Maritime Surveillance from Sentinel-1
A hybrid ship detection and segmentation system combining YOLO11 and SAM2 for vessel monitoring in Sentinel-1 SAR imagery.
Built a hybrid ship detection and segmentation system combining YOLO11 and SAM2 for vessel monitoring in Sentinel-1 SAR imagery, as part of postdoctoral work at UniBw Munich on operational EO exploitation.
Published as “Prompting Foundation Models for Zero-Shot Ship Instance Segmentation in SAR Imagery” (with Francescopaolo Sica and Michael Schmitt), presented at EUSAR 2026. A YOLOv11 detector trained on open SAR datasets localizes ships via bounding boxes, which then prompt the Segment Anything Model 2 (SAM2) to produce pixel-level instance masks — without any mask annotations, fine-tuning, or adapters. Decoupling detection (handled by a SAR-trained model) from segmentation (handled by an off-the-shelf foundation model) lets the bounding-box prompts alone regularize SAM2’s predictions across the optical-SAR domain gap.
On the SSDD benchmark (1,160 images from Sentinel-1, RADARSAT-2, and TerraSAR-X; 2,358 annotated ships), the pipeline reaches a mean IoU of 0.637 (Dice 0.764) — about 89% of a fully supervised baseline — with an overall ship detection rate of 89.2% (92.0% offshore, 87.8% inshore), using only bounding-box supervision, no pixel-level mask labels at all.
