I’m trying to use a UNet in Pytorch to extract prediction masks from multidimensional (8 band) satellite images. I’m having trouble getting the prediction masks to look somewhat expected/coherent. I’m not sure if the issue is the way my training data is formatted, my training code, or the code I’m using to make predictions. My suspicion is that it is the way my training data is being fed to the model. I have 8 band satellite images and single band masks with values ranging 0-n number of classes with 0 being background and 1-n being target labels like this: