GET TO KNOW MAPBIOMAS ALERTA METHOD

MapBiomas Alerta is a system that compiles alerts from different deforestation detection systems in Brazil, all based on remote sensing. The processing of this set of alerts involves the aggregation, validation, and refinement of spatial boundaries based on the use of high-resolution satellite imagery (Planet Scope with 3.7 m spatial resolution), generating reports, and publishing the final results on a single, open-access platform (https://plataforma.alerta.mapbiomas.org).

This section provides a simplified explanation of the MapBiomas Alerta processing methodology. It also highlights some limitations of the method and differences compared to official annual deforestation data.

STEP DESCRIPTION

The MapBiomas Alerta process encompasses the following steps: compilation, validation, refinement, cross-referencing with public data, auditing, and publication of deforestation alerts and reports (Figure 1).

Figure 1. MapBiomas Alerta's methodological process for compiling, validating, refining, cross-referencing data, auditing, and publishing deforestation alerts in Brazil. 

Step 1: Compilation of alerts from existing systems for all Brazilian biomes.

MapBiomas Alerta gathers, organizes, and consolidates deforestation information from various official and independent monitoring systems in Brazil. These systems utilize satellite imagery with resolutions ranging from 10 meters to 60 meters to generate deforestation alerts (Table 1). In 2023, MapBiomas Alerta consulted alerts provided by the following sources and systems on a monthly basis:

Table 1: Deforestation Detection Systems and Sources Used in Brazilian Biomes by MapBiomas Alerta

Name	Source	Biome(s)	Link
DETER (MapBiomas Alerta only processes and analyzes deforestation alerts. It does not handle alerts related to forest degradation, fires, or timber extraction.)	National Institute for Space Research (INPE)	Amazon and Cerrado	https://terrabrasilis.dpi.inpe.br/
SAD	IMAZON	Amazon	https://imazon.org.br/categorias/sad/
SAD Caatinga	Geodatin	Caatinga	n.a.
SAD Mata Atlântica	SOS Mata Atlântica e ArcPlan	Atlantic Forest	https://www.sosma.org.br/iniciativas/alertas
SAD Pantanal	SOS Pantanal e ArcPlan	Pantanal	n.a.
SAD Pampa	GeoKarten e UFRGS	Pampa	n.a.
SAD Cerrado*	Instituto de Pesquisa Ambiental da Amazônia (IPAM)	Cerrado	https://sadcerrado.ipam.org.br/
SIRAD-X	Instituto Socioambiental (ISA) and rede Xingu+	Xingu basin region in the Amazon and Cerrado	https://xingumais.org.br/siradx
GLAD	Maryland University	Pampa	https://glad.umd.edu/
PRODES**	National Institute for Space Research (INPE)	Amazônia, Cerrado, Pampa e Pantanal	https://terrabrasilis.dpi.inpe.br/

* For the period July to December 2023, the SAD Cerrado system was analyzed to identify deforestation alerts with areas exceeding 10 hectares.

** To ensure comprehensive coverage and prevent omissions, polygons from the PRODES deforestation monitoring system were incorporated

The SIRAD-X complements the data from SAD and DETER with deforestation monitoring using radar images from the Sentinel-1 satellite in the Xingu Basin, within the Amazon and Cerrado biomes. The SAD Caatinga was developed by Geodatin in partnership with the State University of Feira de Santana (UEFS) to detect deforestation, primarily in dry forests within the Caatinga biome, but may include some areas within the Cerrado and Atlantic Forest biomes, through the comparison of Sentinel 2 images (with a spatial resolution of 10 meters). The SAD Mata Atlântica and SAD Pantanal, both using Sentinel 2 images, were developed with a focus on forest and savannah formations within their respective biomes, by SOS Mata Atlântica and SOS Pantanal in partnership with ArcPlan. The SAD Pampa, also using Sentinel 2 images, was developed by GeoKarten in partnership with the Federal University of Rio Grande do Sul, and is operational for detecting deforestation in forest environments and is undergoing tests for grassland environments (not yet included in the 2023 Deforestation Annual Report). The SAD Cerrado was developed by IPAM in partnership with the Image Processing and Geoprocessing Laboratory of the Federal University of Goiás (LAPIG-UFG) and MapBiomas, with a focus on forest, savannah, and grassland formations within the Cerrado biome using Sentinel-2 images with a resolution of 10 meters.

