Monitoring maritime activities is important for safety and security, particularly for detecting illegal activities, but remains challenging due to the size and complexity of the maritime domain. The Automatic Identification System (AIS) is commonly used for this purpose, despite its known vulnerabilities. In contrast, space-based radio frequency (RF) observations provide an independent source of vessel activity but lack ship identification and therefore cannot identify vessels on their own. Following the CRISP-DM process, this study applies a Pair Hidden Markov Model (PHMM) to align individual RF detections with AIS trajectories, using temporal and spatial information while modeling noise and timing discrepancies. We evaluate it on NOAA AIS data with synthetic RF detections against nearest-neighbor baselines and analyze its errors with bias tests on trajectory length, sampling density, stationarity and AIS--RF distance. The length-corrected Forward score reaches a top-1 precision of 0.7776 on the validation split of multi-candidate detections, where errors are possible, and an AUC of 0.637 for detecting dark vessels among the detections that pass the prefilter. With a single RF detection, the score mostly reflects the AIS trajectory itself, and its errors favor longer and more stationary trajectories. A log-odds score, which compares the full model with a null model without match transitions, raises precision to 0.8580 and AUC to 0.852. Its length-corrected variant no longer shows a significant bias towards stationary trajectories and short inter-message distances, but still favors candidates farther from the RF detection. A time-weighted nearest-neighbor baseline without fitting remains ahead (precision 0.9013, AUC 0.904), so this single-detection, hand-designed PHMM formulation approaches but does not beat it.
Figures & tables
Fig. 1 : Graphical representation of the PHMM.
Matrix element
Explanation
Value/distribution
π
Initial state distribution: transition from B to an emitting state ( IAIS , IRF , M ).
Uniform, 1/3 .
τ
Transitions from an emitting state ( IAIS , IRF , M ) to termination state E .
Uniform, 0.05 .
ε
Transitions into IAIS and IRF .
fε(Δt,Δd) , see Figure 2 .
δ
Transitions into M .
fδ(Δt,Δd) , see Figure 3 .
TABLE I : PHMM transition parameters. π and τ are uniform values; ε and δ are hand-designed functions of Δt and Δd .
Fig. 2 : Weight surface of fε(Δt,Δd) . This function attains its maximum value for observations that occur nearly simultaneously and at the same location (point a ), suggesting a high probability that they originate from the same vessel. The probability decays sharply with increasing temporal (point b ) or spatial (point c ) separation, reaching its minimum when both differences are substantial (point d )
Fig. 3 : Weight surface of fδ(Δt,Δd) . This function is zero beyond predefined temporal ( 120 seconds) and spatial ( 2 kilometers) thresholds, so pairs outside these thresholds cannot transition into the M state
RF detections
n
Precision
All
17,081
0.8899 a
Single-candidate
10,903
1.0000
Multi-candidate
6,178
0.6955 a,b
Tuning split (20%)
1,236
0.7921 ( α=0.90 )
Validation split (80%)
4,942
0.6938 → 0.7776 c
TABLE II : Evaluation populations. Top: RF detections used for top-1 precision, the bias experiments and the baseline comparison, with the precision of PHMM Forward at each stage. Bottom: vessels (MMSIs), trajectories and RF detections used for the dark-vessel test.
Experiment
Forward (corrected)
Log-odds ( nα )
1. Track length
longer, 2.2×10−14
shorter, 1.0×10−20
2a. Intra-track time gap
n.s.
larger, 0.040
2b. Intra-track distance gap
closer, 7.2×10−8
n.s.
3. Stationarity
higher, 2.0×10−5
n.s.
4a. AIS–RF time gap
farther, 1.8×10−3
n.s.
4b. AIS–RF distance gap
farther, 5.9×10−30
farther, 3.8×10−14
TABLE III : Bias experiments, validation split. Cells give the direction (chosen vs. should-be) and the two-sided p -value where the test is significant at p<0.05 (Welch for 1–3, paired for 4a–4b), and n.s. otherwise. The log-odds column is the nα variant. NN time-weighted (not shown) has no significant bias except in the AIS–RF time gap (4a), which is expected, since its score contains the time difference directly.
Fig. 4 : AIS messages per trajectory for the Correct, Chosen and Should-be groups. The numbers next to the boxes are the means (orange) and medians (blue), and the legend in (c) applies to all panels.
Model
Precision
AUC (dark-vessel)
PHMM Forward ( α=1 )
0.6938
0.604
PHMM Forward (corrected)
0.7776
0.637
PHMM log-odds (raw)
0.8580
0.852
PHMM log-odds ( nα )
0.8571
0.878
NN point-to-segment a
0.6333
0.651
NN (time-weighted)
0.9013
0.904
TABLE IV : Precision (validation split, n=4,942 ) and AUC of the dark-vessel test (detections not rejected by the prefilter).
Fig. 5 : ROC of the dark-vessel test on the detections not rejected by the prefilter, for PHMM Forward ( α=1 and corrected), PHMM log-odds (raw and nα ), NN point-to-segment and NN time-weighted. Legend AUC values match Table IV .
School of Computer Science and Statistics, IRIS Research Group, ADAPT Research Centre, Trinity College Dublin, Ireland · School of Engineering, ADAPT Research Centre, Trinity College Dublin, Ireland