Delineation of burned arid landscapes using Landsat 8 OLI data: a case study of Karaganda region in Kazakhstan

Abstract

In arid lands of Central Kazakhstan, fires commonly occur due to both man-made (e.g. space rocket launches) and natural factors. Remote sensing is the best method to identify burned landscapes and to assess their long-term dynamics. In this paper, we assessed total areas and seasonal features of fires using Landsat 8 OLI data of high spatial and temporal resolution, corrected for atmospheric effects. We defined data requirements, considered various spectral indices, and evaluated their suitability for automated delineation of burned areas. Spectral indices included special burn indices (NBRT, MIRBI, NBR, NBR2, BAI) and vegetation indices (MSAVI, MSAVI2, MTVI2, GEMI3) that allowed us to determine the boundaries and severity of fires indirectly based on the state of vegetation. Our study showed that burned areas were most accurately identified using indices of NBR2 (normalized burn ratio 2) and MSAVI2 (modified soil adjusted vegetation index 2). They were extracted using image segmentation and natural breaks classification methods. We have assessed MSAVI2 and NBR2 segmentation results with machine learning metrics, i.e. for Soyuz drop zone (landing area of rocket stages) precision of NBR segmentation equaled 99.5%, recall − 99.5%, accuracy − 98.5%. For Proton drop zone NBR segmentation precision equaled 99.3%, recall − 99.7%, accuracy − 99.4%. MSAVI2 results are less accurate. The accuracy of segmentation significantly depended on the vegetation state which was predetermined by the frequency and severity of previous fires within certain territory.