Int J Appl Earth Obs Geoinformation 90 (2020) 102119

Contents lists available at ScienceDirect

Int J Appl Earth Obs Geoinformation

journal homepage: www.elsevier.com/locate/jag

The 2019 Brumadinho tailings dam collapse: Possible cause and impacts of T
the worst human and environmental disaster in Brazil
Luiz Henrique Silva Rottaa,*, Enner Alcântarab, Edward Parkc, Rogério Galante Negrib,
Yunung Nina Lind, Nariane Bernardoa, Tatiana Sussel Gonçalves Mendesb,
Carlos Roberto Souza Filhoe
a
São Paulo State University – Unesp, Department of Cartography, Brazil
b
São Paulo State University – Unesp, Department of Environmental Engineering, Brazil
c
National Institute of Education and Asian School of the Environment, Nanyang Technological University, Singapore
d
Institute of Earth Sciences, Academia Sinica, Taiwan
e
University of Campinas – Unicamp, Institute of Geosciences, Brazil

A R T I C LE I N FO A B S T R A C T

Keywords: On 25th January 2019, the tailings dam of the Brumadinho iron mine operated by Vale S/A failed catastro-
Dam collapse phically. The death toll stood at 259 and 11 people remained missing as of January 2020. This tragedy occurred
Mining three years after Mariana’s tailings dam rupture – the most significant tailing dam disaster in Brazilian history.
Contaminated mud Thus far, a systematic investigation on the cause and effect of the failure has yet to be conducted. Here, we use
Suspended particulate matter
satellite-driven soil moisture index, multispectral high-resolution imagery and Interferometric Synthetic
Remote sensing
Aperture Radar (InSAR) products to assess pre-disaster scenarios and the direct causes of the tailings dam col-
lapse. A decreasing trend in the moisture content at the surface and the full evanescence of pond water through
time (2011–2019) suggest that the water was gradually penetrating the fill downwards and caused the seepage
erosion, saturating the tailings dam. Large-scale slumping of the dam (extensional failure) upon the rupture
indicates that the materials of the fill were already saturated. InSAR measurements reveal a dramatic, up to
30 cm subsidence in the dam (at the rear part) within the past 12 months before the dam collapse, signifying that
the sediments had been removed from the fill. Although the information on the resistance level of the tailings
dam to infiltrations is not available, these pieces of evidence collectively indicate that the seepage erosion
(piping) is the primary cause for the chronic weakening of the structure and, hence, the internal “liquefaction”
condition. Upon the collapse, the fully saturated mud tailings flowed down the gentle slope area
(3.13 × 106 m2), where 73 % were originally covered by tree, grass or agricultural tracts. The toxic mud
eventually reached the Paraopeba River after travelling 10 km, abruptly increasing the suspended particulate
matter (SPM) concentration and the toxic chemical elements in the river, immediately affecting the local live-
lihoods that depend on its water. The Paraopeba River is a major tributary of the San Francisco River, the
second-longest river in Brazil reaching the Atlantic Ocean. We anticipate that the environmental repercussions of
this toxic seepage will be felt throughout the entire basin, especially riverine communities located downstream.

1. Introduction several local communities.Subsequently, the contaminants in the Doce

River were delivered to the Atlantic Ocean, after travelling around
A Brazilian mining disaster, also known as the Samarco disaster, 665 km (Aires et al., 2018), seriously affecting coastal environments
occurred on 5th November 2015 at the Fundão dam located around and marine ecosystems (Carmo et al., 2017; Coimbra et al., 2019).
Mariana city, Minas Gerais State. The mud wave produced by the dam Besides the extensive and serious water pollution, the disaster also
failure, estimated at 43 million m³, spread over several kilometers down killed 20 people, caused mass death of fishes, flooded almost 7000 km²
the slope, reaching the the Doce River, the second-longest river in Brazil that covered the Bento Rodrigues district, and adversely affected the
(Garcia et al., 2007), turning it into a “sea” of red mud (Hatje et al., Abrolhos Marine National Park (Francini-Filho et al., 2019). Economic
2017). This river served as a primary source of water and fish for losses were estimated at around US$ 521 million per year (Garcia et al.,

⁎
Corresponding author at: São Paulo State University, Department of Cartography, 19060-900, Roberto Simonsen 305, Presidente Prudente, SP, Brazil.
E-mail address: [email protected] (L.H. Silva Rotta).

https://doi.org/10.1016/j.jag.2020.102119
Received 8 January 2020; Received in revised form 26 March 2020; Accepted 26 March 2020
Available online 10 April 2020
1569-8432/ © 2020 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/BY-NC-ND/4.0/).
L.H. Silva Rotta, et al. Int J Appl Earth Obs Geoinformation 90 (2020) 102119

