PROGRESS TOWARD AUTOMATING GEOREFERENCING AND FEATURE EXTRACTION OF GEOLOGIC MAPS

Geologic maps contain a broad range of information (“data”) – including unit descriptions, structural information, and sample locations, as well as other data used broadly throughout the geosciences. Modern mapping approaches use IT and GIS software that enable organization and maintenance of digital map products that can be readily analyzed and used for decision-making. However, most historic-and report-based maps are stored as non-georeferenced images. These maps can contain detailed information on geology, exploration, and mining activities. Georeferencing maps and extracting geologic and related features is time consuming and requires specialized software and skills.

The U.S. Geological Survey (USGS) partnered with the Defense Advanced Research Projects Agency to address map data needs for mineral resource assessments. Two challenges were presented to the public: 1) automated map georeferencing, and 2) automated extraction of legend-based features. Competitors were provided extensive training and validation datasets sourced from the National Geologic Map Database (NGMDB); they were free to design their own methods using open-source software. The proposed solutions employed image processing techniques and deep learning tools. Results were evaluated using standard metrics for testing model accuracy and top performers were asked to submit technical briefs and code.

Competitors’ solutions for georeferencing and feature extraction show promise even if imperfect. The complexity of the problem was demonstrated in the feature extraction challenge, where different competitors performed best on each of the three feature types - polygons, lines, and points. The USGS is examining options to develop proof of concepts from the competition into prototypes and then final products. This would enable incorporation of geologic and related data from ~100K maps in the NGMDB and thousands of additional maps from USGS publications, industry technical reports, and scientific literature into mineral resource assessments and other applications. Publication of open-source software tools and ‘lessons learned’ from using AI competitions to improve efficiencies in operational tasks will greatly benefit the broader geoscience community.