Monitoring maritime activities remains important for safety and security, especially for detecting illegal operations, but the task grows more complex as the domain expands. The Automatic Identification System, or AIS, provides vessel traffic data and has become a standard tool despite known vulnerabilities such as signal spoofing or deliberate shutdown. Space-based radio frequency observations offer an independent view of vessel activity, yet they lack precise ship identification, meaning they cannot on their own name a vessel or confirm its origin.

This work follows the CRISP-DM process framework to address the fusion problem. The core contribution is a Pair Hidden Markov Model that aligns individual RF detections with AIS trajectories. The model uses temporal and spatial information while explicitly modeling noise and timing discrepancies between the two data sources. The authors evaluate the approach on NOAA AIS data augmented with synthetic RF detections, comparing against nearest-neighbor baselines and conducting bias tests across trajectory length, sampling density, stationarity, and the distance between AIS and RF observations.

On the validation split of multi-candidate detections, the length-corrected Forward score achieves a top-1 precision of 0.7776 and an AUC of 0.637 for identifying dark vessels among those that pass the prefilter. When a single RF detection is considered, the score primarily reflects the underlying AIS trajectory, and errors tend to favor longer, more stationary paths. Introducing a log-odds score, which contrasts the full model against a null model without match transitions, lifts precision to 0.8580 and AUC to 0.852. The length-corrected version of this score eliminates the significant bias toward stationary trajectories and short inter-message gaps, though it still favors candidates positioned farther from the RF detection. A time-weighted nearest-neighbor baseline that requires no fitting remains ahead, reaching precision 0.9013 and AUC 0.904, so the single-detection, hand-designed PHMM formulation comes close but does not surpass this reference.

Maritime monitoring challenges and the AIS/RF gap

Maritime domain awareness serves many purposes: collision avoidance, fisheries management, border security, and detection of illicit operations such as smuggling or unauthorized anchoring. AIS transponders broadcast vessel position, speed, and identity at regular intervals. The system is not without flaws; receivers can be jammed, messages can be fabricated, and vessel operators may disable transponders to go dark. These vulnerabilities motivate the search for complementary data sources.

Space-based RF sensors detect radio emissions from vessels, including those that have turned off AIS. The advantage is independence—AIS and RF provide separate views of the same physical scene. The limitation is that RF observations alone typically report only signal strength, location, and timestamp, without an embedded vessel identifier. Matching an RF signal to a known ship therefore requires aligning the RF footprint with an AIS track, a problem of temporal and spatial correspondence.

The paper frames this as a data-fusion task: given a sequence of AIS positions and a set of RF detections, determine which RF observation corresponds to which AIS segment, and do so despite gaps, delays, or noisy measurements. The authors choose a probabilistic model—specifically a Pair Hidden Markov Model—to represent the alignment as a sequential decision process.

Pair Hidden Markov Model architecture

A Hidden Markov Model represents observed data as generated by an underlying hidden state process. In a standard HMM, the hidden states evolve according to a transition matrix, and each state emits an observation according to a emission distribution. The parameters—transition probabilities, emission densities, and initial state probabilities—are learned or specified by the modeler.

The Pair Hidden Markov Model extends this structure to handle two correlated observation sequences. Here, one sequence consists of AIS trajectory points, the other of RF detection events. The hidden states represent whether a given RF detection matches a particular point on an AIS trajectory, or whether no match exists. Transition probabilities govern how the alignment moves forward through time: an RF detection may match the current AIS point, skip forward, or be declared an false alarm. Emission probabilities capture how well the spatial and temporal features of the RF observation align with the AIS point, typically incorporating distance, time difference, and expected signal characteristics.

In the authors' formulation, the pair structure means the model jointly considers both sequences rather than treating them independently. This joint consideration allows the HMM to reason about timing discrepancies: an RF detection may occur slightly before or after the corresponding AIS report, and the model's transition structure can accommodate such offsets. Noise in either modality—AIS position error, RF geolocation uncertainty—is incorporated through the emission distributions, typically modeled as Gaussian-like densities centered on the expected relationship between the two data types.

The model is not fully learned from data in an end-to-end sense. Instead, the authors treat many components as hand-designed, relying on domain knowledge to set transition structures and baseline emission scales. This design choice limits the degrees of freedom the model must estimate from the available data, which is sensible given the relatively small synthetic RF dataset over which experiments are run.

Experimental setup and evaluation metrics

The authors use NOAA AIS data as the base traffic layer. AIS reports are collected over a defined area and time window, forming the trajectory dataset. Synthetic RF detections are then generated to mimic space-based observations: each RF event includes a location, timestamp, and signal attribute, but no vessel MMSI ID. This synthetic setup permits ground-truth matching, since the authors can associate each RF event with the correct AIS segment by construction.

Baseline comparisons use nearest-neighbor approaches. A simple time-weighted nearest-neighbor assigns each RF detection to the closest AIS trajectory point in combined spatial-temporal distance, without fitting any probabilistic model. This baseline serves as a point of reference because it requires no training or parameter estimation.

Performance is measured through several lenses. The Forward score evaluates how well the model forwards probability mass through the HMM states, producing a ranking of candidate matches. Top-1 precision at the length-corrected threshold indicates what fraction of the highest-ranked matches are correct. The AUC (area under the receiver operating characteristic curve) assesses the model's ability to distinguish dark vessels—those without AIS identity—from normal traffic among the detections that survive a prefiltering stage. Bias tests examine whether error rates systematically depend on trajectory length, how frequently observations are sampled, whether the vessel motion appears stationary, and the spatial gap between AIS and RF positions.

