PaperlayerView exact revision on arXiv

Source-linked AI summary

WAY: Estimation of Vessel Destination in Worldwide AIS Trajectory

Jin Sob Kim, Hyun Joon Park, Wooseok Shin, Dongil Park, Sung Won Han

arXiv:2512.13190v1cs.LGcs.AI

TL;DR

AIS destination estimation must contend with unreliable, irregular data and prior work limited to regions of interest. WAY models global port-to-port trajectories as detailed nested sequences; experiments report improvements over benchmark models, with further gains from Gradient Dropout.

  • Problem

    Global vessel destination estimation is limited by region-of-interest constraints, irregular AIS intervals, and records that are not annotated as port-to-port operations.

    Source: §II · p. 3, §III · p. 4

  • Method

    WAY recasts port-to-port AIS trajectories as detailed grid-wise nested sequences and applies a four-channel model with Gradient Dropout for destination estimation.

    Source: §VII · pp. 13–14

  • Results

    Across benchmark models, Gradient Dropout yields average gains of +1.01% for overall accuracy and +1.92% for F1-score, while WAY improves estimation across trajectory steps.

    Source: C. Experimental Results · pp. 11–12

  • Takeaways & Limitations

    WAY’s expandability to multitask ETA estimation is explored as a potential route toward real-world port-congestion estimation through further improvement in arrival-time identification.

    Source: §VII · p. 14

    The approach is formulated on the premise that an annotated port-to-port trajectory is given.

    Source: §VII · p. 13

Abstract

from arXiv · show

The Automatic Identification System (AIS) enables data-driven maritime surveillance but suffers from reliability issues and irregular intervals. We address vessel destination estimation using global-scope AIS data by proposing a differentiated approach that recasts long port-to-port trajectories as a nested sequence structure. Using spatial grids, this method mitigates spatio-temporal bias while preserving detailed resolution. We introduce a novel deep learning architecture, WAY, designed to process these reformulated trajectories for long-term destination estimation days to weeks in advance. WAY comprises a trajectory representation layer and Channel-Aggregative Sequential Processing (CASP) blocks. The representation layer generates multi-channel vector sequences from kinematic and non-kinematic features. CASP blocks utilize multi-headed channel- and self-attention for aggregation and sequential information delivery. Additionally, we propose a task-specialized Gradient Dropout (GD) technique to enable many-to-many training on single labels, preventing biased feedback surges by stochastically blocking gradient flow based on sample length. Experiments on 5-year AIS data demonstrate WAY's superiority over conventional spatial grid-based approaches regardless of trajectory progression. Results further confirm that adopting GD leads to performance gains. Finally, we explore WAY's potential for real-world application through multitask learning for ETA estimation.

I. INTRODUCTION

AIS destination estimation is challenged by unreliable, irregular observations and prior trajectory-prediction studies constrained to regions of interest. The paper motivates global port-to-port modeling and introduces WAY to process detailed, multivariate trajectories.

I. INTRODUCTION

  • Unstructured destination and ETA reports, errors, and irregular observation intervals make AIS data unreliable for marine-traffic analysis despite its rich trajectory context.The introduction links accurate destination and ETA knowledge to addressing port congestion in advance.Source: §I · p. 1
  • Prior vessel-prediction studies were constrained to regions of interest, while global AIS data had mainly supported traffic and route analyses rather than tracking individual vessels.The authors identify global-scope port-to-port trajectory modeling as the gap to address.Source: §I · p. 1, §II · p. 3
  • The paper proposes annotating meaningful port-to-port segments and recasting AIS trajectories as nested sequences for global destination estimation.The formulation assumes known departure and destination ports for each verified trajectory.Source: §I · pp. 1–2, §II · p. 3
  • WAY represents trajectories as four-channel vector sequences, aggregates channels with attention, and uses a Transformer decoder to deliver past information to the current step.Gradient Dropout is proposed to temper feedback bias when many-to-many training uses a single destination label across sequence steps.Source: §I · p. 2

II. RELATED WORK

  • Real-world AIS records can contain transmission errors, missing values, and contextual misinformation caused by sensor or transmission failures and human mistakes.Prior work addressed these data-quality challenges through preprocessing or explicit error identification and refinement.Source: §II · p. 2
  • Earlier spatial-grid tokenization can mitigate irregular-interval bias, but larger grids lose trajectory detail and smaller grids increase the number of parameters exponentially.The text also notes uncertainty about whether tokenized trajectories adequately represent irregular intervals.Source: §II · p. 3

Contribution

The section frames worldwide AIS destination estimation as a long-range problem ill-served by message-wise or grid-only sequences, and introduces a nested grid representation and WAY with Gradient Dropout.

