Integrating Google Earth Engine and Decametric Sentinel 2 Images for Analysis of Vegetation Pre and Post the Disaster at Brumadinho, Brazil

Abstract

This paper presents the application of the normalized difference vegetation index to assess the vegetation dynamics for the period between years 2017 and 2021 at Brumadinho, MG, Brazil. The normalized difference vegetation index was calculated using a Google Earth Engine script applying Sentinel 2 data with a spatial resolution of 10 meters, to quantify the extent of the affected area and assess the vegetation dynamic after the disaster. The Dwass-Steel-Crichlow-Fligner test for nonparametric data was used for a pairwise comparison between years and the confidence interval was calculated using bootstrap with 9999 repetitions. The total area affected by the dam brake was 2662 ha. The NDVI values presented a statistically significant decrease from 2017 to 2019, with little increase until 2021. Mean NDVI values were 0.314003 [0.31028; 0.317564], 0.339887 [0.336591; 0.343231], 0.145814 [0.144004; 0.1476], 0.1495 [0.147676; 0.15128], and 0.15572 [0.153727; 0.15774] for 2017–2021, respectively. According to the results, we conclude that the vegetation in the affected area did not fully recover.

1. Introduction

In recent decades, there has been an increase in concern about environmental disasters. This concern fosters the need for emergency management to mitigate socioeconomic consequences [1]. Disaster management is defined as the field of science that develops and applies technologies, planning, and management to deal with extreme events. Whether of natural or anthropic origin, events are managed that can kill or injure people and animals, as well as cause extensive damage to properties and communities [2].

Occurred in 25-01-2019, the catastrophe of Brumadinho released 43 million m³ of iron ore tailings enters the list of the biggest mining disasters of history, being classified by number of deaths: Bulgaria, 1966, lead-zinc tailings (488 deaths); Brazil (Brumadinho), 2019, iron ore tailings (363), Chile, 1965, copper tailings (300) and China, 2008, iron tailings (277) [3].

Several challenges arise when dealing with disasters, the main ones being the need to analyze large spatial extensions in a short time [4]. Both in the drafting of evacuation routes, or in the analysis of potential risks, managers deal with aspects related to space. Still, in the context of developing countries, they face challenges related to a lack of financial resources and trained analysts [5]. In this context, geographic information systems (GIS) and remote sensing translate as tools to support geographic analysis, through storage, processing, and access to spatialized information [6]. Several studies assessing the social, economic, and ecologic impacts over flood areas using GIS and orbital remotely sensed data have been deployed [7, 8, 9].

In this context, the normalized difference vegetation Index (NDVI) is a ratio index calculated using the difference between near-infrared reflectance and red reflectance to their sum, being widely used to assess vegetation dynamics before and after disasters [10, 11, 12]. Rotta et al. [13] assessed pre-disaster scenarios and the causes to the dam collapse using satellite-driven soil moisture index, multispectral high-resolution imagery, and Interferometric Synthetic Aperture Radar (InSAR) products to assess pre-disaster scenarios and the direct causes of the tailings dam collapse. Cheng et al. [14] used Landsat 8 operational land imager products to assess sediment concentration and found an increase of sediments in the Paraopeba River due to the mudflow. Gama et al. [15] using Sentinel −1 InSAR data were able to observe persistent trends of deformation on the crest, middle and bottom sectors of the dam, which may have caused the collapse.

Nonetheless, to the researchers’ knowledge, there is a gap regarding the assessment of vegetation recovery for the impacted area for subsequent years and compared to the previous dynamic. This is due to the large quantity of images that would require a large amount of time and processing capacity. For example, for an entire year, there would be 73 Sentinel 2 images to process. In this chapter, we assessed the NDVI reflectance for 2018, 2019, and 2020, a total of 217 images with a spatial resolution of 10 meters. To tackle this problem, we used the Google Earth Engine cloud computing environment.

Therefore, the research question that led the discussion is “how has the vegetation recovered after two years of the disaster?” Therefore, the aim of this work is to assess vegetation dynamics by analyzing the reflectance values for the NDVI index for 2018, 2019, and 2020.

