Replies: 1 comment 5 replies
-
Yes. This is where the scale factors come from.
This is the script used to map from float32 to uint8. We download the data first and preprocess later. Of course, you could also do the preprocessing on GEE itself.
There are many ViT models available. The specific ones we trained are designed for 224x224 px patches. We need patches that are larger than that in order to perform random cropping. Hope this helps! Let me know if there are any other questions I can answer. |
Beta Was this translation helpful? Give feedback.
5 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Hi,
I'm trying to implement one of the new pretrained LandSat Models. More specifically the LANDSAT_ETM_SR models that were pretrained on the SSL4EO-L dataset.
I read the paper (https://arxiv.org/abs/2306.09424) but for me It's hard to understand the preprocessing for the image values.
It says:
"The official scale factors suggested by the USGS to map between Level-1 and Level-2 Landsat imagery and the visualization range
recommended by GEE for each sensor are used to map from float32 to uint8"
From my understanding this is then the resulting preprocessing step, right? (GEE example)
The recommended visualization range for the image after applying the scale factor and the offset (so for the "var image...") is
(min: 0, max: 0.3). Is this also taken into account during preprocessing in GEE?
I would be glad if somebody could help me. Is there maybe a ressource where I can find the necessary scaling/preprocessing steps for the corresponding models, that I am missing?
Another question that maybe rises from my lack of understanding of ViTransformers is:
The image sample size, as stated in paper, should be 264 × 264 px. Why does the ViT require 224x224 images?
Beta Was this translation helpful? Give feedback.
All reactions