Computer vision tools for geophysicists

Geophysics will be one of the most important disciplines in the present drive to explore covered regions of the Earth. As more data are being collected, there is a growing requirement to develop tools which can consistently extract relevant information from geophysical imagery. Computer vision methods are a key platform for this, but are only useful if they are adaptable to different tasks and the outputs are geologically explainable. Here we present a workflow for extracting meaningful information from geophysical imagery which provides a basis for a range of analysis tools depending on the task at hand. The process works by training a computer vision model to generate numeric feature vectors which capture the spatial and compositional information contained in geophysical imagery of any type with any number of image bands. Two model architectures are presented, one utilising contrastive self-supervised learning with a convolutional neural network, and the other using vision transformers. We show how these models can be used to extract quantitative information from magnetics, gravity, and multispectral imagery which can be input to downstream tasks such as unsupervised classification and similarity mapping. Further, these workflows can be applied to more than one image dataset at a time, allowing for novel combinations of data to provide a more comprehensive characterisation of geology. These techniques give geophysicists the ability to efficiently assess the content of large geophysical images in detail with consistent results, providing more resources for critical analysis and decision making.