search

kappazeta / km_predict

S2 full image prediction
Apache License 2.0
20 stars 9 forks source link

Normalization #23

Closed kanlions closed 4 months ago

kanlions commented 1 year ago

I observe that in data generator python file there are separate normalization for UNET and Deeplab V3+. Can you please tell me how those values are generated?

tanyashtym commented 1 year ago

Hi!

Normalisation parameters were calculated for each band over the entire training set. You can find the code here: https://github.com/kappazeta/cloudmask-fit/blob/master/cm_fit/training/train_utils.py

The normalisation parameters for U-Net were calculated for the dataset that covers only terrestrial North European summer conditions. You can find more information here: https://www.mdpi.com/2072-4292/13/20/4100 In turn, we used KappaSet (https://zenodo.org/record/7100327) to train the DeepLabv3+. Therefore, normalisation parameters were calculated through the KappaSet's training set.

If you have more questions, feel free to ask!

indrek-sunter commented 4 months ago

Closing, for there have not been any more questions.