2007), and rigorous and vigilant protocols were proposed to assess affected by the mudflow (Fig. 1). The mud traveled approximately
other dams in a similar condition in the country aiming to prevent 10 km, reaching the Paraopeba River. The mud covered an area of
history from repeating. 3.13 × 106 m2, equivalent to 450 football fields. The collapsed dam
After 3 years of the Mariana tragedy, on 25th January 2019, another (Dam I - Córrego do Feijão Mine) had an area of 4.13 × 105 m2 and
mining dam consisting of iron ore tailings failed catastrophically in the approximately 600 m width.
Brumadinho city, Minas Gerais State, Brazil. The rupture produced
more than 11 × 106 m3 of mining waste – equivalent to a 10-m high 3. Methods
wave of mud that spread 10 km the downhill – that reached the
Paraopeba River; a major tributary of the São Francisco River. Although 3.1. Soil Moisture Index (SMI)
the volume of mud derived from the Brumadinho disaster is less than
the case of Mariana, the event caused far more significant loss of lives. The Soil Moisture Index (SMI) over the tailings dam (Lambin and
As of January 2020, the death toll hit 259 people, while 11 individuals Ehrlich, 1996; Zhan et al., 2004) was used to reconstruct the dynamics
remain missing. of the ponding over the dam since 2011. This information is important
Both collapsed dams belong to Vale S/A, which owns 133 iron ore in assessing the cause of the failure because the dam was decommis-
dams in Brazil, of which 80 % of them (105) are in the Minas Gerais sioned in 2015. Therefore, the pond over the dam is expected to dis-
State. Vale uses both downstream (conventional model) and upstream appear over time. Landsat 5, 7 and 8 images between 2011 and 2019
tailings dams. The upstream tailing dam, the model used both in were used to calculate the SMI. This index is based on the Normalized
Brumadinho and Mariana, is constructed by vertically accumulating the Difference Vegetation Index (NDVI) and Land Surface Temperature
tailings through successive uphill deposition (Kossoff et al., 2014). This (LST), yielding values between 0 (drier soil) and 1 (wet soil), which is
dam filling method is cheaper since only moderate compaction of a calculated as (Zhan et al., 2004):
smaller amount of material is required. The license for such dam pro-
LSTmax − LST
jects are also easily obtained since upstream tailing dams occupies a SMI =
smaller area relative to the downstream tailing dams. This upstream LSTmax − LSTmin (1)
tailing dam model, however, is considered dangerous and risky (Kossoff where, LSTmax and LSTmin are, respectively, the maximum and
et al., 2014), becausethe stability of the tailings impoundments de- minimum values of LST within the image for a given NDVI, expressed
creases substantially once operations cease (Martin and McRoberts, as:
1999). Generally, the disposal management in such a dam aims to form
a uniform deposit, whereby the coarser solids are deposited adjacent to
LSTmax = a1 × NDVI + b1 (2)
the embankment. A supernatant pond is formed by managing the beach LSTmin = a2 × NDVI + b2 (3)
to ensure the run-off, enabling decanting and re-use. The dam stability,
therefore, relies on the earthfill/tailings strength (used to construct where, a and b are empirical parameters defining the dry and wet edges
each bund), the characteristics of the tailings and the maintenance of modeled as a linear fit to the data (Zhan et al., 2004; Parida et al., 2008;
the tailings pond at a safer distance from the structure zone (Cambridge Potić et al., 2017).
and Shaw, 2019). Since 2015, when the Brumadinho dam was deacti- According to Lambin and Ehrlich (1996), the scatterplot of LST vs
vated, it received several positive assessments by the local consultants NDVI results in a trapezoidal shape, and all types of land cover fall
on dam safety and seepage/stability (Porsani et al., 2019). However, within the trapezoid of the LST-NDVI space. The upper envelope of the
despite systematic monitoring, it failed three years after closure, con- trapezoid (upper limit of surface temperature for a given vegetation
flicting with the notion that geotechnical structures become more stable cover) represents the dry condition (warm edge), while the lower limit
over time (Santamarina et al., 2019). represents the wet condition (cold edge) (Parida et al., 2008).
Here, we investigate possible causes of the dam failure and discuss A time series of SMIs was retrieved from TM/Landsat-5, ETM
the environmental consequences of the tragedy. We attempt to under- +/Landsat-7 and OLI/Landsat-8 images (Table 1). The images used are
stand the changes in moisture contents at the surface of the Brumadinho surface reflectance products (L3) at 30 m resolution, which were
tailings dam and hypothesize that the seepage erosion followed by the downloaded from the USGS Earth Explorer site (https://earthexplorer.
liquefaction and gradual removal of the fill have led to the failure of the usgs.gov).
dam. The main environmental impacts caused by the mudflow were We generated a standard deviation map of SMI over time to analyze
also investigated using available remote sensing data. Specifically, we the variation of moisture conditions over the dam. Upon the dam
aim (i) to investigate the possible cause of the dam rupture using var- rupture, the mud flowed into the Paraopeba River, dramatically af-
ious available remote sensing datasets, (ii) to map the impact of mud- fecting the water quality. Remote sensing images were also used to
flows on land use and land cover changes (LULCC), (iii) to compare the compare the SPM concentrations before and after the rupture.
suspended particulate matter (SPM) in the Paraopeba River before and
after the dam rupture, and (iv) to assess the geochemical impact on the 3.2. Interferometric synthetic aperture radar
Paraopeba River.
We obtained 33 scenes from Sentinel-1 Synthetic Aperture Radar
2. Site description (SAR) descending path 53 between Jan 3rd 2018 and Jan 22nd 2019.
Given the regular acquisition frequency of every 12 days, we adopted
The complex of iron dams is located at upstream of the Brumadinho the nearest-2 paring strategy to produce 63 SAR interferograms. The
city in Minas Gerais State, southeastern Brazil (Fig. 1a). The watershed processing was done using NASA JPL’s opensource software ISCE2
of the Paraopeba River (WPR) is located at the center of Minas Gerais (https://github.com/isce-framework/isce2). All interferograms are
State, with the area of 13,640 km2, draining directly into Três Marias geocoded at 15 × 15 m ground pixel spacing using an oversampled
hydroelectric reservoir (Durães et al., 2016). The Paraopeba River has SRTMv3 DEM (Farr et al., 2007) by a factor of two. To compute the
large-scale sand-extraction industrial operations and is located near surface line-of-sight (LOS) displacement time-series from the pairwise
highly populated areas with extensive iron mining activities (Sato and interferograms, we choose the small baseline subset approach
Godinho, 2003). Fig. 1 shows the location of the collapsed dam, as well (Berardino et al., 2002) in the Generic InSAR Analysis Toolbox (GIAnT;
as high-resolution images acquired by Planet constellation satellites Agram et al., 2013). We incorporated simple tropospheric correction by
before and after the dam rupture. removing the topographically-correlated phase (Lin et al., 2010). The
An image acquired on January 29th 2019, shows the area directly standard deviation of the LOS displacement time-series was derived

