Detection of forest fire in Menderes district using a superpixel segmentation based search method

Abstract

Forest fires are one of the most affected natural disasters in our country. According to the forest statistics shared by the Ministry of Forestry and Water Management of Republic of Turkey, 3188 forest fires are occurred just in 2016 [1]. After each forest fire, detection of affected regions is crucial with regard to land management and fast planning. In this respect, remote sensing technologies have become a popular topic. In this work, detection of forest fire regions are investigated using multispectral images which are acquired by Sentinel-2A satellite. For the detection of forest fire regions, classical spectral indices are used. Additionally, a novel method with coarse-to-fine search strategy is proposed. Firstly, forest fire regions are coarsely detected, and detailed regions are detected efficiently using a fine search step. In order to evaluate spatial and spectral information together, a superpixel segmentation based approach is used for both coarse and fine search step. The experimental results of proposed method and other methods are obtained for the forest fire located in Izmir-Menderes region in 1 July 2017 [2]. Obtained results are compared using receiver operating characteristics and the proposed method is found to provide better detection performance than the other methods in the literature.