SAR-based vessel detection in open ocean
Synthetic aperture radar detects metal hulls at sea regardless of cloud cover or darkness, using backscatter physics that no vessel operator can suppress. This page explains how CFAR algorithms, sensor geometry, and sea-state interact to determine what gets found and what does not.
Sensors
- Sentinel-1 (ESA, Copernicus): C-band (5.405 GHz) SAR. Interferometric Wide (IW) mode: 5 × 20 m resolution, 250 km swath, free and open data. Extra Wide (EW) mode: 20 × 40 m, 400 km swath, suited to ocean surveillance at the cost of resolution. Repeat pass every 6 days per satellite; combined A/B constellation (when both operational) gives roughly 3-day revisit at mid-latitudes, shorter at high latitudes.
- ICEYE (commercial constellation): X-band (9.65 GHz) SAR. Strip-map mode at approximately 3 m ground resolution; spot mode at approximately 1 m. Rapid revisit through a growing constellation; tasked passes can be delivered within hours of collection. X-band is more sensitive to smaller targets than C-band at equivalent resolution, but also more affected by rain attenuation.
- Capella Space (commercial): X-band SAR. Spotlight mode achieves approximately 0.5 m resolution, sufficient to resolve deck structure on larger vessels. Sliding spotlight and strip-map modes trade resolution for coverage. Tasked collection latency depends on constellation geometry but is typically sub-12 hours for most ocean regions.
- RADARSAT Constellation Mission (CSA): C-band, three-satellite constellation providing daily coverage of Canada's exclusive economic zone and near-daily global access. Medium-resolution mode at 16 m; compact polarimetry available, which improves discrimination between vessel returns and wave clutter. Operated by the Canadian Space Agency with data available under licence.
Why radar sees what optical sensors miss
Cloud cover obscures roughly 67 percent of the ocean surface at any given moment, which makes optical sensors unreliable for persistent maritime surveillance. SAR illuminates its own scene with microwave pulses and records the energy scattered back. Metal ship hulls, with their flat vertical surfaces and right-angle junctions, act as corner reflectors and return energy far more efficiently than the surrounding sea surface. The ratio between hull return and background clutter is the foundation of every detection algorithm in this field.
The physics does not change with time of day or weather. A 100-metre steel-hulled vessel in calm seas produces a radar cross-section that can exceed 40 dBsm at C-band. That figure drops significantly for wooden or fibreglass hulls, which is a genuine and important limit discussed below.
How CFAR detection actually works
Constant False Alarm Rate detection is the standard algorithmic approach for isolating vessel returns from ocean clutter. The core idea is local: for each pixel under test, the algorithm estimates the statistical distribution of the surrounding pixels (the clutter window), sets a threshold at a fixed multiple of that estimated noise floor, and flags the test pixel as a detection if it exceeds the threshold. The multiplier is chosen to hold the false alarm rate at a predetermined level across the image, regardless of how clutter intensity varies spatially.
Several CFAR variants are in operational use. Cell-Averaging CFAR (CA-CFAR) works well in homogeneous clutter but degrades near coastlines or in heavy ship traffic, where neighbouring pixels may themselves contain targets. Order-Statistic CFAR and its variants are more resistant to these edge cases. The European Maritime Safety Agency's CleanSeaNet service, which runs on Sentinel-1 data, uses CFAR as its primary detection step, and the method is documented extensively in ESA's Sentinel-1 Algorithm Theoretical Basis Documents.
Detection probability for a given vessel size depends on three interacting factors: the sensor's spatial resolution (which determines how much of the hull's cross-section falls within a single resolution cell), the polarisation channel used (VV polarisation is generally preferred for ship detection over open ocean because it suppresses wind-driven Bragg scattering more effectively than VH), and the sea state at the time of collection.
Sea state is the variable that forecasts cannot fix
At low sea states (Beaufort 1-3), the ocean background is dark in SAR imagery and even modest vessels stand out clearly. As wind speed rises above roughly 10-12 metres per second, wave breaking generates foam and increases surface roughness, raising the clutter floor and reducing the signal-to-clutter ratio for smaller targets. In Beaufort 6-7 conditions, a 20-metre vessel can fall below the detection threshold in Sentinel-1 IW imagery even with optimised CFAR settings.
This is not a solvable problem with better algorithms alone. It is a physical constraint. Analysts working in high-latitude or monsoon-affected ocean regions should plan collection windows around forecast sea state and treat detections during rough weather as a lower bound on actual vessel density.
The resolution floor and the small-vessel problem
Sentinel-1 IW mode at 5 × 20 m resolution can reliably detect vessels longer than approximately 30-40 metres in calm conditions. Below that size, the vessel's radar cross-section may not fill a resolution cell sufficiently to clear the CFAR threshold. Wooden and fibreglass hulls, common in artisanal fishing and small-boat smuggling, scatter radar energy diffusely rather than reflecting it back coherently. Their effective radar cross-section can be 20-30 dB lower than a steel hull of the same physical size.