2
L.H. Silva Rotta, et al. Int J Appl Earth Obs Geoinformation 90 (2020) 102119

Fig. 1. Study area: (a) Minas Gerais State in Brazil; (b) Paraopeba River and cities near to the accident; (c) before and (d) after images of the dam collapse, indicating
the dam location and the area directly affected by the mud wave.

Table 1 through a regular bootstrapping approach.

Summary of the TM/Landsat-5, ETM+/Landsat 7 and OLI/Landsat 8 images
used in this study. 3.3. Land use and land cover mapping
Date Sensor Date Sensor Date Sensor
We delineated the affected area by using Planet images acquired
2011-02-09 TM L5 2013-10-28 OLI L8 2016-07-16 OLI L8 before (January 22nd, 2019) and after (January 27th 2019) the dam
2011-04-14 TM L5 2014-04-06 OLI L8 2016-08-17 OLI L8
rupture (Fig. 1). The Planet constellation consists of more than 130
2011-04-30 TM L5 2014-05-24 OLI L8 2016-09-18 OLI L8
2011-06-17 TM L5 2014-06-09 OLI L8 2016-12-07 OLI L8 orbital Earth observation satellites acquiring daily images of the Earth
2011-07-19 TM L5 2014-08-12 OLI L8 2016-12-23 OLI L8 on four bands in the visible and near infrared wavelength range at 3 m
2011-08-04 TM L5 2014-09-29 OLI L8 2017-01-08 OLI L8 spatial resolution (www.planet.com). The affected area was auto-
2011-09-05 TM L5 2014-11-16 OLI L8 2017-03-13 OLI L8
matically vectorized by differencing the Normalized Difference Vege-
2011-09-21 TM L5 2015-01-03 OLI L8 2017-07-03 OLI L8
2011-10-07 TM L5 2015-05-27 OLI L8 2017-09-05 OLI L8
tation Index (NDVI) from images before and after the dam collapse.
2012-09-15 ETM + L7 2015-06-12 OLI L8 2017-10-07 OLI L8 Using the NDVI difference image, the area with more significant land
2013-02-22 ETM + L7 2015-07-30 OLI L8 2018-04-17 OLI L8 cover changes was easily identified, which is the area covered by the
2013-04-27 ETM + L7 2015-08-31 OLI L8 2018-05-03 OLI L8 mud tailings (Fig. 1d).
2013-07-16 ETM + L7 2015-09-16 OLI L8 2018-06-04 OLI L8
We mapped the Land Use and Land Cover Changes (LULCC) of the
2013-08-01 ETM + L7 2015-10-02 OLI L8 2018-07-06 OLI L8
2013-11-21 ETM + L7 2016-03-10 OLI L8 2018-09-08 OLI L8 area using a Support Vector Machine (SVM) algorithm (Mountrakis
2013-07-08 OLI L8 2016-04-11 OLI L8 2018-12-13 OLI L8 et al., 2011). The Planet image acquired on January 22nd 2019 (before
2013-08-25 OLI L8 2016-06-14 OLI L8 2019-01-14 OLI L8 the collapse) was selected for the classification. The SVM approach
yields robust results whilst requiring limited training data. It also has a

3
L.H. Silva Rotta, et al. Int J Appl Earth Obs Geoinformation 90 (2020) 102119

Fig. 2. Planet satellite image form 24 January 2019 (a) and SMI Standard Deviation map from 2011 to 2019 of the Dam I – Córrego do Feijão Mine (b). Three specific
sectors of the dam are considered for analysis: Slope (i), Top-of-bench-1 (ii) and Top-of-bench-2 (iii).

high generalization capability in comparison to other conventional not sensitive enough to map the high level of SPM concentrations in the
classification methods (Mountrakis et al., 2011). Paraopeba River after the dam collapse. The SPM concentration re-
Classes of “tree vegetation”, “grass/agriculture”, “water body” and trieval model used here is calculated by the following equation:
“bare soil” were considered and respective training and test samples
A × ρw
were collected through visual inspection of the images. To evaluate the SPM = ρw +B
1− (4)
classification results, a confusion matrix between the classified image C
and the ground truth data was generated to calculate the Kappa coef- where ρw is the water reflectance; c is established according to Nechad
ficient. Regions of interest (ROIs) of the four classes were created and et al. (2010) and the central wavelength (865 nm = 0.2115); B denotes
used in classification and validation. The “building” class was created measurements errors in reflectance. The parameter A (mg/l) provides
manually because of the high confusion among the other classes. the relationship between SPM and reflectance. We used an A value of
Furthermore, the “Dam I – Córrego do Feijão Mine” class was inserted 289.29 mg/l, adapted from Nechad et al. (2010).
in the LULCC map to identify its location in the area affected by the The SPM was calculated along 60 km in the Paraopeba River
mud tailings. The SVM parameters were empirically tuned, yielding a downstream from the initial point of pollution in the water body, i.e.
penalty value of 100 and RBF kernel function with gamma parameter of from Brumadinho to Juatuba (Fig. 1b). The Paraopeba River was di-
0.25. Furthermore, the one-against-all multiclass strategy was em- vided into twelve regions (5 km each) to show the longitudinal dis-
ployed. tribution of SPM using a boxplot for each date. We also employed SPM
In addition to the LULCC mapping, we calculated slope and terrain median values to compare the pre- and post-collapse SPM conditions in
roughness index (TRI) (Riley et al., 1999) of the impacted area using the river.
the SRTM DEM (30 m). TRI is defined as the mean difference between a
central pixel and its eight surrounding cells.
4. Results and discussion

