Mpanza, M, Wistockk, J & Rupprecht, S 2024, 'Mapping tailings storage facilities associated with abandoned mine sites', in AB Fourie, M Tibbett & G Boggs (eds), Mine Closure 2024: Proceedings of the 17th International Conference on Mine Closure, Australian Centre for Geomechanics, Perth, pp. 615-626, https://doi.org/10.36487/ACG_repo/2415_44
(https://papers.acg.uwa.edu.au/p/2415_44_Mpanza/)
Abstract: Besides tailings storage facility (TSF) structural failure, there are various other threats that result from poor management of TSFs. For example, poor rehabilitation of TSFs can result in air pollution by wind-generated dust affecting human health and wellbeing. Furthermore, water pollution owing to acid rock drainage can result from a lack of TSF lining. Recently some TSFs have commenced posing a safety challenge for communities residing in proximity as they become infested with illegal mining activities. The 2020 Global Industry Standard on Tailings Management (GISTM) hopes to achieve zero harm to people and the environment, with zero tolerance for human fatality. To conform with the GISTM, companies are required to be accountable and responsible, and all stakeholders must prioritise TSF management. Two recent TSF-related cases (the Jagersfontein tailings dam failure and the crime activity by illegal miners at the North Sands Dump, both in South Africa) are significant motivations for effective tailings management. It appears that the significant challenges with TSF management in South Africa result largely from mine operation abandonment and legacy mine sites. This study suggests the mapping of TSFs associated with abandoned mine sites as a starting point for effective TSF management in South Africa. The study uses Faster Region Convolutional Neural Network (RCNN) and the Mask Region Convolutional Neural Network deep learning models to identify TSFs in Gauteng Province. Satellite (Sentinel-2) imagery data was used to train the RCNN models for TSF identification focusing on the red, green and blue band combinations. Sentinel-2 satellite imagery was obtained and processed using ESRI’s Arc GIS Pro software and was used to train and test deep neural networks on the same platform using the Pro Notebook with inbuilt Python scripting. The Mask RCNN was able to accurately identify TSFs (which the model was not trained on) with minimal false positives. The efficacy of the approach is demonstrated by the discovery of 46 TSFs that were not part of the training data. This study highlights the potential of deep learning to assist in TSF management by identifying TSFs which might be ownerless and abandoned.

Keywords: faster RCNN, Sentinel-2, object detection, TSFs, Mask RCNN