A few things that I used that I don't see in your code:

- Data augmentation (randomly flip, rotate, adjust brightness/contrast)

- Train with all negative samples. With each epoch I selected a random negative subset equal to the size of the positive samples.

- Remove or handle no-data values. Values <0, typically it was set to -9999 and can cause issues within the CNN. You can remove those samples or fill no-data values with mean valid data.

- Your dropout value seems to be very high.

It will also be better I think to output raw float probability values rather than int 0 or 1 values.

@backyard baseball In my opinion, it would be more beneficial to output raw float probability values instead of int 0 or 1.