3.4. Suspended Particulate Matter (SPM) mapping 4.1. Moisture in the tailing dam

We used Sentinel 2A images acquired by the Multispectral The SMI indicates the moisture in the study area and was calculated
Instrument (MSI/Sentinel-2) to estimate the Suspended Particulate for 51 Landsat-5, 7 and 8 images between 2011 and 2019. The SMI
Matter (SPM) over the Paraopeba River. With a 10 m resolution at standard deviation map is shown in Fig. 2.
visible and near infrared bands, and 12-bit radiometric resolution, the Based on the SMI standard deviation map of the dam since 2011,
MSI image is suitable to study small-scale inland water, i.e. water different spatial variation patterns are highlighted that is used to di-
bodies with a diameter less than 30 m (Toming et al., 2016). Level 2A vided the dam into three sections: (i) Slope sector, which hosts the
images (i.e., bottom-of-atmosphere corrected reflectance product) were downstream slope of the dam (face); (ii) Top-of-bench-1, which is the
downloaded from the Sentinel’s Scientific Data Hub (http://scihub. area closest to the slope; (iii) Top-of-bench-2, which is sector more
copernicus.eu/) (tile 23KNT in Sentinel-2 mission system). The first distant from the slope that showed relatively higher SMI standard de-
image, acquired on January 22nd 2019, was used to estimate the water viation within the dam. The observed higher moisture in the Top-of-
quality of the river before the dam failure. The second image, acquired bench-2 (iii) was also highlighted by Robertson et al. (2020); these
on February 2nd 2019 (after the dam collapse), was used to assess the authors showed that the water accumulation near the dam crest allows
impact of the mudflows into the Paraopeba River. the deposition of weak tailings near the dam crest.
The images were processed in SNAP (Sentinel Application Platform, The SMI for twenty images from February 9th 2011 to January 14th
https://step.esa.int/main/toolboxes/snap/), applying the semi-analy- 2019 is presented in Fig. 3.
tical SPM model from Nechad et al. (2010) (Eq. (4)). We used the near- Large tracts with a higher level of moisture were observed in the
infrared band (B8 - 845 nm) since the red band located at 665 nm was rear part of the dam (Fig. 3(iii) - Top-of-bench-2) until 2016. At the

4
L.H. Silva Rotta, et al. Int J Appl Earth Obs Geoinformation 90 (2020) 102119

Fig. 3. Soil Moisture Index (SMI) from 2011 to 2019 based on sensors/satellites TM/Landsat-5, ETM+/Landsat7 and OLI/Landsat-8. The white patches represent
areas in the images influenced by clouds (i.e., unusable data).

Top-of-bench-2 (iii), a significant water pond was then formed. The SMI Measuring Mission (TRMM) in Giovanni (giovanni.gsfc.nasa.gov) was
for 5th September 2017; 3rd May 2018; 4th June 2018; and 8th Sep- plotted in the same diagram.
tember 2018 showed a smaller region with a higher moisture value In general, higher SMI values were observed over the Top-of-bench-
(including a particular blue “bullseye” in the 5th September 2017 and 2 (iii) sector of the dam for the investigated period (Fig. 4). The ob-
3rd May 2018 images), which may indicate a site with external water served high moisture contents might have implied that this sector was
entry to the dam that might have contributed to increase the water saturated with water. Another behavior observed is the decrease of SMI
content over the structure. The SMI in the dam by 14th January 2019 throughout the dam during the same period. Between 2011 and 2017,
(11 days before the collapse of the Brumadinho dam) showed that Top- Slope (i) and Top-of-bench-1 (ii) showed similar behavior, with con-
of-bench-2 (iii) was the sector with the highest moisture at that date, sistently lower values compared to Top-of-bench-2 (iii). After 2017,
whereas Top-of-bench-1 (ii) and Slope (i) displayed similar SMI values. Slope (i) and Top-of-bench-2 (iii) presented similar SMI values. In the
However, pixels with low SMI values were observed in Top-of-bench-1 rainy season, comparable values were also noted in the three regions.
(ii), near the transition between the slope and the flat top surface. However, in the dry period, the difference between Top-of-bench-1 (ii)
Fig. 4 shows the SMI for the sectors indicated in Fig. 2 (Slope (i), and the other regions increases, i.e., the SMI of the Top-of-bench-1 (ii)
Top-of-bench-1 (ii) and Top-of-bench-2 (iii)) for all the images analyzed decreased significantly in comparison to Top-of-bench-2 (iii) and Slope
between 2011 and 2019. The images used to calculate SMI were evenly (i). We speculate that the increase in the moisture pattern between the
distributed through time, allowing an biased-free investigation of pos- top and the slope sectors, between 2017 and 2019 (i.e. after the dam
sible trends. Precipitation data acquired from the Tropical Rainfall was decommissioned in 2015), implied that the fill was becoming

5
L.H. Silva Rotta, et al. Int J Appl Earth Obs Geoinformation 90 (2020) 102119

Fig. 4. SMI mean value between 2011 and 2019 for the sectors Slope(i), Top-of-bench-1(ii) and Top-of-bench-2(iii) in the Brumadinho tailings dam. Precipitation
data from TRMM/Giovanni (giovanni.gsfc.nasa.gov).