In addition to the sources of monthly alerts, annual alerts are being included to avoid omissions (PRODES/INPE, Atlas of Forest Remnants/SOS Mata Atlântica and INPE for the Atlantic Forest).

Step 2: Validation and selection of before and after images

Validation occurs in two steps. The first is done in an automated way, eliminating all alerts that overlap with agricultural areas from the annual land use and land cover maps of MapBiomas Brasil or that have already been detected in previous surveys, because it configures the vegetation cover as not being native. The second step is done visually by trained analysts organized into teams per biome, with the support of monthly high resolution Planet mosaics (4 m resolution images). At this point, alerts that correspond to cases of false positives are discarded, with the corresponding record of the reason for rejection (forestry, agriculture, seasonality, etc.). When the alert is considered valid, an image where it is possible to visualize the native vegetation before deforestation and an image where it is possible to see the area that was deforested are selected and purchased with project resources. The purchase of the images considers a minimum area of 500 by 500 m and a surrounding area that helps to contextualize the deforested area.

Step 3: Validation and refinement in high-resolution images

After confirming the deforestation associated with each alert and selecting the pair of high-resolution images, it is necessary to refine the spatial boundaries of the effectively deforested area. This refinement is done through an automated classification process that ensures greater precision in defining the contours of the area where native vegetation has been cleared. The generation of the refined polygon is done using a supervised classification algorithm (Random Forest), which is processed in the Google Earth Engine platform through the Workspace, an application developed by MapBiomas for processing. The classification is performed by collecting training samples from the high-resolution images, both to represent the deforested area and the adjacent non-deforested areas. The final classification results in a refined polygon that undergoes a simplification process to remove excess vertices (Figure 2). Based on the before and after images of deforestation, the interpreter also identifies and records the pressure vector that may have caused the deforestation event (mining, prospecting, urban expansion, agriculture, natural causes, or others).

Figure 2. Example of Planet imagery showing deforestation before and after, and the refined polygon for alert code 927577 from 2023.

Observation

For the SAD Cerrado deforestation detection system, steps 2 and 3 (image selection and validation/refinement of detected polygons) are combined into a single process within the Workspace application. This streamlined approach utilizes Sentinel-2 satellite imagery with a 10-meter resolution to efficiently validate and refine the detected polygons. The decision to combine steps 2 and 3 for SAD Cerrado stems from the high volume of alerts generated by this system.

Step 4: Cross-referencing with public territorial databases

The refined polygons are spatially overlaid with spatial land and surveillance information, including boundaries of Indigenous Lands (ITs), Conservation Units (UCs), quilombola territories, rural settlements, areas registered in the Rural Environmental Registry (CAR) - including declared Areas of Permanent Preservation (APPs) and Legal Reserve (RL) - in addition to areas embargoed by the environmental agency, suppression authorizations and forest management plans from IBAMA's Sinaflor. The alerts are also linked to geographical boundaries such as municipalities, states, biomes, and watersheds. Crossings with special territories are also considered, such as Legal Amazon, area of application of the Atlantic Forest Law, MATOPIBA, AMACRO, Biosphere Reserves, and others. These crossings qualify the alerts and allow the generation of technical reports based on information that is relevant to the user institutions.

Step 5: Auditing

Each refined polygon goes through an audit process done by the technical supervisor of each biome. At this stage, the eventual need to redo some adjustments is evaluated before the final publication of the confirmed deforestation.

Step 6: Publication

All confirmed deforestation polygons are published on the MapBiomas Alert Platform, with weekly updates. Reports are available for each confirmed deforestation and for each intersection of an alert with a property registered in the CAR, SIGEF and SNCI (with an area larger than 0.1 ha). The reports contain the following information:

• deforestation alert code;
• original source of the alert (detection system);
• Biome, State e Municipality;
• desforastation area;
• deforestation area that intersects with the property;
• property code;
• image and date from before deforestation;
• image and date from after deforestation;
• overlap of deforestation with: APP, Legal Reserve, springs, Indigenous Lands, Conservation Units, Sustainable Forest Management Plan, embargoed areas, vegetation suppression authorization, and others;
• simplified description of the coordinates of the deforestation polygon
• MapBiomas land cover and land use history in the evaluated area.

