A neural gym for training deep learning models to carry out geoscientific image segmentation. Works best with labels generated using https://github.com/Doodleverse/dash_doodler
not our problem; either the user prepares their imagery in the way described, or modifies the loss function to deal with a KNOWN (to them) class imbalance