water-saturated until it collapsed. dam failed. This creeping might have been a result of on-going de-
formation within the dam and contributed to the final failure process.
4.2. Deformation in the tailings dam
4.3. Driving forces responsible for the dam failure
The sector of the dam with significant soil moisture had also sunk
over the past few years (Top-of-bench-2, Fig. 4). We also observed the A time-lapse of high-resolution images over the tailings dam since
largest LOS displacement in the region, of up to −25 cm between 2011 (Fig. 6) revealed the dynamics of the drainage at the rear part of
January 2018 and January 2019 (point B in Fig. 5). As there are no the dam, where ponds were formed (Top-of-bench-1 (ii)). These small
concurrent ground-based displacement measurements available, there drainages had gradually incised the surface and the ponds seemed to be
might remain a slight uncertainty in relation to the subsidence and the resultant depression from both the channel gullying and the gradual
westward motion, given the sensing direction of the descending path. subsidence of the fill (as evidenced in Fig. 5). The accumulated water
However, since Top-of-bench-2 is topographically lower than Top-of- had gradually infiltrated through the fill, causing seepage erosion that
bench-1 according to the high-resolution ALOS2 DEM (Fig. 7c), we led to the removal of material, which is often referred to as “piping”.
believe that westward motion is less likely and that the signal was likely Both SMI time-series and the time-lapse high-resolution images
to be causedby vertical subsidence. After accounting for the average suggest that despite the high precipitation in this area (Fig. 4), the
incident angle of 32.5 degrees, the actual cumulative vertical dis- volume of water in the pond gradually decreased throughout the years.
placement was estimated to be approximately -30 cm. Even if we assume that the precipitation was fully balanced by evapo-
Besides the deformation within the tailings dam, we also observed transpiration and there was no net meteoric water gain (net loss is
creeping in the mountainous area abutting the dam (point A in Fig. 5) unlikely because the vegetation is dense there), it is estimated that
which began around August 2018 and continued up to the point the there was still ∼10,000 m3 of water transferred out of the dam between

Fig. 5. Line-of-sight (LOS) displacement time-series from Sentinel-1 SAR interferograms. The LOS displacement map is for the total cumulated displacement between
January 3rd 2018 and January 22nd 2019.

6
L.H. Silva Rotta, et al. Int J Appl Earth Obs Geoinformation 90 (2020) 102119

Fig. 6. Time-lapse of true-colour images, from Google Earth (2011-2018), showing the formation of drainage towards the dam slope and emplacement of ponds at the
planar surface at the top of the fill. This indicates that the water table gradient was reaching the crest of the dam several years before the collapse. Algae blooms
throughout the years also indicate that the stagnant condition of the pond has been persistent over time.

2011 and early 2018. Moreover, comparison of the InSAR displacement seepage. Upon the dam rupture, it initially seemed like that the failing
time-series with the optical images showed that displacement continued started at the rear part of the surface (around Top 2(iii)), where the
even after dried up completely around mid-July 2018 (Fig. 5). This high pond was formed close to the mountain area abutting the dam (point A
subsidence rate is even higher than that caused by groundwater in Figs. 5 and 7b). This failure through instantaneous “slumping” of the
pumping - a manmade, effective and directional process that usually fill first happened in this area and was immediately followed by the
reaches less than 10 cm/yr (e.g., see Ojha et al., 2018 and Riel et al., weakening of the surrounding materials that were saturated with water
2018 for studies in California). (Fig. 7b).
Natural process like regular sediment compaction typically ranges The initial collapse at the rear part of the dam, where it experienced
between a few mm/yr (deltaic plains, Meckel et al., 2007) to a few cm/ dramatic subsidence during 2018 (Fig. 5), triggered similar dis-
yr (reclaimed lands, Erten and Rossi, 2019; Sun et al., 2018). We, turbances and large scale rotational slumping over the other parts of the
therefore, suggest that the mechanism to effectively remove water from fill. The mapping of the boundary of the slumping area through high-
the pond and materials from the fill is through fast seepage and seepage resolution photographs which were taken immediately before and after
erosion (piping), which was the major cause of dam failure. The sub- the collapse indicates that rotational slumping has occurred on the
sidence in the Top-of-bench-2 region resulted from the compaction of surface of the dam (Top-of-bench-1 (ii) and 2 (iii) zones) where the
the void space, which happened after water was permanently removed shear surface is circular (Fig. 7b, d and e). This implies that the water
from the pore space, together with potential material removal with the table was abnormally higher and perhaps reached close to the surface

7
L.H. Silva Rotta, et al. Int J Appl Earth Obs Geoinformation 90 (2020) 102119

Fig. 7. (a) Screenshot from the movie (https://www.youtube.com/ watch?v = sKZUZQytads), immediately before the dam failure on January 25th, 2019. (b) Large
scale slumping upon the dam failure. This indicates that the bulk of the fill was water-saturated before failure. (c) An elevation profile over the dam surface based on
ALOS DEM acquired between 2008-2011. A depression with a depth around 4 m is shown towards the rear part of the fill, where the water pond is present during the
wet seasons (Fig. 6). (d) and (e) High-resolution Planet images prior (Jan 18, 2019) and after the dam failure (Jan 29, 2019).

