TranSAR - Transfer Learning from Medium-Resolution to High-Resolution SAR Imagery

• Applicant: Professor Dr.-Ing. Michael Schmitt
• Subject Area: Geodesy, Photogrammetry, Remote Sensing, Geoinformatics, Cartography
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 552338883

Deep learning for the analysis of image data is mainly driven by the desire to analyze optical imagery, but deep learning for synthetic aperture radar (SAR) data is also gaining more and more momentum. However, the well-known problem of a limited availability of accurately labeled data is even more severe in radar remote sensing than in optical remote sensing due to two main reasons: firstly, SAR images differ strongly from the human optical perception and secondly, SAR data have been less easily available for the public than optical data in the past. While data acessibility has strongly improved with the implementation of the Sentinel-1 SAR mission in the frame of the European Copernicus program, the lack of annotated data remains for very-high-resolution (VHR) SAR data. This project aims to leverage medium-resolution Sentinel-1 C-band SAR imagery, transfer learning, and physics-aware modeling to develop a solution to the label scarcity for VHR X-band SAR data as provided by modern commercial missions such as TerraSAR-X, Capella, or ICEYE. Such VHR data is often used for disaster response mapping, and companies often support the work of NGOs in dedicated data sharing programs.

• DFG Programme: Research Grants
• International Connection: India, Romania
• Cooperation Partners: Professor Dr. Mihai Datcu; Professor Dr. Sudipan Saha