Land Cover and Land Use Classification of Patch-Based Denoised SAR Images


  • Elif Meseci
  • Caner Ozcan Karabuk University
  • Buse Yaren Tekin


Synthetic Aperture Radar (SAR), Remote Sensing, Deep Learning, Despeckling, Land Classification, Image Processing


Land cover and land use information contributes to the research of important environmental issues such as the change of forest areas on the earth over the years, the determination of the increase and decrease in the amount of water, the detection of irregularities on the earth, the determination of damage after natural disasters. Synthetic Aperture Radar (SAR) systems, which can provide images in all weather conditions without being affected by changes in weather events, are preferred for obtaining images. Classification of SAR images is crucial to the analysis of these images. Developing technology allows the recording and interpretation of many high-dimensional SAR images. Since SAR images represent large areas, the objects in the image take up a very small area. The study includes the use of a patch-based approach to classifying these areas. In the proposed study, despeckling was applied to remove speckle noise which may occur in SAR images due to scattering. Denoised and original SAR images were classified with the deep learning algorithm, and comparative results of the classification performance of the noise filtering were obtained. Experimental results have proven that the despeckling significantly impacts classification performance.