Object detection in ultra-high resolution satellite SAR data by adaptation of a YOLO model

This thesis focuses on the application of object detection techniques to high resolution SAR (Synthetic Aperture Radar) images acquired from ICEYE satellites, highlighting the peculiarities and intrinsic challenges of these data—such as acquisi tion geometry, double bouncing phenomena, and the presence of speckle—that make the detection of small targets particularly complex. In the first phase, the work analyzes the use of the YOLO family of algorithms, employed in various configurations (GRD and CSI images) and at different reso lutions, to evaluate performance in terms of speed and accuracy in object detec tion. Subsequently, an approach based on a Transformer foundation model, named Clay, is examined. This model is adapted to SAR images through a fine-tuning process supported by a custom loss function that integrates Complete Intersection over Union (CIoU) and Mean Squared Error (MSE) for bounding box regression, alongside components of Binary Cross Entropy (BCE) for classification. In parallel, a ResNet50-based backbone—trained using learning rate scheduling techniques—is also compared to explore the implications of extended training. The experimental results indicate that while the YOLO approach proves robust in managing the complexities of SAR data, methodologies based on Transformer models require further investigation to fully leverage the unique characteristics of radar imagery. In conclusion, the thesis underscores the complementarity between classical and innovative approaches, outlining future perspectives for optimizing ob ject detection systems in the SAR domain, with potential applications in Earth observation and the monitoring of critical infrastructures and activities.