“Small object detection in satellite SAR images”
Synthetic-Aperture Radar (SAR) is a powerful technology for generating images from radar signals. Renowned for its reliability in remote sensing, SAR functions independently of external light sources, enabling day and night monitoring, and can also penetrate clouds and foliage. The recent rise of commercial providers offering very-high resolution satellite SAR images has unlocked new opportunities in monitoring, including small object detection. However, interpreting SAR data is notoriously challenging, requiring specialised expertise, and the large spatial resolution makes manual analysis slow and labour-intensive. The complexity of this task, combined with its high-value applications, makes it an ideal candidate for AI-driven solutions.
In this talk, we introduce an innovative vision transformer-based, self-supervised, end-to-end SAR object detection model. Our approach stands out by incorporating auxiliary binary semantic segmentation during the fine-tuning phase, enhancing its effectiveness particularly in detecting small objects.”
Bio
Andrzej Kucik is an AI research engineer at Helsing – European defence company aimed at protecting liberal democracies. Before joining Helsing, he was a researcher and an engineer at the European Space Agency (ESA) where he worked on AI applied to Earth observation. He has MA degree in Mathematics from the University of Aberdeen, MPhil in Computer Science from the University of Manchester, and PhD in Mathematics from the University of Leeds. His interests include computer vision, remote sensing, efficient AI, logic and automated reasoning, neuromorphic computing and sensing, as well as LLMs.