Closed yerzhan7orazayev closed 10 months ago
We have not worked on super-resolving non-RGB bands, but it would certainly be doable in this codebase if you have a dataset of high-resolution imagery with NIR bands. You would just need to change the number of desired output channels, and maybe a few other hardcoded values.
Thanks for your reply. Unfortunately, I do not have dataset of high-resolution imagery for non-RGB bands. I was thinking that using your model directly on single channel input of non-RGB image will not give the desired output? If the problem is with number of channels, one can simply replicate up to three channels at input stage.
Yeah, because our model is just trained to go from Sentinel-2 (RGB or all bands) -> NAIP (RGB), it won't know how to produce super-resolved non-RGB bands. I'm not sure what the output would look like if you tried feeding in three channels of non-RGB to our pretrained models. If you try it, please share the results!
Is there any possibility to SR not only RGB images from S2 rather other spectral bands, such as NIR?
Thanks.