Need to obtain shadow free images

[TolgaAktas]

I am looking for a way to get dataloader to fetch a triplet of cloud-free, clouded and cloud mask images of a given area.

1. What do you think might be the best strategy to modify the get_item function?
2. Do you think we can also sample by the the cloud cover ratio?
3. Does torchgeo make use of the information in MTL file?

[adamjstewart]

1. Depends on exactly what you're looking for. If you're looking for both cloud-free and clouded images of the same area, I assume you don't care about the time component? You could create one GeoDataset containing all cloud-free images and a second GeoDataset containing clouded images and cloud masks, then compute the intersection between these two datasets and sample from that. You would need to ignore the time dimension by not specifying it in the filename_regex.
2. We don't currently store metadata about cloud cover in the R-tree, but you could, and then design a custom sampler that weights which image is chosen based on that metadata.
3. We do not currently make use of information in MTL files in any of our existing datasets, although it wouldn't be hard to add a custom dataset that does (I believe it's just an XML file).

[adamjstewart]

Yes, IntersectionDataset is designed solely for the exact purpose you describe. The sampler is only allowed to sample from regions that appear in both datasets, and it will sample from both with the exact same bounding box, resolution, coordinate system, etc.

[TolgaAktas]

Do you know how the channels from two images from two datasets are combined? I am doing the following, but the samples from both images are identical for some reason. np.allclose return True

bands = ["B1","B2","B3","B4","B5","B6","B7","B8","B9","B10", "B11","QA_PIXEL"]
dataset = Landsat8(root=root_path,bands=bands)
dataset.rgb_bands = ["B4","B3","B2"]

cloudfree_dataset = Landsat8(root=cloudfree_path,bands=bands)
cloudfree_dataset.rgb_bands = ["B4","B3","B2"]

all_dataset = dataset & cloudfree_dataset
all_dataset.rgb_bands = ["B4","B3","B2"]
all_dataset.bands = bands

for elem in dataloader:
    break

idx = 0

sample1 = img[idx,:12].float()
sample2 = img[idx,12:].float()

np.allclose(sample1,sample2)

[adamjstewart]

They should just be stacked along the channel dimension. Are you sure root_path and cloudfree_path don't contain any of the same images? It's not clear from your code where img comes from. Can you share enough code and maybe some example images that I can reproduce the problem on my system?

P.S. You shouldn't need to override rgb_bands or all_bands to use the datasets.

[TolgaAktas]

for 0.3.1, without bands definition, it doesn't recognize a valid Landsat8 dataset from the directory. They shouldn't contain any similar images at all.

[adamjstewart]

I would recommend using 0.4.0 if possible. However, 0.4.0 contains a known bug that was fixed in #1135. This will be included in the 0.4.1 release, coming soonish.