Answers are grounded in this paper and cite their sources.
Synthetic Aperture Radar (SAR) can observe oil on the sea surface in all weather, but operational pipelines are built around Level-1 Ground Range Detected (GRD) products of roughly 0.7–1 GB per scene. This paper presents TARS (Timely Anomaly Response System), an event-driven framework in which Automatic Identification System (AIS) telemetry decides where to look, and the Sentinel-1 browse (quicklook) image, about 3,400 times smaller than the GRD product, is classified at that location. A three-layer AIS cascade combines a speed-change filter, HDBSCAN clustering with physical distance, speed and coastline checks, and an Isolation Forest with an explicit alarm budget. On 12,068 real vessels the spatial checks reduce possible-rendezvous flags from 38% (cluster size alone) to 0.09%, and on a day of US traffic the cascade escalates 2.2–2.6% of vessels per cycle. Against expert-labelled slicks from the Eastern Mediterranean and the Arabian/Persian Gulf, whole browse scenes are classified at close to chance (AUC 0.55) and aligning the input with full-resolution statistics does not help, whereas a 25 km crop centred on the vessel is effective. A classifier head trained on browse crops reaches AUC 0.909 within its region, against 0.638 for a model trained on full-resolution chips and 0.840 for gradient boosting on intensity statistics; a Gulf-trained head reaches 0.882 on a held-out year. Scores change little when the slick lies up to 5 km from the vessel. End-to-end latency is set by satellite revisit and publication (about two days), not by TARS processing (about 3 s).
No takes yet. Share an insight, caveat, or question.
Miranda et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: