Mapping burnt areas in the semi-arid savannahs: an exploration of SVM classification and field surveys

Abstract

Accurate and precise mapping of burnt areas in the fire-prone semi-arid savannah zones is relevant for making decisions on the management of environmental resources. In this study, we apply the Support Vector Machine (SVM) algorithm to freely available Sentinel-2A&B data to delineate and map burnt areas and compare our findings with conventional field surveys to enable resource managers decide on the most appropriate method to use when seeking to rapidly conduct post-fire assessment. We surveyed three burnt patches of varying sizes and compared with estimates from the SVM classification algorithm. Accuracy assessment was based on reference data collected from field surveys. We obtained an average overall accuracy of 94.8% ± 5.2% for all kernel functions in the SVM. The classification estimated the average total of the three patches at 42.99 km2 but with variations among the different kernels, while the field measurements produced 42.29 km2. A follow-up field survey showed that the earlier survey either over- or under-estimated the burnt patches. Our micro-level analysis demonstrates that any kernel function in the SVM algorithm can be used with freely available remote sensing data to accurately and cost-effectively map wildfire hazards, especially in resource-poor settings, for efficient decision making when managing environmental resource.