5.1. Results and Accuracy
The largest values of apparent subsidence (up to +80 cm in the 2015–2017 DoD; Figure 5
A) seemed to be generally related to dendritic erosion patterns, small pools of water, and dead treetops. Some of the apparent subsidence might derive from true erosion, due to, e.g., precipitation and snowmelt, but it is probably related at least partly to the ability of the higher-resolution DEM for June 2017 (5 cm/pixel) to represent fine surface features more accurately than the lower-resolution DEM for 2015 (20 cm/pixel). Unfortunately, no ground truth data were available to validate the amount of tailings subsidence. However, it is noteworthy that the distribution of point-to-point distances for the presumably stable service road centered around ±0 cm change, whereas for the tailings surface the distribution shifted +30 cm towards subsidence in 2015–2017.
Considering the total thickness of the tailings deposit (up to ~10 m) and the time since the last deposition (2014), the values of apparent subsidence (up to +80 cm) seem somewhat high to be caused simply by tailings consolidation. However, the general trend in subsidence was at least in the expected direction and values closer to the distribution peak around +30 cm are more easily explained by tailings consolidation and settlement. It should also be noted that the tailings were deposited on top of a natural peat layer at least 1 m thick that can also undergo compaction. Peat soil is strongly compressible under load and it can continue to settle for a considerable time, thus causing some of the measured subsidence. The subsidence could also be affected by groundwater fluctuations and repeated freeze-thaw cycles which, depending on winter conditions, can affect at least the top ~70 cm of the tailings. Freezing and thawing are known to compact soil materials and to increase the rate of consolidation [11
]. The subsidence detected appeared to be less severe along the edges of the impoundment. This might be due to some degree to data uncertainty, as is evident in the distribution of vertical point-to-point distances for the service road (Figure 5
B). It might also be partly related to initially lower thickness of the tailings deposit, and consequently lower potential for consolidation-related settlement of the tailings and less compression of the underlying peat. An additional cause might be migration of materials from the middle of the impoundment towards the lower elevation edges, due to creep and/or, e.g., wind, rain, and snowmelt-related erosion.
Little to no observable change was expected between the June 2017 and August 2017 measuring campaigns, as the period between the surveys was short and the rate of consolidation is likely to have declined significantly since the end of tailings deposition in March 2014. Thus, the DoD for the 2017 datasets should give some indication of data reproducibility, as the gear and flight parameters used remained constant between the campaigns. With regards to uncertainty, the standard deviations for the 2017 road and tailings DoDs were 4.4 cm and 4.7 cm, respectively, which is greatly reduced from the 11.0 cm standard deviation obtained for the 2015–2017 road dataset. In a previous study, Clapuyet et al. [32
] concluded that, all parameters being equal, the reproducibility of topographical UAV-based SfM measurements is very high and that the main limiting factor with regard to accuracy is the GNSS system used in surveying the georeferencing targets. The manufacturers report the accuracy of the RTK GNSS devices used in the present study to be generally around a few centimeters. Thus, the accuracy of the final models in our case was somewhat lower than the presumed GCP accuracy.
Numerous recent studies have focused on the accuracy, reproducibility, and optimization of UAV surveys utilizing SfM in variable locations [32
] and can provide some insights. The accuracy is generally considered to be affected by various parameters, including camera parameters (e.g., focal length, sensor resolution, and exposure time), acquisition parameters (e.g., flight height and ground control network), environmental parameters (e.g., solar illumination and weather), and processing settings [40
]. Also, the characteristics of the UAV (e.g., stability, velocity) can have an influence on the accuracy. It is important to note that different validation studies focus on slightly different aspects of accuracy and precision, and also utilize different types of validation methods [29
]. Validation studies utilizing point-to-raster comparison (as was the case in this study with the 189 ground checkpoints) generally seem to provide much higher root-mean-square errors (RMSEs) than raster-to-raster or point-to-point comparison [29
]. Furthermore, based on 50 validation studies, Carrivick et al. [29
] demonstrated a clear increase in RMSE as the survey range and area increases.
Agüera-Vega et al. [33
] conducted an accuracy assessment with variable terrain morphologies, flight heights, and GCP configurations. The number of GCPs clearly influenced the horizontal accuracy, with an increased number providing improved accuracy [33
]. The vertical accuracy was not influenced by terrain morphology, while both flight height and number of GCPs had a significant influence on the RMSE of the vertical component, with lower flight heights giving higher accuracy [33
]. Clapuyet et al. [32
] reported that increasing the number of GCPs beyond a certain (area-dependent) number provided only negligible improvements in vertical accuracy. Tonkin and Midgley [35
] showed that increasing the number of GCPs beyond a certain number can even increase the vertical RMSE if the positional accuracy of the survey equipment is variable, and highlighted the importance of a uniform spatial distribution of GCPs if highly accurate data are required.
