ORIGINAL PAPER Comparative analysis of unsupervised classification methods for mapping burned forest areas Dilek Küçük Matcı 1 & Uğur Avdan 1 Received: 28 September 2018 /Accepted: 2 July 2020 # Saudi Society for Geosciences 2020 Abstract Forest fires cause a lot of damage in Turkey, due to the geographical location, high temperatures, and low humidity levels. Accurate determination of burned forest areas is crucial for correct damage assessment studies, fire risk calculations, and review of the forest regeneration processes. In this study, we compare the performances of unsupervised classification methods (which have not been used to map burned areas before) of burned area extraction from medium resolution satellite images with K-means. In this regard, the areas affected by fire in the Kumluca and Adrasan regions in 2016, Alanya and Gümüldür regions in 2017 and Athens region in 2018 are determined using Landsat 8 images. For this purpose, Canopy, M-tree, a hierarchical clustering algorithm, and a learning vector quantization which are frequently used in the literature are applied to determine the burned area, and the results obtained are compared with the results obtained from K-means. The results show that unsupervised classification methods can be used to map burned areas. The hierarchical clustering and K-means algorithms provide the most accurate results in mapping burned areas in most of the regions used in the study. Keywords Unsupervised classification methods . Burned area mapping . Landsat 8 . Classification Introduction Forest fires, caused by extreme heat, lightning, burning wood- land, burning stubble, or deliberately burning forests to open space for settlement, cause millions of hectares of forest to disappear every year. Forest fires have a direct and indirect impact on natural resources and ecosystem flexibility and cause economic losses (Camia et al. 2017). According to the statistics of the Turkish General Directorate of Forestry and Water Affairs Ministry of Forestry, 61,313 forest fires occurred in Turkey between 1988 and 2016, burning 307,855 ha of land (OGM 2017). Mapping the damaged forest areas, assessing economic losses and environmental impacts caused by a fire, and monitoring land cover changes are important for modelling the atmospheric and climatic conse- quences of a fire. Reliable and effective methods of monitoring and analysis should be used to assess the impact of fire on the ecosystem (Polychronaki and Gitas 2012). With the development of satellite technologies, the amount of data provided by remote sensing systems is increasing day by day. The correct local use of this data could save lives in many ways, through risk analysis, disaster management, re- search, post-disaster damage assessment, and change detec- tion (Chuvieco 2009). Image classification is one way to obtain useful data from satellite images. In the literature, supervised classification methods are frequently used to map the areas affected by forest fires (Chen et al. 2016; Gitas and Devereux 2006; Hudak and Brockett 2004; Miller and Yool 2002; Rogan and Yool 2001; Sedano et al. 2013). In this method, training data determined by the user is used to classify images and the operation is performed according to this data. Since the suc- cess of this method is dependent on the training data, it is necessary that the classification area is well known. The unsupervised classification approach, another method used to obtain data from satellite images, eliminates this hand- icap of supervised methods. The unsupervised classification approach is based on analysing the spectral values of pixels in Responsible Editor: Biswajeet Pradhan * Dilek Küçük Matcı dkmatci@anadolu.edu.tr Uğur Avdan uavdan@eskisehir.edu.tr 1 Institute of Earth and Space Sciences, Eskisehir Technical University, Iki Eylul Campus, 26470 Eskisehir, Turkey https://doi.org/10.1007/s12517-020-05670-7 / Published online: 22 July 2020 Arabian Journal of Geosciences (2020) 13: 711