Contribution

  • The proposed spatial-grid approach recasts trajectories as nested sequences to mitigate spatio-temporal bias while preserving detailed representation.The paper identifies irregular intervals and coarse portrayal as the respective problems with prior sequence representations.Source: Contribution · pp. 3–4
  • WAY uses multi-channel representations and multi-headed self- and channel-attention, while Gradient Dropout stochastically suppresses length-dependent feedback surges during many-to-many training.The model uses kinematic and non-kinematic context, including ship type and departure port.Source: Contribution · p. 4

III. PROBLEM DEFINITION

  • The task maximizes destination-port probability from the departure port and AIS messages observed through time t, while irregular intervals and unverified, unsegmented records complicate direct application to real-world data.Source: §III · p. 4

A. Spatial Grid Unit Processing

  • Grid-only sequences lose local navigational detail; the nested representation retains each grid unit alongside its original continuous-time subsequence.This preserves the original sequence’s conditional destination-estimation objective while using grid units as processing steps.Source: A. Spatial Grid Unit Processing · pp. 4–5

B. Annotation Framework

The annotation framework constructs verified port-to-port trajectories, and WAY processes them as nested sequences organized by spatial grid. Its representation combines sampled local observations, grid-relative coordinates, time distances, departure port, and ship type.

B. Annotation Framework

  • Port-to-port trajectories are segmented using AIS features, Damerau–Levenshtein distance, and port-wise thresholds, then refined with DBSCAN to detect abnormal shifts.The annotation framework verifies trajectories sharing a departure and destination and removes erroneous movement shifts.Source: B. Annotation Framework · p. 5

IV. METHODOLOGY

  • WAY combines a trajectory representation layer with Gradient Dropout to mitigate feedback bias proportional to data length during many-to-many training.Gradient Dropout is applied only during training and detached for inference.Source: §IV · p. 5

A. Trajectory Representation

  • Trajectories are reorganized into uniform spatial grids, with observations sampled within each region and coordinates expressed relative to grid centers.The number of sampled observations follows a Poisson distribution with parameter λ = 5; time distances measure progression in days from the trajectory’s first observation.Source: A. Trajectory Representation · pp. 5–6
  • The model input also includes the departure port and vessel type alongside the sequence of grid regions.These non-sequential features are identified from the annotation and the ship, respectively.Source: A. Trajectory Representation · p. 6

Description of AIS trajectory components

The trajectory components distinguish global grid locations, local patterns, departure and ship semantics, and day-unit time distances. Spatial Encoding maps coordinates to vectors intended to preserve their positional relationships.

Description of AIS trajectory components

  • Trajectory components comprise unit-area coordinates, local patterns within each area, departure port, ship type, and day-unit time distances.The time distance for each region is measured between its last observation and the trajectory’s first observation.Source: Description of AIS trajectory components · p. 6

1. Spatial Encoding

  • Spatial Encoding transforms coordinates into high-dimensional sinusoidal vectors whose longitude-wise cyclic and latitude-wise diverging distances preserve spherical positional identity.The encoding is inspired by Transformer positional encoding.Source: §1 · p. 7
  • PCA visualizes encoded vectors from the 1 × 1° grids, with the sampled sets distinguished in blue and red.The figure describes the three-dimensional PCA result of the encoded vectors.Source: §1 · p. 7

2. Step-wise Gated Recurrent Unit

The model summarizes local navigational patterns by processing each within-grid subsequence with a GRU. Departure, ship type, and continuous time encodings contribute semantic and temporal information to its four-channel sequence.

2. Step-wise Gated Recurrent Unit

  • A stacked GRU processes each local subsequence and uses its last hidden state to represent that region’s navigational pattern.The GRU is used to capture ship behavioral details and local patterns within uniform grid areas.Source: §2 · p. 7

3. Semantic Representations and Time Encoding

  • Departure-port and ship-type embeddings receive continuous sinusoidal day-offset encodings, forming channels alongside global-space and local-pattern representations.The resulting input is a four-channel vector sequence.Source: §3 · pp. 7–8

B. Channel Aggregative Sequence Processing

CASP aggregates distinct AIS channels and passes sequential information toward the current decision step. Its Multi-head Channel Attention module weights channel representations and reduces them to a single vector at each step.

