Wildland fires are one of the most important disturbances on Earth ecosystems, where the combination of remote sensing data and modern machine learning techniques offer a great potential for detection and monitoring of burned areas. Many algorithms have been developed in parallel to locate burned areas at local and global scales, but more accurate methodologies are still needed to provide precise and valuable information to climate users, environmentalists and public administrations. Gradient boosting-based methodologies have shown a good performance when dealing with segmentation of burned areas.
In this paper, we propose a new method, called MBAGB, based on these methodologies for detecting and mapping burned areas in a multi-temporal setting of satellite imagery.
We illustrate the procedure with the fires occurred in October 2017 in a region covering the North-Central Portugal and the North-West of Spain. Daily satellite images used for the definition of spectral indexes are taken from MODIS satellite products, between September and November 2017. The MBAGB accuracy metrics show the overall goodness of this method, and in particular, a very small number of false negatives in the identification of burned areas.