plotting RGB image of the entire Sentinel2 TILE

[roybenhayun]

the Sentinel2 dataset can render the RGB image of a single sample. at that point the sample already contains pytorch tensors of a piece of the entire Sentinel2 TILE.

we would like to render the RGB image of the entire TILE, once before the actual training starts.
guessing that it should load RGB bands from the .jp2 files (under \GRANULE...\IMG_DATA), then some conversions may occur, then put them into tensors, then some tweaks for matplotlib etc. there could be some work involved.

is there a way in TorchGeo for that?
or possibly can someone share code that does this?

thanks in advance!

[adamjstewart]

Our dataset.plot methods can only accept a sample as input. You could set the size to be the size of an entire tile if you want. You'll likely need to downsample the resolution, otherwise the image is too large to plot.

[roybenhayun]

ok. see working code using the Dataset and PreChippedGeoSampler.
other then being slow, look reasonable? any suggestion?

    ds = Sentinel2(r'SENTINEL2.SAFE')
    print(f"ds: {ds}")  
    width = ds.bounds.maxx - ds.bounds.minx
    height = ds.bounds.maxy - ds.bounds.miny
    print(f"width: {width}, height: {height}")

    fig, axs = plt.subplots(nrows=1, ncols=2, figsize=(2 * 5, 5))

    axs[0].set_title("full input image")
    full_ds_sampler = PreChippedGeoSampler(ds)
    dataloader = DataLoader(ds, sampler=full_ds_sampler, collate_fn=stack_samples)
    print(f"dataloader: {dataloader}")
    for sample in dataloader:
        image = sample["image"]/10000.0
        print(f"image: {image.shape}")
    permuted_image = image.cpu()[0][1:4, :, :].flip(0).permute(1, 2, 0)
    print(f"permuted_image: {permuted_image.shape}")
    axs[0].imshow(permuted_image)

    axs[1].set_title("cropped input image")
    bbox = BoundingBox(ds.bounds.minx + width / 4, ds.bounds.minx + (3 * width / 4),
                            ds.bounds.miny + height / 3, ds.bounds.miny + (2 * height / 3),
                            ds.bounds.mint, ds.bounds.maxt)
    chip_sampler = PreChippedGeoSampler(ds, bbox)
    dataloader = DataLoader(ds, sampler=chip_sampler, collate_fn=stack_samples)
    print(f"dataloader: {dataloader}")
    for sample in dataloader:
        image = sample["image"]/10000.0
        print(f"chip image: {image.shape}")
    permuted_chip_image = image.cpu()[0][1:4, :, :].flip(0).permute(1, 2, 0)
    axs[1].imshow(permuted_chip_image)
    plt.tight_layout()
    plt.show()
    plt.close()

[adamjstewart]

It's slow because each image is massive, if you use:

ds = Sentinel2('SENTINEL2.SAFE', res=1000)

it will be much faster. Then you don't have to bother with bbox cropping.

[roybenhayun]

the bbox cropping is for other purposes that do require cropping. it sneaked into the code...

the res=1000 is fantastic! saves lots of time. many thanks for this tip.