Automated prediction of tailings areas at historic gold mine districts in Nova Scotia using multispectral images and a random forest classifier

Abstract

Satellite imagery can be analyzed to offer a preliminary regional assessment of mine tailings indicators, enabling identification before performing in-depth fieldwork. Nova Scotia, Canada, still retains mine tailings produced in the 1860s to the 1940s in 64 historic gold districts, which exceed soil guidelines for arsenic (As) and mercury (Hg) levels. Tailings data often relies on historical maps, which may not accurately depict the current extent due to wind and rain transportation. This study employs historical data to train a classifier, enabling the classification of multispectral satellite images from Sentinel-2. Both pixel-wise and object-based methods were evaluated, yielding a median F1-score above 0.7 for most tested methods at two case study sites. Accuracy varied at other sites, particularly those with significant proportions of wetland areas. Lastly, we investigate false positives and propose future research to create a more resilient classifier.