B. Channel Aggregative Sequence Processing

  • CASP uses multi-head channel attention, masked multi-head self-attention, and a shared feed-forward network for feature aggregation, sequential processing, and decoding.The modules respectively handle current-feature aggregation, sequential processing, and decoding.Source: B. Channel Aggregative Sequence… · p. 8

1. Multi-head Channel Attention

  • Multi-head Channel Attention compares channels, weights them using pooled summaries, then concatenates heads and max-pools channels into one vector per step.The module uses average and max pooling to form channel summaries and applies the resulting attention weights to the input.Source: §1 · p. 8

2. Masked Multi-head Self Attention

WAY’s CASP blocks use masked self-attention to deliver past trajectory information to each current step, then update channel representations while preserving their identities. Gradient Dropout stochastically controls stepwise loss validity according to instance length.

2. Masked Multi-head Self Attention

  • Masked multi-head self-attention delivers past aggregated trajectory information to the current step while blocking information flow from future steps.It computes attention over sequence steps, scales and mixes past-value representations, and combines the heads into an output sequence.Source: §2 · p. 9

3. Channel Concatenation and Shared Feed Forward

  • The channel concatenation replaces one input channel with the self-attention output, repeatedly enriching that channel while other channels retain their current-step identities.The shared feed-forward module applies two linear transformations with an intermediate ReLU, shared across positions and channels.Source: §3 · p. 9

4. Residual Connection

  • Residual connections preserve the identity of the withdrawn channel while incorporating aggregated channel and sequence representations; layer normalization stabilizes sequential hidden-state dynamics.Residual connections are also used to address vanishing gradients in deep networks.Source: §4 · p. 9

C. Loss Function and Learning Technique

  • Gradient Dropout stochastically controls stepwise loss validity using a sampling ratio inversely proportional to the log-scaled instance length.The technique targets biased parameter updates when very long trajectories contribute excessive feedback toward a single destination.Source: C. Loss Function and Learning Technique · p. 9

V. EXPERIMENTS

The experiments use five years of global AIS trajectories, compare WAY with recurrent and attention-based benchmarks under shared settings, and assess destination accuracy across trajectory progression and macro-F1.

A. Dataset

  • The dataset contains approximately 130K trajectories and 17M AIS messages from 5,103 ships, mapped to 3,243 port classes and split 70%/15%/15% by destination.The AIS data span January 2016 to November 2020 and cover tankers, container ships, and bulk carriers.Source: A. Dataset · p. 10

1. Benchmark Models and Implementation Details

  • Benchmarks include LSTM, GRU, attention-equipped recurrent models, Transformer-decoder, and TrAISformer, with shared layer and hidden-size settings.The spatial grid unit is 1×1 (◦2); models use four stacked layers and hidden dimension 128, while the attention models share additional settings.Source: §1 · p. 10

2. Evaluation Metric

  • Evaluation measures overall accuracy at every grid step, accuracy across four trajectory-progression quartiles, and macro-averaged F1 across destination classes.The F1 measure addresses imbalanced destination labels by averaging class-wise F1 scores.Source: §2 · pp. 10–11

C. Experimental Results

WAY outperforms the benchmark models across trajectory progression, and Gradient Dropout adds gains across different backbones. The comparisons also point to the value of richer trajectory representations and effective aggregation.

C. Experimental Results

  • WAY improves destination-estimation performance over benchmark models at overall and progression-specific steps, with particular improvement before the first quartile.The comparison includes overall scores and trajectory progression; the supplied table caption identifies the benchmark results by quartile and overall steps.Source: C. Experimental Results · p. 11
  • The findings suggest that detailed trajectory representations and effective information aggregation matter for destination estimation.Source: C. Experimental Results · p. 12
  • Transformer-decoder has the highest performance among conventional spatial-token methods, while attention added to recurrent models is uncompetitive in this comparison.The authors suggest that self-attention handles long-term dependencies better, whereas recurrence may impede long-range inter-token correlations.Source: C. Experimental Results · p. 12
  • TrAISformer improves on Transformer-decoder by tokenizing longitude, latitude, SOG, and COG, supporting the value of AIS features beyond spatial coordinates.The two approaches have an identical Transformer-decoder structure, but use different token representations.Source: C. Experimental Results · p. 12
  • Accuracy and F1-score differ across benchmark models because port-to-port operation frequencies are imbalanced by ship type, yet WAY performs clearly on both measures.Source: C. Experimental Results · p. 12
  • Gradient Dropout yields gains across benchmark backbones, averaging +1.01% in overall accuracy and +1.92% in F1-score.The reported comparisons also verify a gain when Gradient Dropout is applied to WAY.Source: C. Experimental Results · p. 12, Table III · p. 12

