Evaluating Classification Models in a Burned Areas' Detection Approach

We present a study to improve automation and accuracy on a Woody Savannah burned areas’ classification process through the use of Machine Learning (ML) classification models. The reference method for this is to extract polygons from images through segmentation and identify changes in polygons extracted from images taken from the same area but in different times through manual labeling. However, not all differences correspond to burned areas: there are also deforestation, change in crops, and clouds. Our objective is to identify the changed areas caused by fire. We propose an approach that employs polygons’ attributes for classification and evaluation in order to identify changes caused by fire. This paper presents the more relevant classifier models to the problem, highlighting Random Forest and an Ensemble model, that achieved better results. The developed approach is validated over a study area in the Brazilian Woody Savannah against reference data derived from classifications manually done by experts. The results indicate enhancement of the methods used so far, and will eventually be applied to more data from different areas and biomes.