Multi-resolution neural networks for subsurface exploration with sparse labels.

Prospectivity mapping, whether for groundwater or rock mineralization, requires the integration of multiple remote data sets, a task well suited to machine learning, and convolutional neural networks in particular.

In general, in subsurface prospectivity mapping there is only one image, with multiple channels (topography, geology, geophysical survey data, for example), with pixel wise training labels.

Until now, most networks trained on small patches from one larger input image, however, with labels at one pixel the question of how to handle the sparse labels arises.

We propose a projected loss-function for training convolutional networks with a multi-resolution structure, including variants of the U-net. Our networks learn from a small number of large map data sets without creating patches. The projected loss-function enables training on labels with just a few annotated pixels, such as water level measurements at sample wells, or ore production quantities, and has no issue with the other unknown labels.

We demonstrate our method for a water prospectivity example in California, and mineral prospectivity in the Canadian arctic.