AIS and RF Data Fusion: Can PHMM Alignment Beat a Simple Baseline?

Study uses paired hidden Markov models to align space-based RF detections with AIS tracks, improving accuracy but still below a simple baseline.
In maritime surveillance, AIS has vulnerabilities and space-based RF lacks identity, making fusion key. A CRISP-DM study uses a paired hidden Markov model (PHMM) with spatiotemporal info to align single RF detections to AIS tracks, tested on synthetic RF from NOAA AIS data. Length-corrected Forward Top-1 accuracy is 0.7776 with dark ship AUC 0.637; log-odds raises accuracy to 0.8580 and AUC to 0.852. Yet a fit-free time-weighted nearest-neighbor baseline does better (accuracy 0.9013, AUC 0.904), showing single-detection PHMM nears but doesn't beat the simple baseline; future work needs more complex models or multi-detection info.
Maritime surveillance is critical for safety and security, especially for detecting illegal activities, but the scale and complexity of maritime domains make this task challenging. The Automatic Identification System (AIS) is a commonly used tool, but its known vulnerabilities cannot be ignored. In contrast, space-based radio frequency (RF) observations provide an independent source of vessel activity data, yet they lack vessel identity information and cannot identify ships on their own. A recent study followed the CRISP-DM process to attempt aligning individual RF detections with AIS tracks using a paired hidden Markov model (PHMM), and evaluated its performance.

Method: How PHMM Aligns RF and AIS
The study employs a paired hidden Markov model (PHMM) that leverages temporal and spatial information while modeling both noise and temporal discrepancies. PHMMs are commonly used for sequence alignment problems, and here they are used to match RF detection points to AIS tracks. The authors synthesized RF detections from NOAA's AIS data and compared the approach against a nearest-neighbor baseline. Error analysis used bias tests to examine the effects of track length, sampling density, stationarity, and AIS-RF distance.
Paired Hidden Markov Model (PHMM) is a variant of the hidden Markov model (HMM) designed specifically for sequence alignment. A standard HMM models the hidden state sequence underlying a single observation sequence, whereas a PHMM simultaneously models two sequences and introduces "paired states" representing correspondences between elements of the two sequences. In bioinformatics, PHMMs are commonly used for protein or DNA sequence alignment; in this study, one sequence is the spatiotemporal coordinates of RF detections and the other is the AIS track points. The PHMM uses state transition probabilities to simultaneously capture the physical constraints of a vessel moving along a track and the noise characteristics of RF detections, thereby finding the best matching path while allowing for temporal offsets and spatial errors.
Performance Evaluation: Accuracy and AUC
On a validation set with multiple candidate detections (where mismatches are possible), the length-corrected Forward score achieved a Top-1 accuracy of 0.7776, and the AUC for detecting "dark ships" (vessels not transmitting AIS) among pre-filtered detections was 0.637. When only a single RF detection is available, the score mainly reflects the AIS track itself, and errors tend to favor longer, more stationary tracks.
AUC (Area Under the ROC Curve) is a common metric for evaluating the ranking ability of a binary classifier, ranging from 0 to 1, where 0.5 corresponds to random guessing and 1.0 indicates perfect discrimination. In this task, AUC measures the probability that the model assigns a higher score to a "true match" RF-AIS pair than to a "false match" pair. Since dark ship detection in maritime surveillance is essentially an imbalanced classification problem—the vast majority of detections correspond to normal vessels—AUC better reflects the model's overall discriminative power across different thresholds than raw accuracy. An AUC of 0.637 indicates the model is only slightly better than random, while 0.852 shows that the log-odds score significantly improves ranking quality.
Log-Odds Improvement and Baseline Comparison
After introducing the log-odds score (comparing the full model against a null model without matching transitions), accuracy improved to 0.8580 and AUC to 0.852. Its length-corrected variant no longer significantly favors stationary tracks and short message intervals, but it still prefers candidates farther from RF detections. However, a fit-free time-weighted nearest-neighbor baseline performed better (accuracy 0.9013, AUC 0.904). This indicates that the single-detection, hand-crafted PHMM formulation comes close to but fails to surpass this baseline.
Log-odds score originates from statistical hypothesis testing and is defined as the logarithm of the ratio of the likelihood of the observed data under the full model to that under the null model. The null model assumes no true matching transition exists between RF detections and AIS tracks, with data generated solely from a background distribution. The log-odds score therefore quantifies the strength of evidence for "a match exists" versus "no match": the higher the score, the more credible the match. Compared to using the Forward score directly, the log-odds score eliminates systematic biases from track length and detection density, making comparisons between candidates fairer—this is the main reason both accuracy and AUC improved simultaneously.
Implications and Limitations
The PHMM approach shows potential for data fusion, especially with the log-odds score improvement, but a simple baseline still has the edge. Future work may require more sophisticated models or more RF detection information to improve performance. The study also highlights the challenges and opportunities of fusing multi-source data in maritime surveillance.
One-line Summary
A study uses paired hidden Markov models to align space-based RF detections with AIS tracks, improving accuracy but still falling short of a simple baseline.
Paragraph Summary
In maritime surveillance, AIS has vulnerabilities while space-based RF observations lack identity information, making their fusion critical. A study following the CRISP-DM process employs a paired hidden Markov model (PHMM) to align individual RF detections with AIS tracks using spatiotemporal information, validated on synthetic RF detections from NOAA AIS data. The length-corrected Forward score achieved Top-1 accuracy of 0.7776 with an AUC of 0.637 for dark ship detection; introducing the log-odds score raised accuracy to 0.8580 and AUC to 0.852. However, a fit-free time-weighted nearest-neighbor baseline performed better (accuracy 0.9013, AUC 0.904), indicating that the single-detection PHMM comes close to but does not surpass the simple baseline, and future work needs more sophisticated models or multi-detection information.
Related articles

Abnormal Error Rate Spike in Claude Mythos 5.2 and Fable 5.2
Claude Mythos 5.2 and Claude Fable 5.2 show an abnormal spike in request error rates, marked as investigating on the official status page. This article covers the incident overview, practical impact on developers, and recommended responses.

ReAct Runtime Context Trimming: Why Claude Code Acts at 70%
A deep dive into ReAct runtime context trimming: why Claude Code compresses at ~70% usage, why tool calls dominate, why messages must be dropped in pairs, and why codebases don't need RAG.

AI Agent vs LLM: How Cursor and Claude Code Actually Work
An in-depth analysis of the essential difference between AI Agents and LLMs: LLMs generate code suggestions, while Agents modify files and run tests via tool calls. Using a Spring Boot bug fix as an example, it breaks down the workflows, permission control, and verification mechanisms of Cursor and Claude Code.