Unsupervised Detection for Burned Area with Fuzzy C-Means and D-S Evidence Theory

Abstract

Mapping the burned area of forest fires can contribute significantly to the understanding, quantification, and evaluation of forest fire severity and its impacts on the forest ecosystem. In this paper, an unsupervised detection for burned region based on the Fuzzy C-Means (FCM) and Dempster-Shafer (D-S) evidence theory with the bi-temporal images is proposed. Specifically, according to difference maps from the delta normalized burn ratio and spectral angle index, the Expectation-Maximization (EM) algorithm is used to separate the study area into the definitely burned region and indefinitely burned region. Then, under the enlightenment of the multi-source information fusion theory, the indefinite region is discriminated against further with FCM and D-S evidence theory. Finally, the final fire-burned map can be inferred from the results obtained from the aforementioned steps. The experimental results on two forest fires with bi-temporal Landsat-8 images have shown the potential of the proposed burned area mapping method, in the field of detecting the forest landscape change based on multispectral remote sensing images.