Advanced analysis of satellite data reveals ground deformation precursors to the Brumadinho Tailings Dam collapse

The study investigates whether advanced satellite InSAR analysis can detect and characterise precursory deformation preceding catastrophic tailings dam failure, and whether the timing of failure could have been forecast. The context is the collapse of Dam I at the Córrego do Feijão iron ore mine, Minas Gerais, Brazil, on 25 January 2019, which released 11.7 million m3 of tailings, caused over 200 deaths, and extensive environmental damage. Despite conventional monitoring (survey monuments, inclinometers, piezometers, ground-based radar), no significant precursors were reportedly detected. The expert panel indicated failure by static liquefaction due to internal creep and loss of suction associated with wet season rainfall. Prior InSAR studies showed settlement-like deformation consistent with consolidation but may have lacked sufficient coverage and accuracy over vegetated surfaces. This study applies the ISBAS technique to Sentinel-1 SAR data from two overlapping descending tracks to overcome coherence limitations and to resolve the temporal evolution of deformation across the entire dam structure leading up to failure.

Existing InSAR techniques have been applied to tailings impoundments, revealing substantial displacement rates, including up to ~25 cm/year at the rear of Dam I’s tailings beach and subtler rates (on the order of tens of mm/year) nearer the centre consistent with consolidation settlement. However, conventional Permanent Scatterer and standard SBAS approaches often suffer from poor coherence over bare soils and vegetated areas, limiting measurement density and accuracy and potentially masking precursory accelerations. Multi-temporal stacking can mitigate atmospheric effects, but coverage remains a challenge. ISBAS, a coherent scatterer approach adapted from SBAS, has been shown to improve coverage and millimetre-level accuracy over vegetated terrains, providing a promising method for comprehensive dam monitoring.

Data: Sentinel-1 C-band SAR Interferometric Wide (IW) mode imagery from two overlapping descending tracks (53 and 155) covering August 2017–January 2019 for Brumadinho Dam I. Track 53: 45 images (12 Aug 2017–22 Jan 2019), incidence angle 32°. Track 155: 45 images (7 Aug 2017–17 Jan 2019), incidence angle 45°. For the Sul Superior site, track 155 provided 59 images (8 Jun 2017–10 Jun 2019). Processing: ISBAS InSAR processing using Terra Motion Limited’s Punnet software. Co-registration of SLC data; differential interferogram generation with temporal and orbital baseline limits (≤1 year, ≤100 m). Phase unwrapping via an in-house implementation of SNAPHU. Pixels required average coherence ≥0.45. Spatial resolution 20 m with multilook factor 7×2 pixels. Outputs geocoded to UTM using SRTM DEM (noting minor geolocation inconsistencies due to DEM age). No smoothing applied during processing; displayed maps smoothed with a 3×3 averaging filter for visualization only. Uncertainty: Number of best-coherence interferograms n_v ≥430 (track 53) and ≥498 (track 155), yielding LOS velocity standard error <0.53 mm/year. For Sul Superior, n_v = 720 (max standard error ~0.41 mm/year). Standard error of velocity ev ≈ 11/√(n_v) mm/year. Comparative analyses: Conventional SBAS processed for comparison (only pixels coherent throughout entire stack). A DInSAR consecutive-pair stacking technique was also implemented to help identify faster deformation. Stereo analysis: Same-side stereo using overlapping descending tracks 53 and 155 to resolve components into up-down (vertical) and lateral towards-away (across-track) using incidence-angle-based equations. Due to small angular difference (~13°), component precision is 2–3× worse than LOS; results mainly indicative of vertical dominance of motion. Time-series extraction: The dam was partitioned into structural sectors (dam wall, front/rear tailings beach). Locations exhibiting distinctive deformation were identified, and representative time-series computed by averaging 25–208 contiguous pixels per location. Uncertainty for each averaged time-series combined the standard deviation of pixel time-series with the maximum ISBAS standard error. Rainfall data: Daily rainfall totals from nearest INMET automatic stations (Ibirité Rola Moça-A555 ~15 km for Brumadinho; Belo Horizonte Cercadinho-F501 ~38 km for Sul Superior) were compiled to compare with deformation. Failure timing prediction: Inverse velocity method applied to smoothed velocity time-series calculated from 12-day displacement intervals for locations showing accelerated deformation (locations 3, 4, 5, 6). Linear regression of inverse velocity versus time was used to predict failure timing, selecting consecutive points with the highest R2 (4–5 points), discarding anomalous points affected by noise/phase wrapping. Sensitivity analysis iteratively assessed how early reliable predictions could be made as successive observations accrued.