Commercial X-band sensors close part of this gap. ICEYE strip-map at 3 m resolution and Capella spotlight at 0.5 m resolution can detect smaller targets and, at the finest resolutions, begin to resolve structural features that help distinguish vessel types. The trade-off is cost and revisit: tasked commercial collections are expensive relative to free Sentinel-1 passes, and a single spotlight image covers a small area. For broad-ocean surveillance, Sentinel-1 remains the workhorse; commercial SAR is the scalpel applied once a target is already of interest.
Practically speaking, no current spaceborne SAR system reliably detects a standard 10-metre fibreglass fishing boat in moderate sea state. Analysts should be explicit about this limit in any product that claims to characterise small-vessel activity.
Polarisation, incidence angle, and what they change
SAR systems transmit and receive in horizontal (H) or vertical (V) polarisation, or combinations. For open-ocean vessel detection, VV (transmit vertical, receive vertical) is generally preferred because the ocean's Bragg resonance scattering, which creates clutter, is weaker in VV than in HH at typical incidence angles. Sentinel-1 acquires dual-polarisation (VV+VH) in IW mode over maritime areas, which allows analysts to compute polarimetric ratios that help separate ship returns from wave features.
Incidence angle also matters. Sentinel-1's IW mode covers incidence angles from roughly 29 to 46 degrees across its three sub-swaths. Vessels near the near-range edge of the swath appear brighter than the same vessel in far-range, because the geometry favours corner-reflector returns at steeper angles. RADARSAT Constellation Mission's compact polarimetry mode provides additional discrimination capability by encoding more information about the scattering mechanism, which can reduce false alarms from wave features that mimic small vessels.
Satellize runs CFAR-based detection pipelines on Sentinel-1 open-archive data and can add tasked commercial SAR passes under client licence, consistent with the same model used in its Tonga crop-estimation analytics work.
Latency, archive depth, and what buyers should ask for
Sentinel-1 data is typically available on the Copernicus Data Space within one to three hours of acquisition for priority areas, and the archive extends back to 2014 for Sentinel-1A. This makes it practical to reconstruct historical vessel presence in a defined ocean area, subject to the 6-day repeat cycle per satellite and the gaps created by acquisition mode scheduling.
For near-real-time alerting, commercial SAR constellations offer shorter latency but require tasking decisions made before the pass. A buyer who wants to be alerted when a vessel of interest enters a defined ocean box needs to decide in advance which sensor to task and at what cadence. That is a different product from a retrospective analysis of archive data, and the two should not be conflated in procurement discussions. Delivered formats range from GeoTIFF detection layers and GeoJSON point files to integrated feeds into maritime picture systems.
Typical figures
| Spatial resolution (Sentinel-1 IW) | 5 × 20 m (range × azimuth) |
| Spatial resolution (Sentinel-1 EW) | 20 × 40 m |
| Spatial resolution (ICEYE strip-map / spot) | ~3 m / ~1 m |
| Spatial resolution (Capella spotlight) | ~0.5 m |
| Revisit (Sentinel-1 A+B combined) | ~3 days at mid-latitudes; shorter at high latitudes |
| Radar frequency | C-band (5.4 GHz, Sentinel-1 / RADARSAT); X-band (9.65 GHz, ICEYE / Capella) |
| Minimum detectable vessel (Sentinel-1 IW, calm sea) | ~30-40 m steel-hulled vessel; smaller targets unreliable |
| Archive depth (Sentinel-1) | 2014 to present (Sentinel-1A); 2016 to present (Sentinel-1B, with gaps) |
| Data latency (Sentinel-1 priority areas) | 1-3 hours post-acquisition via Copernicus Data Space |
| Delivery formats | GeoTIFF detection raster, GeoJSON point detections, CSV with centroid coordinates and estimated vessel length |
Analytics Satellize can run
| CFAR vessel detection layer | Cell-Averaging or Order-Statistic CFAR applied to calibrated sigma-nought imagery; VV polarisation primary channel | GeoJSON point file with detection centroids, estimated length from SAR shadow or bright-pixel extent, confidence score, and acquisition timestamp |
| Sea-state-adjusted detection confidence flag | ECMWF ERA5 or GFS wind-field overlay used to assign a clutter-severity index per detection; detections in high-clutter conditions flagged with reduced confidence | Attribute column appended to detection GeoJSON; summary table of estimated missed-detection probability by sea-state band |
| Historical vessel presence density map | Aggregation of CFAR detections across Sentinel-1 archive passes over a defined area and time window; kernel density estimation on detection centroids | GeoTIFF density raster at 500 m grid, plus PDF report summarising temporal patterns and data-gap periods |
| Tasked commercial SAR detection for named area of interest | ICEYE or Capella spotlight/strip-map collection tasked to client-specified ocean box; CFAR detection run on delivered imagery | Detection report within agreed latency window; GeoJSON plus annotated image chip for each detection above threshold |
| Vessel length estimation from SAR | Bright-pixel cluster extent in azimuth direction used as proxy for vessel length; calibrated against known vessel dimensions in public AIS records for length-class banding | Length-class attribute (small / medium / large / very large) attached to each detection; uncertainty range stated per class |
| Polarimetric discrimination report | VV/VH ratio and dual-polarisation decomposition applied to Sentinel-1 or RADARSAT compact-polarimetry data to separate vessel returns from wave-feature false alarms | Filtered detection layer with false-alarm-suppression log; before/after detection count comparison |
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.