Global Study of Tailings Dam Failures by Statistical and Remote Sensing Methods

Abstract

The growth of the mining industry, mainly driven by increased resource demands, has accelerated the production and storage of mine waste. Tailings comprise the fine-grained fraction of this waste. Due to their intrinsically hazardous characteristics from a geochemical and geotechnical perspective, tailings are conventionally discharged in slurry form behind constructed dams for public and environmental safety purposes. However, tailings dams have been under significant scrutiny following a number of high-profile, catastrophic breach events in, for example, Canada (2014), Brazil (2015 and 2019), India (2017, 2019, and 2021), China (2017 and 2021), and South Africa (2022). This thesis aims to study such phenomena using advanced statistical and remote sensing methods. Specifically, this thesis focuses on quantifying the failure rate statistics via a magnitude-frequency approach, mapping breach-runout processes by GIS analysis of historical satellite/aerial imagery, and analyzing pre-failure deformation patterns using satellite InSAR. By this, the thesis enhances knowledge on how often tailings dam failures occur, the common causal and failure mechanisms for these events, the potential downstream consequences, and the capabilities of satellite technology to monitor, and predict instability in, tailings dams.

Chapter 2 notes that global magnitude-frequency statistics of dam failures in general remain poorly quantified, out-of-date and/or limited in scope in scientific literature. Addressing this gap would give insight into how the hazard-risk of tailings dams has evolved over time in comparison to water-retention dams, and would provide quantitative benchmarks for estimating likelihoods of dam failures and their societal impacts to support risk assessments. The chapter introduces and analyzes new datasets and estimates of the construction and failures of large reservoir facilities (LRFs) and tailings storage facilities (TSFs) worldwide in the period 1965–2020. Long-standing data gaps on LRF failures in China, and subsequently worldwide, and on constructed TSFs worldwide are addressed by new estimation/extrapolation techniques to illustrate the range of uncertainty in the results.

The total number of LRF failures is estimated to have been between 394 and 608. The annual numbers of newly constructed and failed LRFs declined near-proportionally, thus the cumulative failure rate of LRFs stayed fairly constant, falling in the range of 1.2% to 1.8% as of end-of-2020. The rate drops to at least 0.7% when excluding China. The cumulative fatality rate of LRFs reduced over time to 1.2 deaths per constructed facility, and falls in the range of 64 to 98 deaths per failure, as of end-of-2020. Failures of LRFs with very high storage capacities (>200 M m3) have continued to occur since 2016. In comparison, the annual number of TSF failures stayed relatively constant, whereas the annual construction rate of TSFs is estimated to have increased by ~3x, thus the cumulative failure rate of TSFs declined over time. When assuming the lower-estimate of the number of constructed TSFs (6810), the cumulative failure rate is ~4.4% as of end-of-2020. When adopting the upper-estimate (20,230 TSFs), a rate of ~1.5% is obtained, which falls in the same order as the corresponding rate of LRFs. A review of published estimates of existing TSFs worldwide indicates that the “true” rate is much lower than 4.4% and closer to 1.5%. The cumulative fatality rates of TSF failures reduced over time to 0.1–0.3 death per constructed facility and 6 deaths per failure as of end-of-2020, which are lower than those of LRFs. However, the size and the environmental impact of TSF failures have increased on average worldwide, especially since 2014. The rising global rate of failed tailings volumes has been approximately proportional to the rising global rate of tailings production since the 1990s.

Heavy rainfall events and intensifying precipitation patterns are concluded to be statistically important causative variables for the failures of both LRFs and TSFs. This has implications for the design and management of storage capacity, freeboard, facility drainage and spillways under climate change conditions. The results are applicable broadly on a global scale and are conditioned by uncertainties in the data and the methods used to address data gaps. To improve the robustness of future statistical analyses, a more comprehensive public disclosure effort is necessary, particularly with respect to reservoir facility failures in China and constructed TSFs worldwide.

