TRACE

A satellite shows what is on the water; AIS shows what ships report about themselves. TRACE puts both on one map.

System

• Vessels: YOLOv8-OBB trained on DOTA, run over map tiles.
• Oil spills: U-Net with a ResNet-34 encoder trained on a SAR oil-spill dataset.
• AIS: an XGBoost model over 7.17 million transponder records, plus a rule-based risk engine.
• Interface: FastAPI and Leaflet, with short reports written by a vision–language model.

Outcome

1st place at the AEROO International Space AI Competition.

Limits

The AIS anomaly labels were produced by hand-written rules (gaps, speed, turns, status), so the very high anomaly scores in the README measure how well the model learns those rules, not how well it finds real incidents. Tracking, route prediction and multi-image analysis are unfinished.