1.1 Remote sensing applied to analyzes of environmental impacts of disasters

In a post-disaster crisis situation, environmental impact managers and analysts need accurate information in the short term. The information needs to show the spatial and temporal scale of what has happened and what can happen. Based on this information, the necessary resources will be allocated to leverage immediate responses to contain the damage. Access to spatialized information allows the preparation of plans to anticipate contingencies; evaluation of possible scenarios; efficiently and effectively; and actions for recovery and reconstruction of damages [16].

Thus, GIS and remote sensing translate into a key tools in impact analysis and disaster management. This technology has the capacity to aggregate socioeconomic information; images of remote sensing; storage, manipulation, queries, and data analysis; and, more importantly, the visualization of the data [17].

The use of GIS for disaster management can support several aspects of decision-making, such as:

• Disaster prediction: disaster impacts extent and possible impacts to high-importance areas [18];

• Vulnerability analysis: database spatialization regard services such as hospitals, police stations, fire brigades, and shelters, which can be used to mitigate the situation; and other structures with the potential to aggravate the situation, such as dams and effluent treatment plants [19];

• Evaluation of the magnitude of the damages: analyzes of the geographic extent of the impacts caused by the disaster, it assists decision-makers in the elaboration of containment measures for priority areas, such as nurseries, hospitals, schools, and nursing homes [20];

• Identification of hazardous materials: identifies locations that contain materials with potential for contamination, bringing the type of material, quantity, and reactive aspects, such as gas stations and pharmaceutical laboratories [21];

• Human resources: identifies the location and personal information of individuals with skills or experience useful in emergency response situations [22];

• Resources inventory: aggregates data regarding shelters and provision of resources such as food and water. Equipment is also accounted, such as boats, motor vehicles, and kite trucks; brings resources that can be made available by neighboring cities [23];

• Infrastructure: aggregates information regarding roads, airports, bridges, and evacuation routes; in addition to structures such as electrical networks, water supply networks, and sewage collection networks [24].

2. Materials and methods

The disaster occurred in the Brumadinho Municipality, at Minas Gerais State, Brazil. The Córrego do Feijão Dam, the Paraopeba river, and the affected area by the dam break are presented in Figure 1. The impacted area was obtained by the researchers by setting a threshold of values less than 0, converting it in a Boolean image, converting it to a vectoral file, and by overlaying selecting the polygon that matched the crisis image.

Figure 1.

Study area location.

2.1 Socioeconomic and environmental description

With an estimated population of 39,520 [25], a Municipal Human Development Index of 0.747 and GDP per capita of more than R $ 40 thousand, the municipality of Brumadinho belongs to the Metropolitan Region of Belo Horizonte and is inserted in the Quadrilátero Ferrífero where the main economic activity is iron mining. It is crossed by the Paraopeba River and a member of the Aguas Claras and Rio Manso River Basin, a source of supply for about 28% of the population.

The municipality has as a predominant biome the Atlantic Forest and some remnants of Cerrado, protected by the Special Protection Area (APE) of the Rio Manso; Environmental Protection Area (APA) Sul—RMBH; APA Inhotim; APE Catarina; Serra do Rola-Moça State Park, Serra da Moeda Natural Monument, Serra do Rola-Moça Natural Monument and Private Reserves of Natural Heritage—RPPN. About its climate, Brumadinho is located in the zone of influence of the Climate Cwa—Tropical Altitude, with summer rains, hot summers, and dry winters.

2.2 Flowchart with research steps

The steps for the work deployment are presented in Figure 2. The description of each step is presented below. All data were managed with QGis version 3.10 [26].

Figure 2.

Research procedures flowchart.

2.3 Acquisition of geographic data

The georeferenced images were retrieved from the United States Geological Survey. Multispectral imager sensor images from Sentinel-2A satellite with 10 meters resolution were used. Sentinel-2 images used to deploy the real color composite (RGB432) details are described in Table 1.

Satellite/sensor	Wavelength (nm)	Period	Acquisition date	Track
Sentinel 2A—multispectral instrument (MSI)	Blue—458–523	Archive	01/22/2019	131/241
	Green —543–578	Crisis	02/01/2019	131/241
	Red—650–680
	Red edge—698–713
	Red edge—733–748
	NIR—785–899
	NIR narrow—855–875
	SWIR—1565–1655
	SWIR—2100–2280

Table 1.