Scoring variants and their behavior

The base Forward score, length-corrected to account for longer trajectories receiving disproportionate probability mass, reaches top-1 precision of 0.7776 on the validation split. This number reflects the model's ability to correctly identify the matching AIS segment when multiple RF candidates compete for the same AIS point. The associated AUC of 0.637 for dark-vessel detection is modest, indicating that the score alone is not highly discriminating for the hardest cases.

When the analysis considers a single RF detection, the Forward score largely mirrors the characteristics of the underlying AIS trajectory. Longer trajectories and more stationary paths tend to receive higher scores, simply because the HMM has more time steps to accumulate matching probability. This pattern is visible in the bias tests: errors favor candidates with longer observed paths and with motion that appears consistent over time.

The log-odds score takes a different tack. It compares the full PHMM, which includes transitions that allow an RF detection to "match" an AIS point, against a null model that contains no such match transitions. In the null model, every RF detection is forced to either be a false alarm or to align through a default, less-specific path. The log-odds ratio quantifies how much more likely the data are under the matching model than under the no-match default. This variant raises top-1 precision to 0.8580 and AUC to 0.852, a substantial improvement over the base Forward score.

The length-corrected log-odds variant removes the bias observed toward stationary trajectories and short inter-message distances. After correction, the score no longer systematically penalize or reward paths based solely on their length or how closely spaced the RF messages are. However, the score still favors candidates that are farther from the RF detection in spatial terms. This remaining bias suggests that distance, while no longer the dominant factor it was in the uncorrected version, still exerts a moderate influence on the final ranking.

Bias analysis across trajectory and signal properties

The authors run systematic bias tests to understand where the model succeeds and where it systematically misroutes matches. Four dimensions are examined:

  • Trajectory length: longer AIS paths tend to receive higher Forward scores, even after length correction, because the HMM has more opportunities to observe matching evidence.
  • Sampling density: when AIS reports are sparse, the model has fewer emission events to base its alignment on, which can degrade matching accuracy.
  • Stationarity: trajectories where the vessel moves little over time attract higher scores, as the emission model expects consistent spatial proximity between the RF and AIS positions.
  • AIS–RF distance: even after length correction, candidates located farther from the RF detection's geolocation tend to be ranked lower, indicating a persistent spatial effect.

The log-odds score, particularly its length-corrected form, mitigates the stationarity and inter-message-distance biases. The persistent spatial bias, however, remains an area where the model could be further refined, perhaps by incorporating an explicit distance-dependent term in the emission model rather than letting it emerge as a side effect.

Comparison with the nearest-neighbor baseline

A time-weighted nearest-neighbor baseline, which does not fit any probabilistic model, achieves precision 0.9013 and AUC 0.904 on the same evaluation splits. This baseline works by computing, for each RF detection, the closest AIS trajectory point in a combined metric that weights temporal difference and spatial distance. The absence of fitting means the baseline is fast to compute and requires no training data, but it also lacks the probabilistic framing that the PHMM provides.

The PHMM approaches but does not exceed this baseline. The log-odds variant gets closest, raising precision and AUC into the high‑0.8 range, yet the gap to 0.9013/0.904 remains. The authors note that the baseline's hand-designed weighting of time and distance is effective, and that the PHMM's advantage lies in its modular structure—components could be replaced or re‑estimated—but in its current hand‑crafted form it cannot surpass a well-tuned non‑parametric method.

Practical implications for developers

For a developer working on maritime data fusion, the paper offers several takeaways. First, the PHMM framework provides a structured way to combine two heterogeneous streams—AIS and RF—without having to engineer a single monolithic classifier. The pair-state design naturally handles the "match vs. no‑match" decision, and the transition-emission decomposition means each component can be inspected or replaced independently.

Second, the scoring variants demonstrate that modest changes to the probabilistic formulation can have large effects on ranking quality. The log-odds construction, which contrasts a matching model against a sensible null, is particularly effective and worth implementing as a drop-in improvement over a raw Forward score.

Third, the bias analysis is valuable for model debugging. The observation that uncorrected Forward scores favor long, stationary trajectories suggests that a length-correction step should be included in any production pipeline, especially when trajectory lengths vary widely across the fleet. The residual spatial bias reminds us that distance between observation sources cannot be fully eliminated by probabilistic modeling alone; a downstream distance‑thresholding step may still be needed.

Finally, the comparison to the nearest-neighbor baseline sets expectations. Developers should not expect the PHMM to automatically outperform simple, well‑designed non‑parametric methods. Instead, the PHMM's value may lie in its interpretability, its ability to produce calibrated probabilities, and its flexibility when the model components are replaced with data‑learned alternatives.

In practice, a developer might start with the time-weighted nearest-neighbor baseline for quick prototyping, then replace the scoring function with the log-odds PHMM variant for better dark‑vessel detection. Further improvements could come from learning the transition and emission parameters from real RF/AIS pairs, rather than keeping them hand‑crafted, and from incorporating additional features such as vessel type, cargo class, or historical behavior patterns.

The code and data associated with the paper are made available through the arXiv supplementary material, allowing reproducibility and further experimentation. Those interested in extending the approach might explore deep-emission models, where a neural network learns the emission density from raw RF spectra and AIS sensor coordinates, or variational extensions that approximate the posterior over matchings more flexibly.

Read the paper on arXiv