I'd like to apply some machine learning algorithms to satellite imagery that we've collected, but I want to encourage invariance to factors such as sunlight intensity, time of day, atmospheric conditions, and angle of satellite to target. Others could be things such as variations in imagers between satellite hardware for different providers or even within providers.
An example default augmentation often used in deep learning with images is adjusting gamma factor (scaling brightness up and down uniformly), but in reality this doesn't mimic lighting changes the way they usually occur. Typically the changes are more complicated and non-linearly affect each color differently.
I have so far been unable to locate any good methods that have devised random transformations, or really been a potential solution for this issue. Hoping for references/ideas from this crowd.