within both zones. Therefore, we consider that almost the entire dam the dam failure. Lack of sufficient compaction afterwards also should
was fully saturated when it failed, indicating that large scale “lique- have increased the risk of failure.
faction” was already happening across the fill before it collapsed, as
also indicated by the SMI mapping (Fig. 4).
4.4. Land use/cover in affected area by mud
According to Smith (2002), liquefaction is a phenomenon of de-
formation and collapse that could be caused by a triggering event and
The dam rupture directly impacted the ecosystem and human health
collapsible materials (e.g. tailings and water). Upon liquefaction, solid
mainly due to the toxic composition of the mud and the velocity of the
materials lose their mechanical resistance and present fluidal behavior
mudflow that was released and engulfed habited areas, with no chances
(Morgenstern et al., 2016). In case of an earth-fill dam, liquefaction can
of survival during the event. Remote Sensing images provide historical
also be triggered by constant rainfall or irrigations, through infiltration
records, such as pre- and post-disaster scenes. Mapping the differences
processes. After the collapse, the materials were flowing very fast (looks
between the two sets of images, usually carried through change de-
even as a turbulent flow) on a low gradient area (Fig. 8c, also evident
tection techniques, helps the analysis of LULCC and, consequently of-
from the video link in Fig. 7 caption), because they were in super-sa-
fering information regarding the property’s location and assessment of
turated condition, pushed downstream by high pore-pressure generated
affected areas. The purpose of LULCC mapping in the mud-wave
by the large scale slumping in the rear part of the dam.
flooded areas extends beyond the identification of the changes, but also
Although we consider that water-saturated fill triggered the failure,
allows us to quantify environmental losses and impacts, as well as to
more information – such as particle size distribution – which can only
provide support for management and planning actions. Fig. 8a shows
be obstained in the the field, is necessary to confirm the direct cause of
the LULC map of the region before the dam failure, within the extent of

8
L.H. Silva Rotta, et al. Int J Appl Earth Obs Geoinformation 90 (2020) 102119

Fig. 8. The LULC map (a), Elevation (b), Slope (c) and Terrain Roughness Index (TRI) (d) of the area affected by the rupture of the tailings dam.

the mud-wave. mudflows reached the river.

The class “Tree Vegetation” was the most affected by the mud, The first targets hit by the mudflow were the mining workers lo-
comprising 49 % of the analyzed area. This implies a huge loss of the cated in ore treatment site and mine headquarters. The Nova Estancia
ecosystem, with significant impacts on the local flora and fauna. About hotel, which was operating during the disaster, had its employees and
24 % of the affected area comprised of regions with grass or agriculture. guests buried by the mud. The rail bridge was destroyed and washed
Agriculture is an important economic activity for people who live in the away by the mudflow, showing the high stream power of the mud
Vila Parque da Cachoeira. The regions classified as “Bare soil” (11 %) running downhill (Fig. 8a). Based on satellite images it is difficult to
mostly result from mining activities. The tailings dam represents 13 % determine the exact number of buildings inside the area affected by the
of the mapped area. The high difference in altitude between the tailings dam rupture. The industrial facilities at the ore extraction unit and the
dam and the Paraopeba River (> 200 m as shown in Fig. 8b) and the residential areas of Córrego do Feijão and Parque da Cachoeira districts
moderate roughness of the terrain (Fig. 8d) accelerated the mud-flow. were the main populated areas reached by the mudflow. The mud, full
The affected area is mainly a mixture of Tree and Agriculture/Grass. of toxics components, eventually reached the Paraopeba River, causing
Agricultural areas typically have low roughness and mostly occur on serious pollutions in the aquatic system (Thompson et al., 2020). The
flat terrain, whereas trees usually have higher roughness and occupy Paraopeba River is an important tributary of the São Francisco River –
predominantly steeper surfaces (Fig. 8a, c, and d). Therefore, the one of the largest rivers in Brazil. Immediately after the input of the
roughness of the terrain, depending on LULC, affected the rate at which tailings mud, the river’s SPM concentration

9
L.H. Silva Rotta, et al. Int J Appl Earth Obs Geoinformation 90 (2020) 102119

Fig. 9. Suspended particulate matter (SPM) from the affected area and up to 60 km along the Paraopeba River, before (a) (January 22nd, 2019) and after (b)
(February 1st, 2019) the dam rupture.

increaseddramaticallyAccording to Souza (2019), the environmental throughout the first 60 km of the Paraopeba river (Fig. 9a), which is
impacts of Brumadinho dam rupture was massive and difficult to also shown in the boxplot (Fig. 10a). Mean values between 200 and
quantify. 300 mg/l were observed in the river before the dam rupture. On Feb-
ruary 1st 2019 (after the accident) the SPM values displayed high
variability along the river and the mean concentration dramatically
4.5. SPM in the Paraopeba River
increased, overall indicating the direct impact of the tailings mud. The
highest concentrations of SPM (median between 700 and 1000 mg/l)
Fig. 9 shows the SPM concentration maps of the Paraopeba River
were observed over a distance of up to 40 km. There was a decrease in
before (January 22nd 2019) and after (February 1st 2019) the dam
the SPM values after 40 km; however, SPM values were still higher than
collapse. SPM boxplots are shown in Fig. 10. Using the median of the
those before the dam rupture. We propose that this decrease in SPM
SPM concentrations of every 5 km along 60 km downstream, we found a
values was primarily due to the sedimentation of the suspended solids
remarkable difference between the two maps (Fig. 10). Although SPM
in that tract, as the river width becomes wider and flow velocity de-
values of up to 3000 mg/l are observed in the boxplot after the dam
creased (i.e. due to increased width/depth ratio). An increase in the
rupture, only a few points exceeded 2000 mg/l. Thus, to better visualize
SPM values on the map of February 1st 2019 was noted at ∼60 km from
the difference in SPM concentrations, we used a palette ranging from 0
the entry point of mud into the river, close to Juatuba. This increase in
to 2,000 mg/l.
SPM may have occurred due to resuspension of solids caused mainly by
On January 22nd 2019, the SPM remained relatively constant

