Issues with the cloud masks

[mauropk]

I have downloaded the training data at the link that you provide in the readme. However the images in the cloud mask folders do not seem to correspond to actual clouds in the satellite images. Each cloud_mask image contains a "blob" in the middle of the image, even when the corresponding RGB or NIR image does not contain any cloud. Are the correct cloud masks stored anywhere?

[andrzejmizera]

Dear Mauropk,

the clouded RGB images are synthetically generated as described in the "Cloud Removal from Satellite Imagery using Multispectral Edge-filtered Conditional Generative Adversarial Networks" paper and shown in Figure 1 of this publication. The cloud_mask image is combined with a real cloud image and independently fused with a cloud-free RGB image and its corresponding NIR image to generate a training sample consisting of a pair of clouded NIR and RGB images. The implementation of the syntheses can be found in the get_example() method of the NIRRGB2RGBCLOUD_SYNTHETIC class in the MEcGANs/data.py file (lines 126-158).

Best regards,
Andrzej