In both of our 2017 surveys, the GCPs near the middle northern portion of the service road were the least accurate points reported by Agisoft Photoscan during processing. Some of the strongest inaccuracies in the final models were also detected in this area when compared against GNSS ground checkpoints. The origin of the inaccuracies in the southeastern corner, on the other hand, was not clearly apparent in the processing report. Increasing the number of GCPs could have improved the accuracy although, as indicated by the literature, would not always do so. With further processing, one remedy could be to try different combinations of GCPs to find the best fit to the ground checkpoints and also otherwise optimize the processing parameters [38
A further practicality is the flight height, which affects the resolution, flight time, and processing requirements. In the case of a >1 km2
survey area (as in this study), a minimum flight height of 150 m is recommended in order to avoid extremely large datasets and long processing times while providing sufficient coverage [29
]. Furthermore, local UAV regulations should be considered. For example, in Finland, exceeding the flight height of 150 m with a UAV requires a permit from the Finnish Transport Safety Agency. Improved camera quality (in 2017) clearly affected the data accuracy, as shown in Figure 4
. A similar effect has been reported by, e.g., Mosbrucker et al. [39
], who found that improving the information capacity of the camera system (12 megapixel, 380 mm2
sensor replaced by 36 megapixel, 860 mm2
sensor) increased the pixel matching quality, resulting in eightfold greater point density, sixfold greater accuracy, and 50% better precision. Further regression analysis of 67 datasets and related survey metrics from 16 validation studies indicated that photographic scale is the best explanatory variable to predict the accuracy of SfM data [39
Another uncertainty in the present case was the stability of the service road, since there is no definitive certainty that the road has remained stable over the years. However, it is reasonable to assume that the road is among the most stable portions of the structure, considering that it is built on top of glacial till instead of peat and that stabilization work is typically performed in order to ensure stability under vehicle and machinery traffic. The same applies for the outer dams of the impoundment, which were utilized in surveying some of the ground checkpoints. As the road was utilized in co-registration of the datasets, it is at least possible to conclude that the tailings surface has moved relative to the road.
5.2. Application of UAVs for Monitoring Tailings Inpoundments
Tailings storage facilities can cover vast areas and thus require a comprehensive and widespread environmental monitoring system. Each impoundment is unique, tailor-made, and designed. Furthermore, the safety of the impoundment must be ensured by economically feasible means. Overall, the management of tailings impoundments needs several measurement methods for different monitoring purposes. The results from our case indicate that UAVs could be useful, fast, and cost-effective in monitoring the movement of paste tailings in the decimeter range. In the active mine phase, it is economically important to monitor the tailings storage volume in the impoundment and to optimize the disposal technique. The UAV-SfM method can be a useful tool for (i) mapping a tailings surface profile; (ii) calculating impoundment storage capacity; and (iii) predicting future needs more specifically, e.g., the need to raise surrounding dams or other maintenance work. From a geotechnical perspective, the UAV-SfM method can help to optimize the disposal scheme, e.g., it can help decide where to pump the tailings in the impoundment in order to use the storage capacity as efficiently as possible and check the tailings surface slope. The method is suitable for monitoring changes in the tailings surface, but it does not ultimately explain the reason for the changes.
The potential use of the method in management of tailings impoundments is not limited to the active phase. Tailings impoundments have to be closed safely and normally a cover is built above the impoundment to protect and minimize environmental effects. After closure of a mine, regular monitoring with UAVs could help map long-term scale settlements and movements in the tailings impoundment(s). UAVs can also be used early in the design phase for mapping potential locations for a tailings impoundment.
Another key issue related to tailings impoundments is dam safety. While dam safety is governed by different laws, safety guidelines, and industry standards, further improvements are needed to avoid the environmental disasters that still occur [2
]. Dam safety is a key issue around the world and better monitoring methods are needed. As mine tailings areas are vast, dams can be several kilometers in length, but the required accuracy in monitoring dams for possible movement is typically in the range of millimeters up to a few centimeters. Thus, the accuracy achieved in this study with the UAV-SfM method limits its potential use to monitor tailings dam safety. For remote sensing of tailings dams, it could therefore be more reasonable to focus on utilizing, e.g., satellite-based differential radar interferometry, which can detect displacements in the millimeter range, although the spatial resolution is generally much lower (meters to tens of meters) [41
Use of the UAV-SfM method in the field also has some limitations. For example, in sub-Arctic conditions, the time window for measurements is restricted. Snow cover obscures the surface visibility and the tailings surface can be covered with snow for many days per year (around six months at the study site). This means that the UAV-SfM method cannot replace other monitoring activities during wintertime. Heavy precipitation and snowmelt water cover can also prevent or limit UAV measurements. Furthermore, possible problems that can go unnoticed until data processing (e.g., inaccurate or too scarce GCPs) should be considered when working in areas to which access is restricted. One option would be to produce a low-quality model in the field and make supplementary measurements when needed. For the method to be applied in regular maintenance operations and UAV-supported monitoring by mine staff or authorities, the workflow and processing could be streamlined, e.g., in regular operation it would be practical to use fixed and accurately measured GCPs, and the flight, camera, and processing parameters could be optimized over time through experience. Thus, a greater focus on quality control and quality assurance (QA/QC) could help standardize the use of UAVs in monitoring mine environments.