Skip navigation.

Using radar backscatter and AI for better maps of burned area

ESA’s Climate Change Initiative Fire project team introduces a self-adapting algorithm for detecting fire burned area in a paper published online this month in Remote Sensing of Environment. The team say their proposal is particularly helpful for tracking the impact of fire on tropical forests, which are usually shrouded in cloud and difficult to study using optical satellite imagery.

Fires are a natural part of ecosystems but humans can also have a strong role in their frequency and severity, with logging linked to fire occurrence in the tropical forests of Indonesia, for example.

Monitoring changes in the global frequency and extent of land affected by fire is important to better understand fire’s contribution to the build-up of carbon dioxide and methane concentrations in the atmosphere, which cause global warming. The Global Climate Observing System (GCOS) identifies fire disturbance as an essential climate variable for characterizing the climate system. 

Techniques for mapping burned area remotely have typically relied on passive optical and thermal-wavelength sensors, which cannot observe areas obscured by clouds. This new algorithm uses Copernicus Sentinel-1 radar backscatter data to detect changes to the landscape caused by fire, with the advantage that it doesn’t depend on sunlight or cloud cover. 

The algorithm employs artificial intelligence (Random Forest Classifier) and multi-temporal change detection filters to adapt detection thresholds to local scattering conditions, rather than requiring pre-set thresholds to identify areas of burnt vegetation. 

To train the software, the team use active fire data taken from thermal channels of the MODIS sensor to determine thresholds. This enables them to reduce false positives, called errors of commission, produced by other sources of backscatter change such as clear-cutting forest or leaf die-back, says Emilio Chuvieco, Principal Investigator of CCI Fire based at University of Alcalá. The team cross-checked their results against an optical dataset covering over 21 million hectares taken from 18 locations that represent Earth’s major fire-prone biomes.  

The authors show that the proposed algorithm is able to detect small fires, down to one hectare, with greater detection accuracy compared to the current global burned area product, MCD64A1 V6 derived from the MODIS satellite.


Miguel A. Belenguer-Plomer, Mihai A. Tanase, Angel Fernandez-Carrillo, Emilio Chuvieco, Burned area detection and mapping using Sentinel-1 backscatter coefficient and thermal anomalies, Remote Sensing of Environment, Volume 233, 2019, 111345, ISSN 0034-4257,