Sentinel-2 images characteristics.

2.4 3D modeling

For hypsometry mapping and 3D modeling, the Shuttle Radar Topography Mission (SRTM) data were used. The SRTM data were acquired from the Topodata [27] web platform, a Brazilian program that resampled the SRTM data to 30 meters resolution.

The SRTM provided by the Topodata program comes in WGS1984 datum. For analytic and standardization purposes, all data were converted to SIRGAS2000 Universal Transverse Mercator (UTM) zone 21S datum.

2.5 NDVI calculation using google earth engine

For this study, researchers deployed a Google Earth Engine script to calculate normalized difference vegetation index (NDVI) values for the periods of January 1th, 2018 to December 31th, 2018; January 1th, 2019 to December 31th, 2019; and January 1th, 2020 to December 31th, 2020, according to the sentinel 2 constellation revisit period of 5 days. Level 2 (L2) orbital remote sensing products from the Sentinel-2 satellite were used, the L2 algorithm generates ortho corrected, atmospherically corrected, and with bottom-of-atmosphere reflectance. The constellation of two satellites—Sentinel-2A and Sentinel-2B—orbit the Earth at an altitude of 786 km but are separated by 180° to optimize global coverage and revisit times.

The NDVI, proposed by Rouse et al. [28], is the most common ratio index for remote sensing environmental analysis. It allows us to identify areas with dense vegetation and areas with no vegetation. It is calculated using the near infrared (NIR) and red bands, applying the following equation:

$$\mathit{\text{NDVI}}=\frac{\left(\mathit{\text{ρNIR}}-\mathit{\text{ρRed}}\right)}{\left(\mathit{\text{ρNIR}}+\mathit{\text{ρRed}}\right)}$$E1

The NDVI presents values that vary from −1 to 1, where values below 0 are considered non-vegetated areas, such as bare soil, mud, or water, and as closer to 1 get, the healthier the vegetation.

This index presents high accuracy for comparisons in time and spatial scales, considered suitable for analyses of areas affected by floods. Flooded areas generally present the characteristics of water or mud presence. These surfaces present low reflectance patterns in the red and NIR spectrum, while vegetated areas present higher reflectance patterns [7].

2.6 Statistical analysis

The descriptive statistics were calculated and confidence interval was obtained using the Bootstrap method with 9999 repetitions. The Shapiro–Wilk test was used to verify normality, results (p < 0.001) show no adherence to normality. The Dwass-Steel-Crichlow-Fligner test, described by Dwass [29], Steel [30, 31], Douglas and Michael [32] for independent nonparametric samples, was used to test if there was a significant difference between years. The pairwise comparison was deployed using NDVI values for 2018, 2019, and 2020. The test assumed alpha equals 0.05.

3. Results and discussion

3.1 Hypsometry of the study area

The Córrego do Feijão dam was located at the right top corner of Figure 3. As it can be noticed, the dam was located at a higher altitude than the Paraopeba river. This river provides drinking water for several municipalities along its course, among other types of usage such as industries, irrigation, and for livestock. The presented data are key for the study of tailing dam location, since it provides information for decision-makers to prevent the installation of these dams near human settlements. The SRTM data can be used to model the impacts of dam breaks and the main areas that will be affected. This method can be used in the environmental licensing of tailing dams.

Figure 3.

3D digital elevation model representation of the study area.

The digital elevation model (DEM) deployed using SRTM data is key for flow accumulation analysis [33]. Mubareka et al. [34] present a study, using SRTM data to predict settlement location and density at a 90 m resolution. The terrain representation in 3D allows decision-makers to tackle priority areas to focus efforts.

3.2 Environmental impacts in Brumadinho, MG, Brazil

In addition to the 363 confirmed deaths, the mud devastated homes, buildings, businesses, and families, leaving only the trail of destruction and uncertainties. Some of the families that lived there had small land where they developed subsistence agriculture as their main source of income, just as fishermen and rural communities lost all their material possessions.

