Skylight has developed AI models capable of estimating key vessel attributes — such as length, heading, speed, and type — using vessel detections from a range of public satellite sources. This is especially useful for analyzing “dark” vessels that do not transmit AIS data and can help users infer additional details about their identity and behavior. 

Methodology

While it is generally not possible to determine the exact identity of a vessel from satellite imagery alone. Computer Vision (CV) models like those developed by Skylight can, however, provide additional insights into what the vessel is (vessel length, vessel type) and what the vessel is doing (heading, speed), which can help narrow down vessels of interest in an operational context.

The table below summarizes the key vessel attributes that Skylight can detect across different satellite data sources.

To estimate these key vessel attributes, Skylight uses computer vision models powered by deep learning (DL) capabilities, which is like a high-dimensional pattern recognition engine. When a DL model estimates vessel attributes, it doesn't follow a manual of rules like a human analyst would (e.g. the vessel length gives an indication about the vessel type). Instead, it uses feature hierarchies directly drawn from the raw pixels of satellite imagery. To analyze a vessel, the model processes the image through dozens of pixel layers, each capturing a different level of meaning. This enables it to track thousands of micro-features simultaneously to refine its estimates, where a human could only reasonably consider three or four variables to make the estimate.

The drawback of this approach is the “black box” problem: while we understand the inputs (satellite pixels) and the outputs (e.g., vessel attributes), the millions of intermediate computations are not easily interpretable. More specifically, DL models do not provide human-readable explanations for how much specific features of an image influenced the final prediction. What Skylight does to make the data useful for operational purposes is provide performance metrics (accuracy, error ranges) for each attribute so Skylight users can make informed decisions. For the purposes of enforcement and legal actions, these AI-derived attributes serve as a starting point. An independent human analyst is required to validate the imagery since they can explain their exact logic and the model cannot. 

Type Estimation

Skylight models can provide vessel type estimations on electro-optical (Sentinel-2, Landsat-8/9) and SAR (Sentinel-1) imagery. Our models can recognise the following vessel types: cargo, tanker, fishing, service, passenger, enforcement, search & rescue, and pleasure. 

While a human looks at vessel characteristics across two or three variables (presence of containers, deck features, vessel length) to estimate the vessel type, Deep Learning (DL) relies on an analysis of the spatial distribution of pixel intensities to predict the vessel type – for example a container ship, with stacked containers, tends to exhibit high-frequency textures, whereas a tanker’s flat deck produces lower-frequency, more uniform pixel patterns. 

You can find a detailed breakdown of the vessel speed estimation accuracy for each sensor below:

• Sentinel-2
• Landsat-8/9

Length Estimation

Skylight models can provide vessel length estimations on electro-optical (Sentinel-2, Landsat-8/9) and SAR (Sentinel-1) imagery. 

For AI, estimating vessel length is treated as a segmentation task. The model must determine which pixels belong to the ship and which correspond to the surrounding environment (for example, water, wake, or shadow). It analyzes the image using hierarchical features, such as sharp changes in pixel brightness to delineate the vessel, and then derives its length based on these detected patterns.

You can find a detailed breakdown of the vessel length estimation accuracy for each sensor below:

• Sentinel-2
• Landsat-8/9