OBJECT-BASED BURNED AREA MAPPING USING SENTINEL-2 IMAGERY AND SUPERVISED LEARNING GUIDED BY EMPIRICAL RULES

This paper presents a methodology for burned area mapping using Sentinel-2 imagery, which tries to minimize-and conditionally eliminate-user interaction. The methodology employs an object-based image analysis approach, using the Mean-Shift segmentation algorithm. A small portion of representative image object is automatically selected to form the training set, by means of the fuzzy C-means (FCM) clustering algorithm. Subsequently, a pre-fire and a post-fire image are used for calculating a number of well-known burned area indices and their difference is employed for labeling a portion of the selected training patterns (the most unambiguous ones) through a set of empirical rules. The user can subsequently classify any remaining training patterns or accept the automated classification, which is performed through the Support Vector Machine (SVM) classifier. The latter considers the subset with the most informative object-level features, which are obtained by means of a supervised feature selection algorithm.