D. Further Experiments

WAY (base) achieves the highest evaluation score without the large parameter count of conventional grid-token models; Gradient Dropout also improves estimation. The paper explores ETA multitask learning, but inconsistent arrival-time labels limit its accuracy.

D. Further Experiments

  • Removing any represented feature type reduces performance, with navigational patterns contributing particularly to the improvement; the proposed channel aggregation method outperforms the compared alternatives.The channel-aggregation comparisons include dimension-wise concatenation, Cross Attention, and single-headed Channel Attention.Source: D. Further Experiments · p. 12, Table IV · p. 13, Table V · p. 13
  • WAY (base) achieves the highest evaluation score without a massive parameter count, while WAY-Tiny and WAY-Small outperform other benchmarks with 2.5–20.5 times fewer parameters.The comparison includes conventional grid-token models and TrAISformer.Source: Table VI · p. 13, D. Further Experiments · p. 13

VI. DISCUSSION

  • WAY-Mul reduces arrival-time errors from 4.26 days to 2.90–3.03 days while maintaining destination-estimation performance.Source: Table VII · p. 13, §VI · p. 13
  • Inconsistent arrival-time labels limit ETA estimation; congestion and shipping deadlines can add waiting time, and the authors call for a different annotation framework.Source: §VI · p. 13

VII. CONCLUSION

  • WAY processes trajectories as a 4-channel vector sequence, preserving global spatial identities, local patterns, semantics, and irregular time progression.The conclusion contrasts this detailed representation with the coarse portrayal of conventional grid-based approaches.Source: §VII · p. 14
  • Gradient Dropout improves estimation not only for WAY but also for benchmarks using conventional spatial grids.Source: §VII · p. 14
  • The authors present accurate destination and arrival-time estimation as a potential contribution to mitigating port congestion, conditional on consistent arrival-time labels.Source: §VI · p. 13, §VII · p. 14

Appendix

The appendix describes extracting destination candidates from human-entered AIS text, then validating candidates and trajectories against later vessel movements. It also outlines refinement of trajectory data.

A. Port Label Annotation

  • AIS destination text can contain misspellings, varied languages, delayed updates, and unrelated information, so candidate destinations are matched to port identifications using Damerau–Levenshtein similarity.Source: A. Port Label Annotation · p. 17
  • The candidate-extraction procedure regularizes destination expressions into N-grams and selects port identifications whose similarity score exceeds the 0.75 threshold.The score is based on Damerau–Levenshtein distance and gram length.Source: A. Port Label Annotation · p. 17
  • A later validation procedure checks nominated ports against trajectory continuity and uses position and speed over ground to confirm arrival moments.The pseudocode identifies the candidate-validation procedure and its positional-status step.Source: A. Port Label Annotation · p. 17

EXTRACT TRAJECTORY(X{1 : T})

The passage describes separating and validating port-to-port trajectory sections, then refining trajectories by detecting illogical movement edges. Validation uses positional status and human-entered port candidates, while refinement removes detected anomalies and triggers re-validation.

VALIDATE TRAJECTORY(S{1 : U})

  • Validation uses positional tags, human-entered port candidates, and a maximum three-day gap between continuous observations; acceleration must occur at least once.Tag 0 indicates acceleration and null indicates staying still; candidate and status checks help exclude sections that may combine multiple ship operations.Source: VALIDATE TRAJECTORY(S{1 : U}) · p. 18

EXTRACT TRAJECTORY(X{1 : T})

B. Trajectory Refinement

  • Trajectory refinement targets illogical movements caused by transmission errors, including anomalies beyond AIS fields’ explicit default error values.Source: B. Trajectory Refinement · p. 18
  • DBSCAN detects edge-wise anomalies from time, spatial distance, and speed-over-ground residual vectors, using epsilon 0.15.Deleting detected edges can invalidate an arrival, so the affected sequences must be re-validated.Source: B. Trajectory Refinement · pp. 18–19

Related papers

Loading 2512.13190v1…