SAR2SAR DENOISE METHOD ON LAND USE AND LAND COVER IN MALAYSIA USING SENTINEL-1 IMAGERY

Although synthetic aperture radar (SAR) images are often regarded as greyscale images, special care must be exercised when interpreting the images for land use and land cover (LULC). SAR has long been regarded as a viable alternative to optical images because it interacts with ground features in a variety of ways and is less influenced by weather conditions. Unlike optical images, SAR images suffer from speckle noise; therefore, accurate LULC mapping with SAR data is important. In this work, the recently proposed SAR2SAR denoise method has been employed. The self-supervision method is based on a deep learning model that can generate a speckle-free image with few references in a short amount of time. The proposed method has been applied to evaluate five categories of LULC in Malaysia with ground range detected (GRD) Sentinel-1 data: dense forests, paddy fields, urban areas, cleared lands, and water bodies. The results showed that the use of false-color-based denoised VV and VH composites from SAR2SAR denoise method significantly improved the visualizations of LULC classes as much as optical imagery.