How Detectron2 helps make mines safer and more efficient

Mining sites can produce anywhere from 100 to 1,000 meters (or nearly the length of nine football fields) of drill core per day and generate hundreds of reference images per week, depending on their setup. All of these samples have to be carefully analyzed by geologists not only for rock quality but also for indications of potential hazards, such as faults and fractures — both of which are crucial to the design and engineering of mines, as well as their efficiency and productivity. It’s a labor-intensive manual process that was in need of a machine learning (ML) solution.

Australian technology services company DiUS partnered with Solve Geosolutions to develop that ML solution. Datarock, a SaaS solution targeted at the mining industry, leverages various PyTorch tools, including PyTorch-based object detection library Detectron2, to train ML models with geological imagery. Detectron2 is designed to support a wide range of image analysis models for both image classification and object detection. It also offers a modular design and support for panoptic segmentation, which allows it to perform the kinds of sophisticated object recognition tasks found in cutting-edge research and novel commercial and enterprise applications.

Datarock allows mining operations to use ML and computer vision to analyze the geology of mineral deposits and streamline the arduous process by leveraging the wealth of images produced by a drill site and turning them into a dataset — processing the raw images and segmenting important geological information within them. Utilizing Detectron2, Datarock’s developers were able to create models capable of conducting rock quality designation (a metric used to understand the strength of rock), rock fracture prediction, and other high-value but difficult-to-collect geological analyses.

According to DiUS, the Datarock platform has already processed more than 1 million meters’ worth of drill core images. DiUS and Solve Geosolutions say Detectron2 is particularly advantageous because it is up to four times faster at training than previously developed models. We plan to continue to refine and upgrade Detectron2 by enhancing its flexibility and implementing new models to support new and novel use cases like Datarock.

Learn more about the PyTorch-enabled tools supporting this work.