Applying convolutional neural networks to extremely sparse image datasets using an image subdivision approach