Chapter 3 focuses on the downstream mass flows resulting from TSF failures (i.e. tailings flows). The chapter is supported by a new global database of 63 tailings flow cases that have been remotely analyzed through a compilation of satellite imagery, digital elevation models and literature. The synthesis provides insight into the influence of impoundment conditions, preconditioning and trigger variables, failure mechanisms and the downstream environment on tailings flow behavior. The database also sheds light on the limitations of data quality and availability in the public domain. Magnitude-frequency statistics indicate that TSF failures that have produced catastrophic tailings flows with total outflow volumes of ≥1 M m3 have occurred at a mean recurrence interval of 2–3 years over the period 1965–2020. Weather hazards and impoundment drainage issues are identified as major causative variables. The occurrence of liquefaction and/or the incorporation of free water are sufficient conditions to trigger extremely rapid, highly mobile behavior. Travel path confinement and steeper bed slopes enhance flow velocities (peak of 25–30 m/s) and kinetic energy, whereas flow mobility appears to be exacerbated along major rivers. Although general trends may be observed in empirical observations, it is concluded that such efforts are prone to substantial uncertainty due to the complexity and variability of site conditions (that are typically unaccounted for in broad statistical approaches) as well as poor data availability and/or quality for many of the selected cases. This highlights the importance of performing site-specific investigations through numerical models, laboratory tests and field observations to better predict post-breach behavior (ideally within a probabilistic framework) when undertaking site assessments.

Chapter 4 focuses on satellite InSAR technology, which has grown in popularity due to its ability to detect millimeter-scale displacements in infrastructure along the line-of-sight (LOS). The literature on InSAR applications on tailings dams is relatively limited compared to other engineering sectors. This has led to limited understanding of (i) whether InSAR can be as accurate or representative as on-the-ground instrumentation, (ii) whether failures of tailings dams can be predicted in advance using InSAR, and (iii) what conditions and variables the quality of InSAR results and monitoring/prediction efforts depend on. To help fill this research gap, open-source Sentinel-1 data is analyzed to undertake a ground-truth assessment at a test site and a forensic investigation of 7 failure cases (2017-2019). The methodology involves the use of a cost-saving commercial software with an automated Persistent Scatterer (PS) workflow, with comparison to a proprietary software, implemented with a dual PS and Distributed Scatterer (DS) algorithm with advanced noise-filtering proficiency, for the ground-truth test site and one forensic case study.

The chapter delivers key considerations that are of practical value. Firstly, commercial InSAR software with Sentinel-1 data provides reasonable accuracy when monitoring consolidation and lower-scale deformation (< 50 mm/yr) but is ineffective in larger or accelerating deformation regimes (> 50 mm/yr), where advanced proprietary algorithms are more appropriate. Secondly, environmental conditions strongly influence the quality of Sentinel-1 PS-InSAR. The best results are in dry, bare-earth or urban terrains, whereas vegetation, water, and snow/ice cover worsen PS-InSAR data quality. Thirdly, Sentinel-1 InSAR is a useful hazard-screening tool in active mine sites with large TSFs or multiple TSFs, or in legacy mines with abandoned TSFs, as it may help guide where to undertake targeted investigations. Fourth, sub-horizontal movements in the north-south direction are generally poorly captured due to the polar orbit of SAR satellites. Lastly, not all TSF failures show clear warning signs for weeks in advance, and the revisit interval of SAR satellites prevents detection of instantaneous failure modes. Therefore, long-term monitoring programs should ideally be integrated with a combination of remote sensing methods and field instrumentation. Future avenues of research could involve case-study comparisons between medium-resolution Sentinel-1 and high-resolution TerraSAR-X data and ground-truth comparisons between different proprietary algorithms.

Chapter 5 synthesizes the main contributions of this thesis and delivers a concluding outlook on the progress of the tailings industry in reaching its aspirational goal of zero-harm. This commentary is provided in the context of the September 2022 Jagersfontein TSF disaster in South Africa, after which the Church of England investors issued an urgent call for action on (i) a global registry of all TSFs including privately owned and legacy sites and (ii) a global monitoring system that integrates satellites as well as ground sensors, where available, to monitor the “highest-risk” dams. The focus of this thesis has naturally aligned with these goals and the results have contributed to both initiatives. While the Jagersfontein incident offers a sobering reminder that much work remains to be done, a point of optimism is that thesis is just one of many ongoing research projects on tailings management and safety (e.g. liquefaction susceptibility, dam breach-runout modelling, water management, design criteria) – all of which are necessary to facilitate a safer, more sustainable mining industry.