Post-Publication Alert Cancellation and Correction

Under certain circumstances, alerts published on the MapBiomas Alerta Platform may be corrected or even canceled. Whenever there is a formal indication or a well-founded request pointing to possible errors associated with the alerts, whether from environmental agencies or platform users, the technical team carries out a new thorough analysis of these alerts. 

This analysis is carried out by verifying Planet imagery, but also, if necessary, various other complementary information such as images from other satellites (Sentinel, Landsat, etc.), high-resolution satellite images available on Google Earth, in addition to the MapBiomas Land Use and Cover Mapping. In cases where it is confirmed that the published alert does not in fact represent a deforestation event/conversion of native vegetation (regardless of timber yield, regularity or responsibility), the alert is canceled. This means that it is removed from the platform's map and statistics, with the reason for cancellation being indicated. The polygon of the removed alert is kept in the database only for individual consultation through its identification code, where the reason for its cancellation is recorded. 

In some cases, spatial delimitation rectifications of the alert may be carried out, always with the aim of better representing the deforestation event in question (Figure 3). Similarly, if any error or problem is related to the images linked to the alert polygon, new images can be selected and updated on the platform. All rectifications are recorded in the system and the information is made available publicly on the platform, including the date on which the alert was rectified.

Figure 3. Example of rectification of spatial boundaries of the alert after publication for alert Code 564078, detected in the year 2022.

The MapBiomas Alerta platform does not analyze or determine the legality or regularity of the deforestation alerts it presents. The platform's primary function is to detect and confirm the loss of native vegetation, regardless of legal or regulatory considerations.

LIMITATIONS OF THE METHOD 

As every method, MapBiomas Alerta has some limitations that must be considered when applying its data:

1. Processing time - the importation of alerts from their sources (detection systems) occurs monthly, with the exception of DETER alerts, which occur every 15 days. As part of the alerts processing is done individually and visually by trained analysts, the validation and processing time depends on the biome and the time of year. It can vary from 30 to 90 days from the date of detection by the source system to publication on the MapBiomas Alerta platform. The purpose of MapBiomas Alerta is to increase certainty about confirmed deforestation data and provide ready reports for remote surveillance. Rapid field surveillance operations aimed at flagrant deforestation can be planned directly with the pre-existing detection systems.

2. Alert Omissions – deforestation are validated and refined based on the existence of an alert previously captured by a third-party deforestation detection system. The possible omissions of these systems in detecting deforestation also affect the alerts evaluated by MapBiomas Alerta. 

At the beginning of the project, most of the Brazilian biomes did not have a monthly monitoring system, and the main source of alerts used was GLAD. GLAD is a global system that uses images from Landsat satellites to automatically flag areas where forest cover has been disturbed. The system covers the entire tropical region. However, the alerts detect changes more reliably in areas with at least 60% forest cover, making them more useful in dense tropical forests. Therefore, it had omissions because it was not adjusted to the specific characteristics of each Brazilian biome. 

To avoid omissions, MapBiomas supported the development of Deforestation Alert Systems (DAS) adapted for each Brazilian biome by several universities, research institutions and civil society organizations:

• SAD Caatinga, which started operating in 2020, developed by the MapBiomas na Caatinga team (UEFS and Geodatin);
• SAD Mata Atlântica, developed by SOS Mata Atlântica and ArcPlan, was deployed in 2021 for four river basins (Tietê, Jequitinhonha, Iguaçu and Miranda/Aquidauana), where 2,126 alerts were identified in addition to the GLAD alerts, and has been operating for the entire biome since January/2022;
•  SAD Pantanal, developed by SOS Pantanal and ArcPlan to monitor deforestation alerts in forest and savanna formations, deployed in late 2021 (where 103 alerts were identified), operating monthly as of January 2022;
• SAD Pampa, developed by GeoKarten and UFRGS, which in 2022 operated complementarity to GLAD;
• SAD Cerrado, developed by IPAM in 2022, when there was still a partial addition of alerts. It has been operational since January 2023. During the period from January to June 2023, all generated alerts were validated, regardless of size, while for the period from July to December 2023, only alerts above 10 hectares were considered.

