Here we present a summary of the method developed and applied in Collection 3 fire scars mapping in Brazil (1985-2023).
For more methodological details, access the ATBD in this LINK.

GENERAL CHARACTERISTICS

All the fire scars mapping in Brazil was based on mosaics of Landsat satellite images with a spatial resolution of 30 meters. The mapping period was from 1985 to 2023, with monthly and annual data on fire scars covering the entire Brazilian territory.
The process was carried out collaboratively among MapBiomas institutions, using artificial intelligence with machine learning algorithms (deep learning) on Google Earth Engine and Google Cloud Platforms, which offer a huge cloud processing capacity, beyond local servers for additional processing.
The work was organized by biome and region, with samples collected in burned and unburned areas to train the algorithm by region. In addition, reference maps were used, such as the MODIS burned areas product (MCD64A1 - https://lpdaac.usgs.gov/products/mcd64a1v006/) with a spatial resolution of 500 m, GABAM (Global Annual Burned Area Map - https://gee-community-catalog.org/projects/gabam/) with 30 m resolution, fire hotspots, and burn scars from INPE (https://terrabrasilis.dpi.inpe.br/queimadas/bdqueimadas/).

2. METHOD OVERVIEW

The image processing and classification routines used to map burned areas in Brazil followed six stages:

1. Definition of classification regions by biome: The biomes were divided into regions to increase the accuracy of the classification.

2. Creation of annual Landsat mosaics: An annual statistical approach was used to create quality annual mosaics by composing all 16-day images into a single mosaic, selecting the pixel with the lowest NBR (Normalized Burn Ratio) value.

3. Collection of training samples: Spectral signatures of burned and unburned areas were collected from the annual mosaics to make up the model's training sample set.
4. Classification with a Deep Learning model: The DNN (Deep Neural Network) model was trained using the collected samples and the RED, NIR, SWIR1 and SWIR2 spectral bands to classify burnt areas.
5. Post-classification: Masks and spatial filters were applied to improve accuracy and reduce noise in the classification results.
6. Validation with reference data and visual checks: The classification results were validated using reference data, along with visual checks.

2.1. DEFINITION OF REGIONS BY BIOME

For each biome, regions were defined for the collection of training samples and the classification of burned areas by region, with the aim of obtaining a more accurate classification based on edaphoclimatic factors and regional vegetation. The following regions were defined for each biome:

2.2. ANNUAL MOSAICS 

The classification was carried out using surface reflectance (SR) mosaics from the USGS Landsat Collection 2 (Tier 1) (30m × 30m), built for each year from 1985 to 2023. All available scenes from the Landsat 5 (from 1985 to 2011), Landsat 7 (from 1999 to 2012) and Landsat 8 (from 2013 to 2023) satellites were evaluated, with a return interval of 16 days.

2.3. SAMPLE COLLECTION

We created a spectral library based on the manual delineation of burned and unburned areas to be used as training samples. These samples were stratified by Landsat sensors (collected in different years) and by each biome. Training samples were collected from all 28 classification regions, ensuring that the different spectral characteristics present in each region were represented.
Thus, we obtained a sample set per sensor and for the 28 regions of Brazil, to be used for training the classification model.

2.4. CLASSIFICATION

The classification model used was the Deep Neural Network, which consists of computer models based on mathematical calculations capable of performing machine learning and visual pattern recognition.
The burnt area mapping algorithm consisted of two phases: training and prediction. Based on the training samples of burnt and unburnt areas, the following spectral bands were used as input for the burnt area classification model: red (RED - 0.65 µm), near infrared (NIR - 0.86 µm) and shortwave infrared (SWIR 1 - 1.6 µm and SWIR 2 - 2.2 µm). These Landsat spectral bands were chosen based on their sensitivity to fire events. 

The training data was divided into two sets: 70% of the samples used for training and 30% for testing.

2.5. POST-CLASSIFICATION

After training and testing the model, the classification was applied to fortnightly Landsat images for the entire analysis period (1985 to 2023). A spatial filter was applied to remove noise and fill in small empty gaps: areas less than or equal to 1.4 hectares (16 pixels) were removed, and empty gaps less than or equal to 5.8 ha (64 pixels) were filled in as burnt areas.
After evaluating the classification results, post-classification filters were also applied to remove pixels that were in the following land cover and land use classes from MapBiomas Collection 8 in the biomes:

• Amazon: Water, Urban Area, Mining, Beach, Dune and Sand Spot
• Caatinga: Water, Urban Areas, Rocky Outcrop
• Cerrado: Water, Urban Areas and Mining
• Atlantic Forest: Water, Urban Area, Rice, Mining, Beach, Dune and Sand Spot
  • Addittional for regions 6 and 7: Soy, Temporary Crop; Sugar Cane and and others Temporary Crops.
• Pampa: Water, Urban Areas, Rice, Mining, Beach, Dune and Sand Spot, Soy, Other Temporary Crops, Mosaic of Uses;
• Pantanal: Soy, Cotton and Other Temporary Crops.

To obtain the information of the month in which the fire scar was mapped, postclassification processing was performed to retrieve the information of the date of the pixel that was burned from the date of the pixel in which the annual mosaic was constructed from the minimum NBR.

2.6. RATING ASSESSMENT

Evaluations of the classification of fire scars were carried out using Landsat images, with visual inspection, statistical analysis and relationship with land use and cover from MapBiomas. In addition, they were compared with reference maps, including FIRMS (1km), GABAM (30m), MODIS MCD64A1 (500m), FIRE CCI (250m) and INPE hotspots (1km).