On May 20, 2025, Janik Hoffmann will present his Master thesis on Deep Learning-driven SAR Image Super-Resolution” at 13:00 in seminar room 3, John-Skilton-Str. 4a.
From the abstract: Synthetic Aperture Radar (SAR) is a microwave remote sensing system that is used for various applications of land and water surface monitoring including target detection or scene classification in a variety of contexts such as military surveillance, crisis management or environmental monitoring. This thesis employs a Convolutional Neural Network with residual learning to increase the spatial resolution of TerraSAR-X StripMap imagery based on High-Resolution SpotLight reference images. Furthermore, a point target detection algorithm was applied to the super-resolved StripMap data to evaluate if super-resolution can enhance the number of homologous image points in a SAR stereo pair. A quality assessment of the super-resolution results on point targets revealed significant increases in PSLR and spatial resolution by a factor of up to two. This clarifies the applicability of modern super-resolution networks, that were actually developed for processing natural RGB images, to SAR data. Furthermore, it shows that super-resolution can potentially bridge the resolution gap between StripMap and High-Resolution SpotLight imagery to facilitate applications that require large coverage and higher resolution at the same time. This was approved by testing the super-resolution data for point target detection which is a critical part in GCP retrieval as a typical SAR application. It was shown that the number of identified homologous image points in a stereo image dataset could be enhanced. Future studies should test more powerful deep learning algorithms for super-resolution that are more explicitly tailored to a specific task and implement super-resolution algorithms in the processing workflow of GCP retrieval.
1st supervisor: Prof. Dr. Tobias Ullmann
2nd (external) supervisor: Dr. Wolfgang Koppe, Airbus Defence and Space GmbH
From the abstract: Synthetic Aperture Radar (SAR) is a microwave remote sensing system that is used for various applications of land and water surface monitoring including target detection or scene classification in a variety of contexts such as military surveillance, crisis management or environmental monitoring. This thesis employs a Convolutional Neural Network with residual learning to increase the spatial resolution of TerraSAR-X StripMap imagery based on High-Resolution SpotLight reference images. Furthermore, a point target detection algorithm was applied to the super-resolved StripMap data to evaluate if super-resolution can enhance the number of homologous image points in a SAR stereo pair. A quality assessment of the super-resolution results on point targets revealed significant increases in PSLR and spatial resolution by a factor of up to two. This clarifies the applicability of modern super-resolution networks, that were actually developed for processing natural RGB images, to SAR data. Furthermore, it shows that super-resolution can potentially bridge the resolution gap between StripMap and High-Resolution SpotLight imagery to facilitate applications that require large coverage and higher resolution at the same time. This was approved by testing the super-resolution data for point target detection which is a critical part in GCP retrieval as a typical SAR application. It was shown that the number of identified homologous image points in a stereo image dataset could be enhanced. Future studies should test more powerful deep learning algorithms for super-resolution that are more explicitly tailored to a specific task and implement super-resolution algorithms in the processing workflow of GCP retrieval.
1st supervisor: Prof. Dr. Tobias Ullmann
2nd (external) supervisor: Dr. Wolfgang Koppe, Airbus Defence and Space GmbH
