Feature selection for burned area classification in the Castelo de Paiva region
Nov 2018
Burned area mapping is a fundamental activity for the study of wildfires, but good quality data requires
long in situ campaigns performed by specialized personnel. The use of machine learning algorithms and the
availability of high-quality remote sensing data bring new opportunities to make the process more precise
and expedite. This paper elicits the best combinations of features to be used in machine learning pixel-based
burned area mapping, obtained by LASSO regression. The feature selection is applied to ground-truth
obtained in the region of Castelo de Paiva, Portugal, in 2016.
The selected number of features is 9, from an initial set of 51 features, and includes besides a new index
a mix of prefire, postfire as well delta differences of spectral indices (namely dNBR). The selected features
can be used with Sentinel 2, Landsat, and MODIS imagery.
In parallel, the identified features have been fed to several classifiers, namely a multilayer neural network,
gradient boost and extreme gradient boosting, support vector machines and K-nearest neighbour classifiers
to validate the choice performed by the LASSO regression. Accuracies of 96% and kappa of 0.86, are
consistently obtained for Sentinel 2 imagery, and Landsat 8 also scores very well; as expected, MODIS results
in score reduction to the coarser spatial resolution. Confusion maps are presented to visualise the quality of
the obtained results, as well as to pinpoint existing problems with the ground-truth.
Extreme Gradient Boosting shows to combine very high classification metrics with very efficient
processing, possible via the Graphics Processing Units (GPU) implementation of XGBoost. A simple
processing architecture is proposed to support an automatic classification system based on the publicly
available satellite imagery, supported by the benchmarks obtained in this paper. A set of Sentinel 2 granules
covering Portugal can be processed in less than 10 minutes in consumer hardware, for each satellite passage.