U-Net convolutional neural network models for detecting and quantifying placer mining disturbances at watershed scales

Abstract

Placer mining is a mineral extraction method in floodplains that involves the removal of earth material to access mineral-laden sediments, a process that can have significant and long-term impacts on aquatic ecosystems. Given the widespread nature of mining, new tools are required to monitor the potential watershed-scale ecological impacts of placer mining. This study adapted and evaluated a deep learning model – a U-Net convolution neural network, and compared it to a traditional image classification method – random forests (RF) – to detect and quantify the area of post-placer mining disturbance at the watershed scale. Overall, both random forest and U-Net models performed well at classifying digitized image samples where placer disturbances were mapped. Sensitivity in placer classification was high, with both modelling frameworks achieving at least 75% accuracy in the classification of digitized placer samples in 7 out of 12 modelling scenarios. Misclassification of non-placer pixels as placer was highly variable among different models, data configurations, study sites, and time periods. Commission errors (i.e., incorrectly classifying a non-placer pixel as placer) were typically the result of models labelling water areas or forest areas as placer – errors which may have only marginal practical significance. In general, U-Net models performed better in terms of minimizing misclassification errors, whereas RF models performed slightly better in classifying known placer pixels. We conclude with discussions on the advantages of deploying U-Net and RF models for placer detection, challenges that may be encountered in operational systems that employ the models, and identifying outstanding issues which need to be addressed in future placer modelling studies.