10
L.H. Silva Rotta, et al. Int J Appl Earth Obs Geoinformation 90 (2020) 102119

Fig. 10. SPM Boxplots from sections extracted at 5 km interval along the Paraopeba river before (January 22nd 2019) (a) and after (February 1st 2019) (b) the dam
rupture. SPM median values are given in (c) for efficient comparison.

precipitation in that region, which has increased the turbulence locally. respectively. R.G.N. acknowledge FAPESP n° 2018/01033-3. L.H.S.R
acknowledges the Coordination for the Improvement of Higher
5. Conclusion Education Personnel – (CAPES) for the postdoctoral fellowship. E.P.
acknowledges the SUG-NAP (3/19EP) of the Nanyang Technological
In this paper, we investigated the possible cause of the impacts of University. We acknowledge NASA for remote sensing data (Landsat
the Brumadinho dam collapsed on the 25th January of 2019 in Brazil. and SRTM) used in this study.
The time series analysis of satellite-driven soil moisture index showed
that the water accumulated over the Brumadinho dam surface during a References
protracted period increased its moisture contents and accelerated see-
page erosion (piping) through the fill. We provide solid evidence of the Agram, P.S., Jolivet, R., Riel, B., Lin, Y.N., Simons, M., Hetland, E., Doin, M.P., Lasserre,
seepage erosion from the top through the fill, which chronically wea- C., 2013. New radar interferometric time series analysis toolbox released. Eos Trans.
Am. Geophys. Union 94, 69–70.
kened the structure and likely led to the collapse of the dam. The rapid Aires, U.R.V., et al., 2018. Changes in land use and land cover as a result of the failure of a
subsidence rate measured by InSAR analysis (even after the pond dried miningtailings dam in Mariana, MG, Brazil. Land Use Policy 70, 63–70.
out) and the large-scale slumping upon the rupture collectively in- Berardino, P., Fornaro, G., Lanari, R., Sansosti, E., 2002. A new algorithm for surface
deformation monitoring based on small baseline differential SAR interferograms.
dicates that the dam was undergoing the liquefaction process internally. IEEE Trans. Geosci. Remote Sens. 40, 2375–2383.
The mudwave stemmed from the dam collapse killed more than 250 Cambridge, M., Shaw, D., 2019. Preliminary reflections on the failure of the Brumadinho
people and devastated huge forest and agricultural areas. The toxic mud tailings dam in January 2019. Dams Reserv. 29, 113–123.
Carmo, F.F., et al., 2017. Fundão tailings dam failures: the environment tragedy of the
reached the Paraopeba river after traveling ∼10 km from the destruc- largest technological disaster of Brazilian mining in global context. Perspect. Ecol.
tion site and the SPM concentration dramatically increased near the Conserv. 15, 145–151.
affected area reducing the light penetration into the water column, Coimbra, K.T.O., Alcântara, E.F.C., Souza Filho, C.R., 2019. An assessment of natural and
manmade hazard effects on the underwater light field of the Rio Doce River con-
affecting the biological productivity.
tinental shelf. Sci. Total Environ. 685, 1087–1096.
Durães, M.F., Mello, C.R., Beskow, S., 2016. Sediment yield in Paraopeba River Basin –
CRediT authorship contribution statement MG, Brazil. Int. J. River Basin Manag. 14, 367–377.
Erten, E., Rossi, C., 2019. The worsening impacts of land reclamation assessed with
Sentinel-1: the Rize (Turkey) test case. Int. J. Appl. Earth Obs. Geoinf. 74, 57–64.
Luiz Henrique Silva Rotta: Conceptualization, Methodology, Farr, T.G., et al., 2007. The shuttle radar topography mission. Rev. Geophys. 45, RG2004.
Formal analysis, Data curation, Writing - original draft. Enner Francini-Filho, R.B., et al., 2019. Remote sensing, isotopic composition and metage-
Alcântara: Conceptualization, Methodology, Formal analysis, Writing - nomics analyses revealed Doce River ore plume reached the southern Abrolhos Bank
Reefs. Sci. Total Environ. 697, 132038.
review & editing. Edward Park: Conceptualization, Methodology, Garcia, L.C., Ribeiro, D.B., Roque, F., de, O., Ochoa-Quintero, J.M., Laurance, W.F., 2007.
Formal analysis, Writing - review & editing. Rogério Galante Negri: Brazil’s worst mining disaster: corporations must be compelled to pay the actual
Formal analysis, Writing - review & editing. Yunung Nina Lin: environmental costs. Ecol. Appl. 27, 5–9.
Hatje, V., et al., 2017. The environmental impacts of one of the largest tailing dam fail-
Conceptualization, Methodology, Formal analysis, Writing - review & ures worldwide. Sci. Rep. 7, 1–13.
editing. Nariane Bernardo: Formal analysis, Writing - review & Kossoff, D., Dubbin, W.E., Alfredsson, M., Edwards, S.J., Macklin, M.G., Hudson-Edwards,
editing. Tatiana Sussel Gonçalves Mendes: Data curation, Writing - K.A., 2014. Mine tailing dams: characteristics, failure, environmental impacts and
remediation. Appl. Geochem. 51, 229–245.
review & editing. Carlos Roberto Souza Filho: Conceptualization, Lambin, E.F., Ehrlich, D., 1996. The surface temperature–vegetation index space for land
Methodology, Writing - review & editing. cover and land-cover change analysis. Int. J. Remote Sens. 17, 463–487.
Lin, Y.N., Simons, M., Hetland Eric, A., Muse, P., DiCaprio, C., 2010. A multiscale ap-
proach to estimating topographically correlated propagation delays in radar inter-
Declaration of Competing Interest
ferograms. Geochem. Geophys. Geosyst. 11, 1–17.
Martin, T.E., McRoberts, E.C., 1999. Some Considerations in the Stability Analysis of
Upstream Tailing Dams. AGRA Earth & Environmental Limited, Canada.
The authors declare that they have no known competing financial
Meckel, T.A., Ten Brink, U.S., Williams, S.J., 2007. Sediment compaction rates and
interests or personal relationships that could have appeared to influ- subsidence in deltaic plains: numerical constraints and stratigraphic influences. Basin
ence the work reported in this paper. Res. 19, 19–31.
Morgenstern, N.R., Vick, S.G., Viotti, C.B., Watts, B.D., 2016. In: Samarco, S.A., Vale, S.A.
(Eds.), Fundão Tailings Dam Review Panel: Report on the Immediate Causes of the
Acknowledgements Failure of the Fundão Dam. BHP Brasil Ltd., Mariana, Brazil. Available online: http://
fundaoinvestigation.com/the-panel-report/ (Accessed 23 March 2019).
E.A. acknowledge the Sao Paulo Research Foundation (FAPESP) Mountrakis, G., Im, J., Ogole, C., 2011. Support vector machines in remote sensing: a
review. ISPRS J. Photogramm. Remote Sens. 66, 247–259.
under grant 19/00259-0. E.A. and C.R.S.F. acknowledge the Brazilian Nechad, B., Ruddick, K., Park, Y., 2010. Calibration and validation of a generic multi-
National Council for Scientific and Technological Development – CNPq sensor algorithm for mapping of total suspended matter in turbid waters. Remote
for the research grants: 309712/2017-3 and 303169/2018-4, Sens. Environ. 114, 854–866.

