|Title:||Single Image Super-Resolution for SAR Images||Language:||English||Authors:||Dimitrov, Philip||Qualification level:||Diploma||Advisor:||Sablatnig, Robert||Assisting Advisor:||Zambanini, Sebastian||Issue Date:||2021||Citation:||
Dimitrov, P. (2021). Single Image Super-Resolution for SAR Images [Diploma Thesis, Technische Universität Wien]. reposiTUm. https://doi.org/10.34726/hss.2021.88085
|Number of Pages:||114||Qualification level:||Diploma||Abstract:||
Single image Super-Resolution (SR) is a method to get a high-resolution image out of a single Low-Resolution (LR) image. SR is used in different domains, such as medical imaging, satellite imaging, and security imaging. Using SR compared to LR images speeds up training convergence and boosts recognition and segmentation accuracy. Apart from increasing the resolution of LR images, SR is able to denoise. Given the image acquisition hardware, there are physical restrictions on the image quality and resolution.SR helps overcome those limitations. This work focuses on the application of SR for Synthetic Aperture Radar (SAR) C-Band images captured by the Sentinel-1 satellites of the Copernicus Mission conducted by the European Space Agency. Earth Observation (EO) tasks, such as land cover estimation, detection of oil spills, land surface temperature,and soil moisture depend on the quality of the given remote sensing images. LR and noise impairs the underlying models, therefore SR is significant for earth science. This thesis investigates state-of-the-art SR approaches on SAR C-band images based on deep neural networks. An EO task for pixel-wise land cover segmentation is proposed in order to assess the suitability of SR for SAR images. Results of upscaling SAR images by a factor of 2 or 4 are evaluated based on image quality metrics (PSNR, SSIM) and an EO segmentation model. Furthermore, it is assessed if the SR methods can handle unseen temporal and spatial conditions and if adversarial training can further enhance the results.The final evaluation shows that SR for SAR C-band images is viable for upscaling by a factor of 2 and that unseen temporal and spatial conditions are manageable. In contrast,for SR by a factor of 4 to handle unseen temporal and spatial conditions, additional effort (re-training the EO model on the SR images) is required. Results indicate that adversarialtraining can improve both classification and image quality metrics. By keeping only the LR images, SR by a factor of 2 or 4 can reduce the necessary storage by a ratio of 4 or16, respectively. This work lays the ground for future research in the field of single image SR for SAR C-band images.
|Keywords:||Super Resolution; Synthetic Aperture Radar; Earth Observation||URI:||https://doi.org/10.34726/hss.2021.88085
|DOI:||10.34726/hss.2021.88085||Library ID:||AC16336807||Organisation:||E193 - Institut für Visual Computing and Human-Centered Technology||Publication Type:||Thesis
|Appears in Collections:||Thesis|
Files in this item:
checked on Oct 26, 2021
checked on Oct 26, 2021
Items in reposiTUm are protected by copyright, with all rights reserved, unless otherwise indicated.