This project is about removing cloud and haze obstacles from satellite images using deep learning frameworks.
These satellite images all are 16-bit depth, 4 channel RGB-NIR images, where the pixel values represent raw values of Level 4 Top of atmosphere reflectance (TOAR).
We tested out a new idea of a NDVI/NDWI loss function, which uses geoscience metrics as a loss function that gives it a cross channel color stabilizaiton. This remains to be tested thoroughly. Here is an example of our result:

Before / After (Satellite Image 4 channel TOAR):

Before / After (NDVI Heat Map):

Before / After (Satellite Image 4 channel TOAR):

Before / After (NDVI Heat Map):

Here is our overall architecture:

Please use the shell script main provided "main_trainNDVI_NDWI_s2.sh".
Command could be like:
sh main_trainNDVI_NDWI_s2.sh 0 L1C_England_augmented 0 80 0.0001 8 NTIRE2018_Multiscale_NDVI_NDWI 512 The argument L1C_England_augmented is the dataset pathname of your own dataset.