Land Cover Classification for Synthetic Aperture Radar Imagery by Using Unsupervised Methods

Land cover classification is an important application in remote sensing and it plays a critical role in urban planning, land cover change monitoring and agricultural monitoring. Synthetic aperture radar (SAR)has long been recognized as an effective sensing tool for land cover monitoring, because of its ability of capturing images day and night without affected by weather conditions. On the other hand, the interpretation of SAR imagery and to get many labeled SAR images are still a challenging problem for remote sensing. Therefore, the aim of this paper is the implementation of unsupervised classification methods which require unlabeled data and comparison of their performances. In order to achieve this goal, Vertical - Vertical (VV)and Vertical - Horizontal (VH)polarization Sentinel-l SAR images are used. The acquisition year of these images is 2018. Principal Component Analysis (PCA), Kernel PCA, Eigenface and Autoencoder feature extraction methods and also user defined features are implemented for unsupervised classification and the results have been reported and discussed. The performances of these methods are compared by using the cluster validity indices which is the criterion for unsupervised classification, and also the optimum cluster number is determined. Results demonstrate that, KPCA and Autoencoder methods are better for VH polarization. Also, user defined features and Autoencoder are better for VV polarization.