Once you train your baseline network, try to innovate with augmentation, network architecture, etc.
Remark: I haven't encountered any real dataset with paleokarst. So, if you know of any open seismic for this problem, please write it down in the comments. It would help to validate the results of NN.