SAR-based ship detection and vessel classification
Synthetic aperture radar detects vessels by their radar cross-section regardless of cloud, darkness or deliberate AIS silence. Polarimetric signatures and object geometry then allow classification by vessel type, with honest limits on small craft and fixed infrastructure ambiguity.
Sensors
- Sentinel-1 (ESA): C-band (5.405 GHz) SAR with IW mode ground resolution of approximately 5 x 20 m, dual-polarisation (VV+VH). Free and open archive back to 2014. Revisit over most ocean areas is 6 days with a single satellite; the two-satellite constellation achieved 3-day revisit before Sentinel-1B failed in 2021. Sentinel-1C launched in late 2024 is restoring that cadence.
- ICEYE: X-band SAR constellation of more than 20 satellites. Stripmap mode delivers approximately 3 m resolution; Spot mode reaches 1 m. Revisit to any point can be less than 24 hours with tasking. Particularly useful for detecting smaller vessels that fall below Sentinel-1's effective detection threshold.
- Capella Space: X-band SAR with Spotlight mode resolution down to approximately 0.5 m, enabling detection of vessel superstructure features useful for type classification. Tasking latency can be under 60 minutes in some orbits. Archive is commercial and relatively shallow compared with Sentinel-1.
- Umbra: X-band SAR offering published resolution down to 0.25 m in its highest-resolution Spotlight mode, the finest commercially available as of 2024. At that resolution, deck equipment, crane positions and hull geometry become distinguishable, substantially improving classification confidence for large vessels.
What a metal hull gives away
Radar energy scattered from the ocean surface follows Bragg scattering physics: small periodic waves return energy at predictable angles, producing the relatively dark, low-backscatter appearance of calm open water in a SAR image. A steel hull disrupts this entirely. Metal surfaces produce specular and dihedral reflections that return far more energy to the sensor, appearing as bright elongated targets against the dark sea. This contrast is the physical foundation of SAR-based ship detection, and it works identically at 3 a.m. in a North Atlantic storm as it does at noon in the tropics.
The radar cross-section (RCS) of a vessel scales roughly with its physical size and the angle of its surfaces relative to the radar look direction. A supertanker can present an RCS of tens of thousands of square metres; a fibreglass fishing dinghy might present less than one. That gap is where the method's honest limits begin.
Detection: what the algorithms actually do
The standard detection approach is constant false alarm rate (CFAR) processing. The algorithm compares each pixel's backscatter against a local background window, flagging pixels that exceed the background by a statistically significant margin. CFAR handles spatially varying sea clutter reasonably well, though high sea states and rain cells at C-band can raise the noise floor and suppress detections of smaller targets.
Sentinel-1's IW mode, at roughly 5 x 20 m, reliably detects vessels longer than about 30 m. Below that, detection probability degrades. ICEYE and Capella at 1-3 m resolution push that floor down considerably, but even at 1 m, a small wooden or fibreglass hull with low RCS may not clear the CFAR threshold. No SAR system reliably detects kayaks or small rigid inflatables from orbit. That is a physical limit, not a processing failure.
Classification: reading geometry and polarisation
Detecting a bright target is straightforward. Saying what kind of vessel it is requires more. Two sources of information are available in most modern SAR data: object geometry and polarimetric features.
Geometry means the length, width, orientation and aspect ratio of the detected target, plus any secondary bright returns from superstructure elements such as cranes, masts or deck containers. At Umbra's 0.25 m resolution, these features are genuinely legible. At Sentinel-1's 5 m range resolution, length estimation is feasible but width and superstructure detail are not. Polarimetric features come from dual- or quad-polarisation modes. The ratio of cross-polarised (VH) to co-polarised (VV) backscatter differs between vessel types because hull geometry and superstructure materials scatter energy differently across polarisation channels. Tankers with smooth hulls behave differently from container ships stacked with metal boxes. The differences are real but subtle, and classification accuracy reported in peer-reviewed literature typically sits in the 70-85% range for broad vessel categories (tanker, container, bulk carrier, fishing vessel) under good imaging conditions. Finer sub-type distinctions are harder and less reliable.
Machine learning approaches, particularly convolutional neural networks trained on labelled SAR chips, have improved on traditional feature-based classifiers, but they inherit the resolution constraints of the underlying data. A model trained on high-resolution X-band imagery does not transfer cleanly to Sentinel-1 C-band chips.
The wind turbine problem and other ambiguities
Fixed offshore infrastructure, particularly wind turbines, presents a persistent ambiguity. Steel monopole foundations and nacelles produce dihedral and trihedral reflections that are, in terms of backscatter intensity and spatial extent, nearly indistinguishable from small vessels in moderate-resolution SAR. A 100-turbine offshore wind farm generates 100 bright targets that a naive CFAR detector will flag as ships.