- ISBAS provided near-complete displacement coverage across the dam structure (99.5%) compared to only 3.3% using conventional SBAS, enabling comprehensive spatial assessment. - Average LOS velocities over Aug 2017–Jan 2019 indicated widespread deformation across the tailings beach. Track 53 showed magnitudes typically on the order of several mm/year (up to ~8 mm/year downward/away from satellite) immediately behind the crest; track 155 showed similar spatial patterns with slightly lower magnitudes due to larger incidence angle. - Time-series revealed nonlinear deformation behaviour: after ~10 months of slow settlement, accelerated displacement commenced from mid-October to December 2018, correlating with increased rainfall. Cumulative LOS displacement reached up to ~−15 mm at the back-centre of the beach (location 3) by 17 Jan 2019, and up to ~−18 mm at the front of the beach (location 5) by 22 Jan 2019. - The dam wall centre (location 1) showed no significant sustained deformation, only minor seasonal fluctuations (~±2 mm) in phase with the rainy season, likely elastic response to groundwater/pore pressure variations corroborated by piezometer data. - Two arms of deformation were observed extending down the front of the dam wall, coincident with the eventual failure extent. The right toe (location 4) exhibited subtle deformation starting mid-October 2018 (following rainfall) and accelerated into January 2019, about two weeks before collapse. - The late-2018 accelerations at multiple locations (3, 4, 5, 6) are inconsistent with consolidation settlement (which should decelerate over time). Instead, they suggest wetting-induced suction reduction, decreased soil strength, internal strains/creep, and susceptibility to static liquefaction. - Stereo analysis indicated the deformation on the dam wall and tailings was predominantly vertical, consistent with the expert panel findings. - Inverse velocity analysis correctly forecast the failure timing interval at all four analysed locations. Predicted intervals: location 3 (days 513–537 since 19 Aug 2017), location 4 (512–537), location 5 (515–527), location 6 (523–551). The actual collapse occurred at day 524, within all intervals. - Sensitivity analysis showed earliest reliable predictions: as early as 51 days before collapse at location 4 (from 5 Dec 2018). Three overlapping failure time-interval predictions across different locations could have been available 44 days before collapse, offering ~40 days of actionable notice before the start of the predicted interval, even considering ~1 day ISBAS processing per new acquisition. - Case comparison at Sul Superior dam (Gongo Soco): detected downslope motion on pit walls (<15 mm/year) with reactivation in Dec 2018 and continued into Jun 2019; dam and tailings exhibited uniform, decelerating settlement consistent with consolidation (no anomalous pre-failure behaviour). A spoil heap also settled (up to ~15 mm/year). No evidence of degrading stability conditions at the dam over Jun 2017–Jun 2019.

The findings demonstrate that advanced ISBAS InSAR can detect subtle, spatially distributed precursory deformations over vegetated tailings dams that conventional techniques may miss. The observed late-2018 accelerations, correlated with rainfall, provide evidence of wetting-induced suction loss and internal creep that likely precipitated static liquefaction, directly addressing the research objective of identifying pre-failure signals. The successful inverse velocity forecasts across multiple locations show that, when implemented as part of an early warning framework, satellite InSAR time-series can provide actionable lead time (weeks) for intervention. The variation in predicted intervals between locations underscores the need to monitor the entire dam structure and to employ multiple SAR viewing geometries where possible. The comparison with Sul Superior highlights that uniform, decelerating trends are indicative of benign consolidation, whereas spatially differential and accelerating trends signal potential instability. Operationally, integrating ISBAS InSAR with in situ geotechnical monitoring can enhance reliability, given InSAR’s limitations for very rapid failures and coherence constraints. Even with modest processing latency, systematic monitoring and inverse velocity analyses could have flagged risk and supported escalation of ground-based inspections or mitigation measures well before failure.

The study shows that ISBAS processing of Sentinel-1 data revealed precursory, non-consolidation deformation across Dam I’s wall and tailings beach prior to the 25 January 2019 collapse. Accelerated movements began in late October 2018 following increased rainfall, and inverse velocity analyses accurately forecast the failure time window, in some cases over a month in advance. The results emphasize the importance of monitoring the entire tailings dam and leveraging multiple SAR geometries. Application to the Sul Superior dam did not reveal anomalous pre-failure behaviour, instead showing uniform consolidation settlement. The work underscores satellite InSAR’s value as part of integrated, cost-effective dam monitoring and early-warning systems that require no ground infrastructure. Future research should focus on improving component decomposition with enhanced viewing geometries, automating real-time inverse velocity alerts, integrating multi-sensor data (e.g., GNSS, piezometers, ground-based radar), and addressing coherence challenges due to water and vegetation.

- Only descending same-side Sentinel-1 tracks were available, with small incidence-angle separation (~13°), limiting stereo decomposition accuracy and the ability to resolve horizontal components. - InSAR is restricted in capturing very rapid or sudden failures due to potential loss of coherence, phase ambiguity, and satellite revisit intervals. - Coherence can be degraded by water coverage and vegetation, potentially limiting measurement density; although ISBAS mitigates this, it does not remove the limitation entirely. - Geolocation inaccuracies may arise from using the SRTM (2000) DEM, causing small inconsistencies in mapped features (e.g., lake outlines). - Rainfall data were obtained from stations 15–38 km away, which may not perfectly represent site-specific precipitation. - Lack of direct access to ground monitoring datasets limited cross-validation beyond published summaries. - Inverse velocity forecasts depend on the quality and number of late-stage observations; sensitivity varied by location and track, and some points required exclusion due to noise or phase wrapping.