Data-Efficient Spotlight SAR Image Formation for Ship Detection

Sammanfattning

Spaceborne Synthetic Aperture Radar (SAR) provides a method for producing high-resolution images of the Earth’s surface, independent of meteorological conditions and external light sources. However, forming these images is generally a computationally expensive process. This is particularly important for on-board and real-time processing where memory usage, processing time, and power consumption are limited. In this thesis, the effect of reduced raw input data and algorithmic parameters in a Polar Format Algorithm (PFA)-based pipeline is investigated with respect to final image quality and the performance of a YOLOv8s-detection model. The evaluated configurations include varying aperture fractions, decimation in both data dimensions, and different interpolation kernels. The formed images are evaluated using established image-quality metrics and runtime measurements. Finally, a deep learning model is applied to assess how the different configurations affect ship detection performance. The results show that moderate aperture reduction can preserve visual image quality and ship detection performance while significantly reducing computational cost. Decimating the data does not exhibit the same behavior, even when the amount of data reduction is comparable, and instead leads to a substantial decrease in both visual quality and detection performance. This indicates that not only the amount of data, but also which data subset is used, matters for the final image quality. Furthermore, the image-quality metrics do not align with the visual and detection-based results, suggesting that they are not sufficient on their own for evaluating image quality in this specific application. In conclusion, the results demonstrate the trade-off between computational cost and task-relevant SAR image quality, supporting future work on streaming on-board SAR processing.