Following the catastrophe, the National Human Rights Commission carried out a mission aimed at promoting qualified listening and proposing emergency actions for the affected populations. During the mission, the possible impacts resulting from the rupture of Córrego do Fundão tailings dam were raised, among them the mortality of specimens in all trophic chains; harming the conservation status of species already listed as endangered and the entry of new species into the list of threatened species, and undermining the structure and function of associated aquatic and terrestrial ecosystems [35].

Figure 4 presents the archive image acquired 01/22/2019 from Sentinel MSI (left) and the crisis image from 02/29/2019 before the Córrego do Fundão tailing dam failure (right).

Figure 4.

Archive image from 22 to 01-2019 on the left panel and crisis image from 01 to 02-2019 on the right panel of the study area.

3.3 Vegetation loss due to Córrego do Fundão tailings dam failure

The Atlantic Forest biome prevailed in the area affected by the tailings Dam I failure. This biome holds up to 8% of the world’s species and is recognized as a hotspot for biodiversity conservation. As a result of anthropic activities, in recent years, its occupied territory has plummeted to less than 15% [36], thus being a global conservation priority [37].

Although mining activity is predominant in the area, besides the Atlantic Forest parks protection, activities, such as agriculture and livestock, both for subsistence, were responsible for a small portion of land use and occupation [38]. Of the affected area, the major affected area corresponds to Atlantic Forest vegetation, and the rest are activities carried out by the population, such as housing and agriculture. It should be noted that food production, food security, and community health were strongly and directly affected since natural resources such as soil, water, and ecosystem interactions were compromised.

With the Vale S/A tailings dam failure in Brumadinho (MG), 11.7 million cubic meters of high silica and iron sludge were dumped under an area of Atlantic Forest native vegetation.

Figure 5 presents the area affected by the dam break, representing a vegetation loss of 2662 ha. The difference can be identified due to high reflectance values in the NIR band by mud-covered areas. Vegetated areas have high absorbance of the green spectrum. As seen in the NDVI Archive histogram, peaks of reflectance can be noticed in 0.8 μm, which shows a high quantity of green vegetation in the area, after the dam break, it can be noticed a significant decrease in peak values, now in 0.1 μm.

Figure 5.

Vegetation dynamics and histograms of NDVI values for years 2017–2021.

The NDVI is the most used remote sensing index for vegetation loss studies [39], for example, in the studies of Mariana environmental disaster and analysis of native Atlantic Forest loss [13]. We used Google Earth Engine to calculate the NDVI, using the affected area polygon, for years 2018, 2019, and 2020. It can be noticed, in Figure 6, lower values for NDVI in the crisis year.

Figure 6.

NDVI values for years 2017–2021.

The mean values, with lower and upper bootstrap confidence intervals, for NDVI for year 2017 were 0.314003 [0.31028; 0.317564], for NDVI in year 2018 values were 0.339887 [0.336591; 0.343231], for year 2019 the values were 0.145814 [0.144004; 0.1476], for year 2020 values were 0.1495 [0.147676; 0.15128], and for year 2021 the values were 0.15572 [0.153727; 0.15774]. The Dwass–Steel–Crichlow–Fligner test for nonparametric data presented a statistically significant (p < 0.001) decrease in vegetation when comparing 2018–2019 and 2020 values. The descriptive and inferential statistics are presented in Tables 2 and 3.

	2017	Bootstrap	2018	Bootstrap			2019	Bootstrap		2020	Bootstrap		2021	Bootstrap
		Lower conf.	Upper conf.		Lower conf.	Upper conf.		Lower conf.	Upper conf.		Lower conf.	Upper conf.		Lower conf.	Upper conf.
N	8817	8817	8817	8817	8817	8817	8817	8817	8817	8817	8817	8817	8817	8817	8817
Min	−0.23			−0.1			−0.16			−0.18			−0.16
Max	0.63			0.62			0.54			0.59			0.57
Sum	2768.56	2735.74	2799.96	2996.78	2967.72	3026.27	1285.64	1269.68	1301.56	1318.19	1302.06	1333.9	1373.03	1355.41	1390.86
Mean	0.314003	0.31028	0.317564	0.339887	0.336591	0.343231	0.145814	0.144004	0.1476	0.1495	0.147676	0.15128	0.15572	0.153727	0.15774
Std. error	0.001848	0.001825	0.00187	0.00171	0.001691	0.001729	0.000945	0.000929	0.000961	0.000935	0.000918	0.000951	0.00105	0.001033	0.001067
Shapiro–Wilk p	<.0001			<.0001			<.0001			<.0001			<.0001

Table 2.

