Sentinel-2 difference between 'super-resolution' and 'resampling'

Hello friends.
question 1:
I have a Sentinel 2 image that I took in SNAP Resample using band-2 (Blue 10 meter), but when I export a single band like (SWIR), the pixel size doesn’t change, i.e it remains the same 20 meters.
What is the reason?

question 2:
and what is the difference between ‘super-resolution’ and ‘resampling’ in SNAP?

question 3:

Is the following method correct?

1 - Mosaicking
2 - Get a subset
3 - Super Resolution (Optic > Sentinel-2 super-resolution)

In advance thanks for your guidance.

Maybe you have selected nearest neighbor resampling so the look stays the same?

The super resolution is more advanced than the resampling. Depends on your use of the data what makes more sense.

1 Like

How have you exported the band? Have you really used the new resampled product when exporting the band?

Regarding your workflow. I don’t know if the super-resolution can work with a mosaicked product. If yes, then the steps are in the right order.

The super-resolution is explained here:
Super-resolution of Sentinel-2 images (brodu.net)
In brief, resampling does the resampling for each band independently while the super-resolution considers other bands when creating interpolated values.

Please note, there is also a S2 Resampling in SNAP. This one is like the generic resampling but considers the view angles differently. They need a special treatment when they are interpolated.

2 Likes

Thank you very much. Your tips can help me.
And I’m proud to have friends like you.
And I tried two methods, Bilinear and Bicubic, and got the right result.

I wish the best for you.
:bouquet: :bouquet:

Thank you for your answer.
I extract a single band (SWIR). And I exported resampled images to .tif.

I used @ABraun advice and tried two methods of Bilinear and Bicubic and got the answer.
And thank you very much for your explanation of the difference between the two methods.

Hello friends,

Does this mean one cannot resample using the Nearest neighbor? Cubic or bilinear are normal good for continuous data.

Thanks in advance.

does your question refer to the “super resolution” operator for Sentinel-2 or in general?
Nearest Neighbor resampling means that the pixel size is changed but the values remain the same. So visually there will be no difference. Bilinear or others calculate a new value based on local statistics.

Thank you ABraun, yes that is what I understand however when I run NN using the SNAP tool the 20m pixels do not change. They remain 20m instead of downscaling to 10m.

how did you check it?

A simple way was to put one of the 10m bands, e.g. band 2 on top of the “resampled” band 5. The pixels do not look the same. So I didn’t have to calculate it as visually the two looked unequal.

I’m trying to understand what you mean. Can you please give an example with a screenshot?

The image uploaded is a resampled product, from 20m to 10m. The pixel size is still 20m, showing that the resampling somehow did not work?

well, if you selected “Nearest Neighbor” resampling, the pixel value of the new pixels stays the same. So one 20x20 m pixel is split into four 10x10 m pixels with the same value. Accordingly, you cannot see the change.

2 Likes

I get it now. Thank you soo much. My next question; What is then the point of downscaling using NN, as it still keeps the values (pixels?) the same? Also, can I combine the resampled 10m band with the other 10m bands, since they have somewhat the same resolution?

Depends on why you want to do it in general. Resampling to a higher spatial resolution never brings you more information, it just geometrically alters the data so that, for example, you can stack it with other data of same resolution.

1 Like