How it works: Matching satellite vessel detections to AIS

As part of the Using Satellite Data for Analysis in Starboard series, this article explains how satellite collections are supported by software analytics in the automatic detection of vessels within an image or scan, and how this is presented in the Starboard platform.

Even the smallest footprints of the highest resolution satellites can be several square kilometres in size, which makes spotting vessels by eye very tedious, and almost impossible at scale or for small craft.

Can you see the 60m long vessel in this high resolution EO footprint? The answer is shown later in this article.

This is why computerised analysis is applied to EO and SAR satellite imagery to automatically detect potential vessels within the images and estimate their size, type, speed and heading. The vessel detection algorithms behind this process are usually developed and applied by the satellite provider themselves, or through an intermediate provider.

In Starboard, automated detections are displayed as transparent boxes over the apparent vessel.

Where a vessel (or vessel-like shape) is automatically identified in satellite imagery, those detections are provided alongside the raw satellite data into Starboard, where they will appear as transparent boxes around the vessel, to highlight the location in the wider image. The image detection boxes are then processed further in Starboard, with a colour being applied based on whether it could be correlated to any nearby AIS broadcasts.

⬜ AIS match: Match score greater than 0.5

🟨 Possible dark vessel: Match score 0.5 or less

🟥 Likely dark vessel: Match score is not defined

When there are numerous detection boxes obscuring the clarity of the underlying image, their visibility can be toggled off by selecting the relevant ‘eye’ (visibility) icon in the Added satellite data panel. The same function can be used for the underlying raw imagery as well, when there are multiple images taken at the same time which need to be separated for individual inspection.

Try toggling the 'eye' icons to see how it changes your map visualisations

The colour classification of the detection boxes are determined by Starboard’s probabilistic method.

The probabilistic method accounts for the positional uncertainty of both the satellite detection and AIS-reported positions. Each potential match is scored between 0 and 1, indicating the likelihood that the detection corresponds to a given AIS vessel. Vessels that cannot be correlated to an AIS signal are assigned as likely dark vessels.

Probabilistic matching was introduced to Starboard in May 2025. Please note that any detections matched prior to May 2025 do not account for proximity and may present a different colour classification in vessel detections.

Detection correlation can also be applied to bespoke datasets such as Vessel Monitoring System (VMS) data, or operational surveillance data. Chat to our team about your organisation’s specific data integration requirements.

Correlating RF detections with aerial surveillance images during Operation Nasse 2022.

Uncertainty calculation: The estimated AIS position at the time of detection is derived by interpolating between the vessel’s known positions before and after the satellite data capture. The uncertainty around this interpolated location is shaped by how far the vessel could plausibly travel in that time. We also consider the geolocation accuracy of the satellite sensing system to compute the final match probability.

A satellite vessel detection matched to an AIS transmitting vessel is selected. Hover your mouse over the matched vessel name (coloured blue) in the vessel detection panel to the left of your screen. This will produce a light blue ellipse on the map, representing the area of uncertainty for the vessel position.

A possible dark vessel detection is indicated by a yellow square, as the location of the detection is only just inside of the area of uncertainty (blue ellipse) of the vessel position.

Several possible matches: Our data interpolation and uncertainty logic helps to avoid false matches even in busy areas. If multiple vessels have very similar match scores, all of their names are displayed in the detection panel for manual differentiation by the analyst. This is usually more common near ports or anchorages.

Commercial EO and SAR satellite data providers may integrate the automated vessel detections as part of an image order, but this is not always guaranteed (it may need to be purchased separately). If you’re uncertain what this means for you, we can assist with assessing your needs and communicating them to the provider(s).

Automated vessel detections are typically delivered to the end user as an image chip, showing a cutout of the detected vessel and its immediate surroundings on the map. The detection algorithms behind this function can however, make mistakes.

This SAR image chip from an automatic vessel detection in a port area is typically how satellite data is delivered in most MDA platforms.

To help analysts identify false detections that may have been missed by the algorithms, Starboard also integrates the full image footprint directly into the map, as shown below.

The same SAR image chip as above + the raw satellite image integrated into the map shows additional vessels missed by the algorithm.

Algorithm failures in vessel detections (false detections) are sub-categorised into false positives or false negatives:

Understanding the rate of false detections is particularly important for operations with the goal of dark vessel detection. Knowing that there will be false detections in almost every satellite footprint (in particular for low resolution and RF sensors) highlights the importance of combining different forms of remote sensing and vessel tracking data (AIS, VMS, coastal radar, etc) according to your specific objective–which will likely change in every case.

False positives typically occur when the detection algorithms misinterpret environmental noise or fixed infrastructure as vessels. False negatives are more complex. The following table highlights some of the common causes, with a ‘tick’ representing where false negatives are more likely to occur for each sensor type and category.

The uncertainty of vessel presence caused by false detections has significant operational consequences. Analysts will rarely justify a costly follow-up response to a detection that could be a false positive. This is particularly relevant for low-resolution and RF sensors which often produce a number of false detections due to a lack of image/detail–but can also be present in high resolution imagery as well.

The image at the start of this article contained a false negative, which had to be manually identified by an analyst (yellow box + inset) via its AIS.

Automated vessel detections also have higher error-rates in nearshore waters. Land has to be masked from the detection algorithm to avoid false positives, but masking too generously, or using an inaccurate basemap can lead to chronic false negatives. Starboard is one of the only MDA integration providers to include the entire satellite image in the map (not just the image chip), which enables analysts to determine where detection failures have occurred in areas of high traffic or environmental complexity.

An uncorrelated SAR detection in a high traffic area can be compared with the surrounding environment to help determine its potential validity.

Radio Frequency (RF) scans present a particular challenge for the evaluation of detections with respect to false positive or negative rates, because there is no visual information to help make an assessment. Moreover, the geolocation uncertainty of RF detections is typically larger than that of SAR or EO.

To alleviate the uncertainty around RF detections, some providers can analyse the components of electronic vessel emissions and assign unique fingerprints to vessels, which can then be archived and used for future matching and tracking. This requires numerous scans of a consistently emitting signal, which is not guaranteed and attracts a higher financial cost.

Demonstrating the uncertainty of RF detections in high traffic areas, which are visualised as mathematical ellipses of probability.

Continue to the next article in this series to explore more advanced techniques for analysing RF and other satellite (EO and SAR) metadata in Starboard.