A new method for burnt scar mapping using spectral indices combined with Support Vector Machines

In present paper, to effectively improve the accuracy of burnt scar mapping in hilly areas, a new method by combing multiple spectral indices with Support Vector Machines (SVM) has been put forward. Firstly, the Landsat TM image was geometrically and atmospherically corrected. Secondly, two widely used thematic-oriented spectral indices, namely, the vegetation index SAVI and the water index MNDWI were derived. Thirdly, to minimize the confusion between burned areas and low reflectance objects (e.g., water and shadows) before next step, by adopting a looser criterion, the above mentioned indices were used to mask out as much as possible water and unburned vegetated pixels. Finally, images of four widely and successfully used vegetation indices specially designed for burnt area mapping, including BAI, VI3, NBR, and GEMI, together with SAVI and MNDWI, were gathered and stacked as input, and the SVM was applied to extract burnt scars. The resultant image was evaluated and a sound performance was achieved, with an overall accuracy of 90.1% and a Kappa coefficient of 0.8034.