11
L.H. Silva Rotta, et al. Int J Appl Earth Obs Geoinformation 90 (2020) 102119

Ojha, C., Shirzaei, M., Werth, S., Argus, D.F., Farr, T.G., 2018. Sustained groundwater loss Panel-Report-ENG.pdf > . (Accessed 24 March 2020).
in California’s Central Valley exacerbated by intense drought periods. Water Resour. Santamarina, J.C., Torres-Cruz, L.A., Barchus, R.C., 2019. Why coal ash and tailings dam
Res. 54, 4449–4460. disasters occur. Science 364, 526–528.
Parida, B.R., Collado, W.B., Borah, R., Hazarika, M.K., Samarakoon, L., 2008. Detecting Sato, Y., Godinho, H.P., 2003. Migratory fishes of the São Francisco River. In: Carolsfeld,
drought-prone areas of rice agriculture using a MODIS-derived soil moisture index. J., Harvey, B., Ross, C., Baer, A. (Eds.), Migratory Fishes of South America. World
GISci. Remote Sens. 45, 109–129. Fisheries Trust/The World Bank/International Development Research Centre,
Porsani, J.L., Jesus, F.A.N., Stangari, M.C., 2019. GPR survey on an iron mining areas Ottawa, pp. 195–232.
after the collapse of the Tailing Dam I at the Córrego do Feijão Mine in Brumadinho- Smith, M., 2002. Liquefaction in Dump Leaching. Mining Magazine.
MG, Brazil. Remote Sens. 11, 1–13. Souza Jr., P.A., 2019. Open data could have helped us learn from another mining dam
Potić, I., Bugarski, M., Matić-Varenica, J., 2017. Soil moisture determination using remote disaster. Sci. Data 6, 1–2.
sensing data for the property protection and increase of agriculture production. In: Sun, Q., Jiang, L., Jiang, M., Lin, H., Ma, P., Wang, H., 2018. Monitoring coastal re-
Responsible Land Governance: Towards an Evidence Based Approach 2017 World clamation subsidence in Hong Kong with distributed scatterer interferometry.
Bank Conference on Land and Poverty. The World Bank - Washington DC, March Remote Sens. 10, 1–23.
20–24. Thompson, F., et al., 2020. Severe impacts of the Brumadinho dam failure (Minas Gerais,
Riel, B., Simons, M., Ponti, D., Agram, P., Jolivet, R., 2018. Quantifying ground de- Brazil) on the water quality of the Paraopeba River. Sci. Total Environ. 705, 135914.
formation in the Los Angeles and Santa Ana coastal basins due to groundwater Toming, K., Kutser, T., Laas, A., Sepp, M., Margot, S., Nõges, T., 2016. First experiences in
withdrawal. Water Resour. Res. 54, 3557–3582. mapping lake water quality parameters with Sentinel-2 MSI imagery. Remote Sens. 8,
Riley, S.J., DeGloria, S.D., Elliot, R., 1999. A terrain ruggedness index that quantifies 1–14.
topographic heterogeneity. Intermount. J. Sci. 5, 1–4. Zhan, Z., Qin, Q., Wang, X., 2004. The application of LST/NDVI index for monitoring land
Robertson, P.K., Melo, L., Williams, D.J., Wilson, G.W., 2020. Report of the Expert Panel surface moisture in semiarid area. IGARSS 2004 - IEEE International Geoscience and
on the Technical Causes of the Failure of Feijão Dam I. . Available at < https:// Remote Sensing Symposium.
bdrb1investigationstacc.z15.web.core.windows.net/assets/Feijao-Dam-I-Expert-

12