The standard mitigation is to cross-reference detections against a database of known fixed infrastructure positions. Vessels move between passes; turbines do not. Multi-temporal coherence analysis can also separate stationary from transient targets, though this requires image pairs from the same orbital geometry. Oil platforms, aquaculture cages and navigational buoys create similar false-positive problems. In congested coastal waters, these ambiguities are not edge cases; they are routine. Any operational detection system that does not explicitly mask known infrastructure will produce misleading outputs.
Combining SAR with AIS: what the gap reveals
SAR detection on its own tells you a vessel exists and roughly where it is. The more operationally interesting question is whether it should be there. Cross-referencing SAR detections against Automatic Identification System (AIS) broadcast positions identifies three categories: vessels that match (broadcasting and detected), vessels detected but not broadcasting (dark), and AIS positions with no corresponding SAR detection (spoofed or ghost signals). The dark-vessel fraction is the primary target for maritime surveillance applications.
AIS-SAR fusion is addressed in detail on the sibling page covering AIS vessel tracking and dark-ship detection. The point here is that SAR provides the independent physical observation against which AIS can be validated, not the other way around. SAR does not depend on a vessel choosing to transmit anything.
Satellize runs SAR-based maritime detection analytics for government clients, fusing open Sentinel-1 passes with commercial tasking where revisit or resolution requirements exceed what the free constellation can deliver.
Choosing the right SAR source for the job
Sentinel-1 is the right starting point for wide-area persistent monitoring over large ocean regions. It is free, the archive runs back to 2014, and 5 m resolution is sufficient for detecting and roughly sizing most commercial vessels. Its weakness is revisit: 3-6 days is too slow to track a vessel that is actively manoeuvring, and the 2021 loss of Sentinel-1B left gaps that are only now being closed.
ICEYE suits applications where revisit matters more than cost, such as monitoring a specific port or chokepoint daily. Capella and Umbra are the tools of choice when classification confidence matters, when the target is a specific vessel of interest, or when you need to distinguish a fishing vessel from a patrol boat in a contested area. The tradeoff is cost per image and the narrower swath of high-resolution Spotlight modes, typically 5-10 km, against Sentinel-1's 250 km IW swath. There is no single answer; the appropriate constellation depends on the area, the target size and the decision the data is meant to support.
Typical figures
| Spatial resolution (Sentinel-1 IW) | ~5 x 20 m (range x azimuth) |
| Spatial resolution (ICEYE Spot) | ~1 m |
| Spatial resolution (Umbra Spotlight) | Down to 0.25 m |
| Radar frequency | C-band (Sentinel-1, ~5.4 GHz); X-band (ICEYE, Capella, Umbra, ~9-10 GHz) |
| Revisit (Sentinel-1, single satellite) | 6 days at mid-latitudes; improving toward 3 days with Sentinel-1C |
| Revisit (ICEYE constellation) | Sub-24 hours to any ocean point with tasking |
| Minimum detectable vessel (Sentinel-1) | Approximately 30 m length under moderate sea state; smaller targets unreliable |
| Swath width | 250 km (Sentinel-1 IW); 5-15 km typical for commercial Spotlight modes |
| Archive depth | Sentinel-1: back to 2014 (open). Commercial X-band: variable, typically 2019 onwards |
| Delivery formats | GeoTIFF (detected target chips), GeoJSON (vessel position polygons), CSV (detection lists with metadata) |
Analytics Satellize can run
| Dark vessel detection layer | CFAR detection on SAR amplitude image, cross-referenced against AIS position database; unmatched detections flagged as dark | GeoJSON alert feed with vessel centroid, estimated length, time of detection and AIS-match status |
| Vessel density heatmap | Aggregated CFAR detections over multiple Sentinel-1 passes gridded to 0.1-degree cells | GeoTIFF raster updated per new pass, suitable for GIS overlay or web map ingestion |
| Vessel type classification | CNN classifier applied to high-resolution SAR chips (ICEYE or Capella); trained on open benchmark datasets such as SAR-Ship and HRSC | Detection report with per-vessel type probability scores and confidence flags |
| Fixed infrastructure mask | Multi-temporal coherence analysis plus cross-reference against published offshore wind and platform databases to suppress false positives | Updated infrastructure exclusion polygon layer, delivered as GeoPackage |
| Vessel length and heading estimation | Geometric fitting to detected bright-target extent in SAR amplitude; heading inferred from wake or target orientation relative to look direction | CSV attribute table appended to detection GeoJSON |
| Tasked surveillance of vessel of interest | Commercial SAR tasking (ICEYE or Umbra) on specified area of interest; Spotlight acquisition for maximum resolution; change detection between passes | Annotated image report with position, heading and visible superstructure features |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.