Descriptive statistics for NDVI values according to the period of 01-01-2018 to 31-12-2018, 01-01-2019 to 31-12-2019, and 01-01-2020 to 31-12-2020.

Comparisons		p	Comparisons		p
2017	2018	1	2018	2020	<0 .001*
2017	2019	0.082*	2018	2021	0.004*
2017	2020	0.006*	2019	2020	0.808
2017	2021	0.019*	2019	2021	0.994
2018	2019	0.027*	2020	2021	0.969

Table 3.

Inferential statistics for NDVI values pairwise comparison to the period of 2018–2021.

*

Values which reject the null hypothesis with alpha equals to 0.05.

Table 3 displays the inferential statistics for the NDVI values for the period between 2018 and 2021. The results clearly show that the vegetation did not recover from the disaster in January 2021, with significant difference between the years pre and post the disaster, with exception of the year 2018.

Silveira et al. [40] used Landsat 8 images of before and after the Mariana, MG disaster to detect vegetation loss in the affected area. Their conclusions were that the index produced “highly accurate maps of areas affected by post-dam-failure flooding in the region. This approach can be used in many other contexts for rapid and accurate assessment of such land-cover change.” (Figure 7).

Figure 7.

NDVI and precipitation (mm.Month⁻¹) for the period between 2017 and 2021.

3.4 Disaster management

Disasters are events that escape to normality, involving large negative environmental, economic, and social impacts. Their environmental and socio-environmental consequences can be reversible or not [41]. Its origin can be natural or anthropogenic, it is currently considered that they are, in general, products of interrelation between human activities and natural phenomena [42, 43, 44] as an example of the disaster that occurred in the municipality of Brumadinho, MG in February 2019.

Disaster management presents itself in a cycle divided into mitigation, preparation, response, and recovery. Mitigation implements measures that can eliminate or reduce the degree of risks and hazards. Preparation is where actions are taken in advance in order to develop effective mechanisms to respond to the event. It is followed by the response that is the actions to be implemented as soon as the event occurs minimizing the damage and the recovery phase, that is, the reestablishment of the area and its actions to reach normality [45, 46, 47, 48].

In the last 10 years, according to the Université Catholique de Louvain’s EM-DAT [49], 10 disasters were registered in Brazil between 2010 and 2019 including industrial, transportation, and assorted. Among them is the disaster in the municipality of Mariana (MG) with a total of 25 deaths and the Brumadinho disaster in early 2019 with 363 deaths. Furthermore, it must be accounted for the loss of Atlantic Forest native vegetation, water quality degradation, and livestock deaths as environmental developments to this catastrophe.

In response to these events, Brazil has a specific legislation for disasters, where the Federation, states, municipalities, and organized society have defined roles. The Law 12,608—2012 [50] establishes the responsibilities, goals, and directives of planning with the objective of reducing disasters. This law foresaw the zoning for land use and occupation, creation of a database for risk areas monitoring, prevention, response, and mitigation plans. Nevertheless, its efficiency was absent in disasters such as those mentioned above, it is exposed that there is a need for structuring and implementing disaster management plans.

3.5 Environmental regulation regarding the tailings dam I

The Tailings Dam I, broken in January 2019, belonged to the Córrego do Feijão Dam Complex, it was 87 meters high, medium size, and stored iron ore. According to the Brazilian National Dam Security Plan (PNSB), the dam was categorized as low risk and had high potential damage. This complex belongs to Vale, the largest Brazilian mining company. In this complex, there are still five small dams classified as low risk, as seen in Figure 8, with potential damage between medium and high, Figure 9 [51]. The dam’s potential damage is a function of the human life’s potential loss and economic, social, and environmental impacts. In addition to these, the municipality of Brumadinho has another 20 small and medium-sized dams, between low and medium risk [52].

