R. Kavasseri
South Dakota State University, South Dakota, United States
Keywords: Synthetic Aperture Radar, Maritime Domain Awareness, AIS-Denied Tracking, Vessel Detection, Maritime Intelligence
Automatic Identification System (AIS) transponders, the primary mechanism for tracking vessels can be spoofed, jammed, or disabled, creating blind spots in contested waters. Existing SAR ship detectors often depend on intensity thresholds that vary with sea state, producing false alarms in rough conditions or missed detections under calm-water calibration regimes. We present a physics-informed residual framework for non-cooperative vessel detection using open Sentinel-1 SAR and ERA5 wind reanalysis. For each scene, expected ocean backscatter is estimated with CMOD7, the C-band geophysical model function relating VV-polarized return to wind speed, incidence angle, and wind direction. Then, a lightweight U-Net learns structured residuals between predicted and observed backscatter and outputs per-pixel uncertainty. Detection becomes a calibrated z-score test against the residual field, enabling one threshold across changing ocean conditions. Phase 1 validation across 27 Sentinel-1 scenes shows CMOD7-conditioned residuals are near-Gaussian (β=2.26, σ=0.599 dB) with a learnable 1.3 dB incidence-angle bias. In an April 30, 2026 Sentinel-1C Strait of Hormuz pass, corridor occupancy fell 63% versus May 2025 while total vessel count remained stable. A prototype application, Seahound, is under development toward deployment as a spoof-resistant radar verification layer for DOD maritime domain awareness and supply-chain security.