Sentinel-2: use 10,000 as scale factor

[osgeokr]

My suggestion is to apply a scale factor that converts DN to Reflectance of Sentinel-2 images.
Pls refer to this URL: https://docs.sentinel-hub.com/api/latest/data/sentinel-2-l2a/

Surface Reflectance values lies usually between 0.0 and 1.0. However, Specular effects on surface or clouds could lead to values higher than 1.0.
Pls refer to this document(p.40.): https://sentinel.esa.int/documents/247904/685211/Sentinel-2-MSI-L2A-Product-Format-Specifications.pdf

If the brightness adjustment is needed, it would be conceptually desirable to apply a seperate brightness factor. This is equivalent to adjusting the alpha value in image processing, regardless of the scale factor. Pls refer to this document: https://docs.opencv.org/3.4/d3/dc1/tutorial_basic_linear_transform.html

# image(reflectance) = image(DN) / 10000
bright = 5
adjusted_image = image * bright

The concept will be clearer if the brightness adjustment method is separated from the scale factor, such as torchvision.transforms.functional.adjust_brightness.

Thank you.

See #1025

[adamjstewart]

Can you also fix docs/tutorials/custom_raster_dataset.ipynb? You don't have to open a new PR, you can just add another commit to the same branch and it will automatically show up.

[adamjstewart]

I think we'll still want to clamp it to the 0–1 range (so keep the max=1). Otherwise, matplotlib may get confused whether the scale is supposed to be 0–1 or 0–255.

[osgeokr]

@microsoft-github-policy-service agree

[osgeokr]

I agree. That seems like a wise choice.

[calebrob6]

Thanks @osgeokr, do you mind calling .plot() on a Sentinel2 image with this change and posting what the result looks like?

[osgeokr]

Thanks @osgeokr, do you mind calling .plot() on a Sentinel2 image with this change and posting what the result looks like?

In custom_raster_dataset.ipynb, the current scale factor is set to 6000. 6000 is assigned to the left figure and 10000 is assigned to the right figure. My opinion is to distinguish between scale factor and brightness factor in TorchGeo. Because this can be confusing to remote sensing majors like me. thank you.

[adamjstewart]

10K looks better for those glaciers, but probably worse for agricultural land. But that's fine, we can always add an additional scale factor like you suggested. We should prob update the figure in the notebook though. Let me do that before I forget.

[adamjstewart]

I don't seem to have permission to push to your branch. Can you either enable commits from maintainers or re-run the notebook yourself to update the plots?

[calebrob6]

Thanks @osgeokr! I agree.

[adamjstewart]

I'll just update the notebook in a follow-up PR. Thanks for the fix!