An Automatic Process Chain for Detecting Burn Scars Using Sentinel-2 Data

Abstract

The advantage of the new Sentinel-2 mission is its high geometric resolution (10 m) with a high temporal resolution (~ 5 days) at the same time. To make use of this for the monitoring of burn scars, this paper presents a new method for the automatic detection of burn scars on the basis of Sentinel-2 data. In order to develop a globally applicable method, 108 mono- and bi-temporal features of seven global distributed test sites are analysed in a discriminant analysis to examine their potential for the identification of burned areas. Based on these results, a two-phase algorithm is proposed. The first phase serves to identify seed pixels which have a high probability to be related to burned areas. Due to its simplicity, a threshold-based approach is used for this purpose. In the second phase, a refinement of the burned areas is accomplished by examining the neighbourhood of the seed pixels through a region growing method based on Support Vector Machines. The validation of the proposed approach shows encouraging high user accuracies of 95 - 99%. The producers accuracies vary between 80 - 90%.