Figure 8.

Tailing dams risk categories according to the Brazilian National dam Security Plan.

Figure 9.

Tailing dams damage potential according to the Brazilian National dam Security Plan.

In 2017, the Environmental Impact Report (EIA/RIMA) was registered at the State Environmental Foundation of Minas Gerais.

The EIA/RIMA is a direct product of the Environmental Licensing process. It is a regulatory instrument in which the government, represented by environmental agencies, authorizes and monitors the installation and operation of activities that appropriate from natural resources or that are considered effective or potentially polluting. It is mandatory for entrepreneurship, to seek environmental licensing from the competent agency, from the initial stages of its planning and installation until its effective operation.

Environmental impact study and the environmental impact report are included in this context as a regulatory requirement, established by the National Environmental Council (CONAMA) Resolution 001/86. It consists of in situ studies in soil, water, and air to verify if the prospective area contains environmental liabilities. Furthermore, present studies regarding the socio-economic-environmental relationship will be affected by the implementation of the enterprise.

The EIA/RIMA proposed licensing the operational continuity of the Córrego do Feijão dams until the year 2029. It was approved in November 2018 by the Superintendency of Priority Projects (SUPPRI) linked to the State Secretariat of Environment and Sustainable Development of the State of Minas Gerais. Palagi and Javernick-Will [53] states that a major constraint for policy-making and application is “institutional norms and cultural beliefs considerably narrowed the range of options post-disaster decision-makers perceived as viable, appropriate, or compassionate.”

Vale has expressed in an official note regarding the risk management actions that had been implemented before and during the event, such as alarms, leakage routes, and maintenance and monitoring of dams. Regardless of these actions, it is not yet known exactly what triggered the rupture of the dam. As in the Mariana disaster that occurred in 2015, some hypotheses were raised about the disaster in Brumadinho. To solve any doubts regarding this and other possible dam breaks, not only of Vale but other mining companies, the Public Prosecutor’s Office continues to follow the developments, as well as the documents issued where possible problems regarding the dam structure were confirmed.

As a consequence of this act, a stronger posture is expected by the government agencies. On the contrary, the rural and economic development-focused parties in the Brazilian senate propose the flexibilization of the environmental laws. For example, reducing the bureaucratic criteria for the environmental licensing, main regulation instrument in Brazil. In this sense, integrating technology and decision-making is key for disaster management.

4. Conclusions

The vegetation in the area affected by the dam break in Brumadinho did not present a statistically significant recovery according to NDVI values. The total area affected by the dam brake was 2662 ha. The NDVI values presented a statistically significant decrease from 2017 to 2019, with little increase until 2021, NDVI values for year 2017 were 0.314003 [0.31028; 0.317564], for year 2018 the NDVI values were 0.339887 [0.336591; 0.343231], for year 2019 the NDVI values were 0.145814 [0.144004; 0.1476], for year 2020 the NDVI values were 0.1495 [0.147676; 0.15128], and for year 2021 the NDVI values were 0.15572 [0.153727; 0.15774]. The study shows that GIS technology is key for post disaster impacts assessment and monitoring of vegetation recovery and aiding strategic planning of decision-makers. GIS presents low cost and rapid response tools for disaster management, allowing it to focus efforts in priority areas. Furthermore, GIS must be integrated in disaster prevention policies, integrating environmental licensing development processes.

Regarding the information gathered during the study, it is concluded that although Brazil has regulatory instruments, they must be revised for efficient application.

Once the company responsible had employed the disaster management tools, such as GIS-based decision-making, with prevention actions, rapid response to events, and recovery, much of the damage would be minimized and even avoided. Still on disaster management, this is also government responsibility. Decisions must be community driven, where agile responses to events and community well-being are guaranteed, regardless of responsibility.

Approaching all environmental impacts, besides the loss of Atlantic Forest vegetation, directed studies on aquatic and terrestrial ecosystems are of great relevance to explain and elucidate all the damage caused, supporting future discussions about legislation of dam failures and other projects that endanger environmental health. These discussions must focus on prevention other than remediation.