MapBiomas Fire Method
Here we present a summary of the method developed and applied in Collection 1 of the mapping of fire scars in Brazil (1985-2020).
For further methodological details, access the ATBD (Algorithm Theory Base Document) on this LINK
- GENERAL CHARACTERISTICS
The entire mapping of fire scars in Brazil was based on mosaics of images from Landsat satellites with a spatial resolution of 30 meters. The mapping period was from 1985 to 2020, with monthly and annual data on fire scars covering the entire Brazilian territory.
The entire process was carried out collaboratively between MapBiomas institutions and with artificial intelligence based on the use of machine learning algorithms (deep learning) on the Google Earth Engine and Google Cloud Storage platform, which offer immense processing capacity in the cloud.
The work was organized by biomes and regions, with samples collected in burned and unburned areas for training the algorithm by regions, in addition to the use of reference maps, such as burned areas of the MODIS product (MCD64A1 - https://lpdaac .usgs.gov/products/mcd64a1v006/) of 500 m of spatial resolution and heat source data from INPE (https://queimadas.dgi.inpe.br/).
- METHOD OVERVIEW
The first step consists of defining the regions by biome for classification. The second stage consists of building the annual mosaics for classification, and the third stage of collecting training samples from burned and unburned areas in Landsat image mosaics on the Google Earth Engine platform. For the classification step, the Google Cloud Storage Bucket platform was used, where the mosaics for classification and the training samples were stored, in addition to the use of virtual machines for processing the classification algorithms.
The diagram below illustrates the main steps and platforms used in the process of classification of fire scars in Brazil.
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 objective of obtaining a more precise 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 image mosaics of Landsat satellites (Surface Reflectance) (30 x 30 meters) constructed for each of the years 1985 to 2020. All available scenes from Landsat 5 (from 1985 to 1998 and from 2003 to 2011), Landsat 7 (1999 - 2002 and 2012) and Landsat 8 (2013 - 2020), all with an interval of 16 days.
2.3. SAMPLE COLLECTION
This step consists of collecting polygons of training samples from burned and unburned areas in the annual Landsat mosaics for each region of the biomes and for a sensor (Landsat 05/Landsat 07 and Landsat 08).
Thus, a sample set per sensor and for the 21 regions of Brazil was obtained, to be used in the training of the classification model.
The classification model used was the Deep Neural Network (deep neural networks) which consists of computational 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 burnt and unburnt training samples, the following spectral bands were used as input to the burnt area classification model: red (RED - 0.65 µm), near infrared (NIR - 0.86 µm) and short wave infrared (SWIR 1 - 1.6 µm and SWIR 2 - 2.2 µm). In addition to these selected spectral bands, the Landsat bands 2 to 7 and the NDVI (Normalized Difference Vegetation Index), NBR (Normalized Burn Index) and Delta NBR indices were tested. These spectral Landsat bands were chosen based on their sensitivity to fire events. The training data input was divided into two sets: 70% of the samples used for training and 30% for testing.
After training and testing the model, the classification was applied with Landsat images fortnightly for the entire period of analysis (1985 to 2019). A spatial filter was applied to remove noise and fill small empty gaps: areas smaller than or equal to 1.4 hectares (16 pixels) were removed, and empty gaps smaller than or equal to 5.8 ha (64 pixels) were filled as areas burned.
After evaluating the classification results, post classification filters were also applied, removing pixels that were in the following land cover and land use classes from the MapBiomas Collection 6 in the biomes:
Amazon: Water, Urbanized Area and Rocky Outcrop;
Caatinga: Water and Rocky Outcrop;
Cerrado: Water and Urbanized Area;
Atlantic Forest: Water, Urbanized Area and Rice;
Pampa: Rice, Soybeans, Other Temporary Crops, Urbanized Area and Water;
Pantanal: Water, Soybeans and Other Temporary Crops;
As Deep Learning methods require large computational processing, implementations were based on graphics processing units (GPUs) and specialized hardware components to perform parallel arithmetic operations. Access to GPUs in a virtual machine environment was implemented on the Google Cloud Platform (https://console.cloud.google.com), a set of cloud computing services provided by Google.
To obtain the information of the month the fire scar was mapped, post-classification processing was performed to retrieve the date information of the pixel that was burned, from the date of the pixel in which the annual mosaic was constructed from of the minimum NBR.
2.6. RATING ASSESSMENT
Evaluations of the classification of fire scars were carried out with monthly Landsat images, with visual inspection, statistics and relationship with land use and land cover, in addition to comparison with reference maps (MODIS and INPE).
The validation of the burnt area mappings was carried out by biome, considering the years 2007, 2011 and 2019. Each of the biomes was divided into areas of 2 km by 2 km, spatially integrated with the data of burnt areas from the FIRMS (Fire Information for Resource Management System - https://firms.modaps.eosdis.nasa.gov/), for the same years, where the sampling units were grouped into 4 categories: without fire occurrence, with up to 32% fire occurrence, with occurrence of fire between 32% and 70% and with occurrence of fire above 70%.
Then, the validation sampling units for each category, year and biome were randomly selected.
For each selected sampling unit, a Landsat image of minimum annual NBR was extracted. Then, automatic segmentation of all images from the selected sampling units was performed. The segments were interpreted in order to clarify which had the occurrence of fires. After the interpretation of all segments, their centroids were obtained (validation points)
The validation points (fire and non-fire) were spatially integrated with the MapBiomas Fogo mappings, making it possible to assess the qualities of the mappings and obtain the accuracy of the classification.