In a complementary manner, the annual deforestation identified by PRODES Amazon and Cerrado were used to identify omissions from the monthly monitoring systems of these biomes. With the same objective of reducing omissions, the annual deforestation identified by the Atlas of Forest Remnants of SOS Mata Atlântica/INPE were also included, rejecting those that overlap with alerts already validated.

It is also worth noting that deforestation monitoring systems present minimum detection areas and, therefore, may omit some deforestation. For example, alerts smaller than 6.25 hectares are not detected in the Amazon (DETER Amazon) and those smaller than 1 hectare are not detected in the Cerrado (DETER Cerrado). The utilization of multiple data sources for the same region aims to reduce omissions in deforestation detection.

3. Deforestation Speed Underestimated – when validating and refining an alert, a search is performed for a pair of good quality Planet satellite images of before and after deforestation. The "before" image is the most recent available in the period up to 12 months before detection, and the "after" image is the one closest to the end of deforestation. The presence of clouds can increase the period between before and after images by days, weeks, and even months. This does not change the statement that deforestation occurred in the period between the two images, but it does affect the calculation of the average speed at which deforestation actually occurred.

4. Automatic Polygon Delimitation – polygons delimiting the refined alerts are established by a process of automatic classification of the area of change between the two images, i.e., the place where the native vegetation was suppressed. In delimiting the deforestation polygon, areas with signs of previous change or with small clusters of trees that were maintained in the midst of deforestation are removed. In 2020, a procedure was developed to minimize the small islands within the polygons in the refinement step, as well as the excess of vertices *which correspond to the points that form the polygons). 

5. Limitation to Non Woody Native Vegetation – the detection of non-forest vegetation suppression, such as grassland vegetation, for example, has limitations in the alert-generating systems, whose methods focus on identifying where forest vegetation has been suppressed. Except for SAD Cerrado, which has calibrated detection for both forest formations and savannas and grasslands. However, when suppression of non-forest vegetation also occurs in the alert area or in adjacent areas for other biomes, the use of high-resolution images allows its registration during the alert refinement phase. Therefore, most of the non-woody vegetation clearings detected since 2019 occurred occasionally, whenever observed around woody vegetation alerts. Therefore, current detection systems still underestimate the suppression of non-forest native vegetation.

DIFFERENCES FROM THE OFFICIAL ANNUAL DATA  

When comparing deforestation data from MapBiomas Alerta with official deforestation data from PRODES (Program for Monitoring the Brazilian Amazon Forest by Satellite), it is crucial to exercise caution due to inherent differences between the two systems. These differences, summarized in Table 2, can lead to variations in reported deforestation estimates.

Table 2. Differences between official deforestation data systems and MapBiomas Alert in 2023

Subject	PRODES Amazon	PRODES Caatinga, Cerrado,Atlantic Forest, Pantanal e Pampa	ATLAS Atlantic Forest	MapBiomas Alerta
Minimum Mapped Area	6,25 ha	1 ha	3 ha	0,3 ha
Area Calculation	releases rate that estimates deforestation also in unobserved areas	data represents the sum of the observed areas	data represents the sum of the observed areas	data represents the sum of the observed areas
Analysis period	August 2019 to July 2022	August 2019 to July 2022	October 2018 to September 2022	deforestation detected between January and December 2019, 2020, 2021, 2022 and 2023.
Image capture window	July to September 2019, 2020 ,2021 e 2022	June to September 2020 , 2021 e 2022	July to November 2019, 2020 ,2021 e 2022	July 2018 to December 2023
Territorial Range	Legal Amazon	boundaries of the biomes at scale 1:250,000 (for the Cerrado, subtracting the overlap area with the Legal Amazon)	area of application of the Atlantic Forest Law (biome + enclaves in the northeast)	IBGE biome boundaries at scale 1:250,000
Mapped Vegetation Type	primary or existing forest vegetation in 1988 (excludes cerrado areas and non-forested areas in 1988)	forest and savanna vegetation existing in 2000	primary or existing forest vegetation in 1985	primary vegetation and may include secondary vegetation