Comparative analysis of machine learning methodologies for mapping global-scale mining

Keywords: Mining, geopolitics, remote sensing, machine learning, critical minerals
Abstract Type: Poster Abstract

Abstract

Extraction of critical minerals is key to enabling a transition to a sustainable future. Mapping these resources globally has been a challenge for researchers, as critical mineral development often occurs in areas prone to conflict, poor governance, and limited regulation. Given the resulting limited means of doing global analysis on these important resources, scholars have attempted to create global geodatabases through an aggregation of industry and government data, using it to identify spatial information on mining areas. Our early validation and analysis of expert-delineated mining boundaries suggests significant limits to manually mapping the extent of extractive areas and inconsistencies both within and across studies. Employing machine learning of remotely sensed data can allow for a scalable and repeatable approach to identify these mining areas. This is not without challenge. Given the diversity in features at mining sites, it can be difficult trying to use a single model to identify mine area boundaries if it were to include one or several open pit mines, tailing ponds, built infrastructure, and other features. By using a multi-stage process to identify common features, however, researchers can get towards greater accuracy in identifying these areas. This research will provide a comparative analysis of various machine learning methods, from unsupervised classification to deep learning methodologies like the Segment Anything Model to assess their relative capabilities to detect critical mineral extraction. This analysis will provide insights for researchers looking to understand the use of remote sensing to detect complex land use/land cover disturbances.