you add a tests/data/deepglobelandcover/data.py script that creates the fake data and zip file? See other tests/data/*/data.py scripts for examples.

[saumyasinha]

I made a small change in the draw_semantic_segmentation_masks function inside torchgeo/torchgeo/datasets/utils.py. The classes variable was earlier just using the unique values from the mask, but for segmentation scenarios where the image's mask doesn't include all the given classes such as an image with only "rivers" or "forest land" but not "urban land"...this function doesn't plot the mask right. I think this is because classes should be set to all possible class labels so that class_masks can be appropriately created and colors can be correctly assigned.

I may not have done a great job in describing the issue, but I am happy to talk more about it. This fixed the plots for me, but there might be more elegant ways of doing this.

[calebrob6]

Can you check to see what other plot methods use draw_semantic_segmentation_masks then test one of them with this change?

[saumyasinha]

Can you check to see what other plot methods use draw_semantic_segmentation_masks then test one of them with this change?

I did check for Potsdam and Vaihingen. Vaihingen had a few label masks that did not include all the classes, and I saw incorrect plots with the existing code, but correct ones when I included my change. I could show sample images when we meet next.

[calebrob6]

Can you post them here?

[adamjstewart]

Can you rebase your branch so that CI runs the new minimum version tests? There's a button on GitHub or you can do it manually from the command line. Also, can you add this dataset/datamodule to the documentation? You'll need to add a section to docs/api/datasets.rst and docs/api/datamodules.rst under non-geospatial datasets. You'll also need to add a line to the table in docs/api/non_geo_datasets.csv.

[saumyasinha]

Can you post them here?

Yeah!
The following plots are a sample from Vaihingen dataset, that has 5 out 6 classes in their segmented outputs. Attaching them here:
The first one is plotted with the existing draw_semantic_segmentation_masks, followed by the one plotted with my suggested change and the third one is the manually downloaded mask image (for reference - we are looking at top_mosaic_09cm_area30.tif).

[ashnair1]

Can you check to see what other plot methods use draw_semantic_segmentation_masks then test one of them with this change?

I did check for Potsdam and Vaihingen. Vaihingen had a few label masks that did not include all the classes, and I saw incorrect plots with the existing code, but correct ones when I included my change. I could show sample images when we meet next.

OSCD and XView2 also use draw_semantic_segmentation_masks

[calebrob6]

So this PR uncovered some unexpected behavior in torchgeo.datasets.utils.draw_semantic_segmentation_masks. When we added this function with the xView2 dataset it looks like we added a little hack to ignore the "0" background class when generating the masks to pass to torchvision's draw_segmentation_masks so that the background would be transparent. E.g. xview2.py has 3 colors in the colormap while having 4 classes -- https://github.com/microsoft/torchgeo/blob/main/torchgeo/datasets/xview.py#L63.

Generally, I think a good solution would be to force each dataset to have a color for each class (even background classes) then add an "ignore" class as an argument to draw_semantic_segmentation_masks. We will need to update OSCD and XView2 to implement this. @adamjstewart, thoughts?

[calebrob6]

This should be good to go now!

[adamjstewart]

Almost there! Just a few remaining questions

[adamjstewart]

Running the tests results additional files being created during zip file extraction. We should add these to the repo so that running the tests doesn't result in new files in git status.

[adamjstewart]

#651