A Graph Learning-Driven Method for Multi-Ship Collision Risk Prediction in Complex Waterways

Abstract

The proactive identification of emerging collision risks is pivotal for maritime traffic safety, particularly in congested hub ports where multi-ship encounters exhibit complex spatiotemporal dependencies. Conventional risk assessment methods, predominantly predicated on instantaneous geometric indicators, often fall short in capturing the systemic evolution of risk. To address these limitations, this study proposes an Improved Spatio-Temporal Graph Convolutional Network (IST-GCN) framework for the short-term forecasting of ship collision risk. The framework models maritime traffic as a rule-integrated dynamic interaction graph, where edge weights are adaptively modulated by navigational rules and the Collision Risk Index (CRI). By leveraging historical observation windows, the model forecasts the maximum collective risk level over a subsequent prediction horizon, categorizing traffic scenes into three ordinal levels: Low, Medium, and High. A comprehensive case study utilizing real-world Automatic Identification System (AIS) data from the core waters of Ningbo–Zhoushan Port demonstrates the efficacy of the proposed approach. The IST-GCN achieves a superior prediction Accuracy of 92.4% and an F1-score of 0.91, significantly outperforming representative baselines including Long Short-Term Memory (LSTM), Temporal Convolutional Network (TCN), and standard ST-GCN. Notably, by explicitly encoding COLREGs-based interaction logic, the framework reduces the False Alarm Rate (FAR) to 8.5% in complex crossing and merging scenarios. These findings indicate that the IST-GCN serves as an interpretable, reliable, and early-warning decision-support tool for intelligent maritime supervision and modern Vessel Traffic Services (VTS).

1. Introduction

With the continuous increase in global trade volume, the shipping industry plays a vital role in international transportation. However, maritime traffic safety in busy waterways (i.e., environments characterized by restricted physical topologies, heterogeneous traffic, and highly interdependent dynamic situations prone to cascading risks) remains a critical and persistent challenge for coastal administrations and port authorities [1,2,3,4]. In hub ports and major waterways, the traffic density is typically high and strongly heterogeneous, involving vessels with diverse sizes and operational roles [5,6]. Under such conditions, collision is not an occasional anomaly but a systemic operational concern that directly affects navigational safety and port efficiency [7,8,9]. With the ongoing development of smart shipping, there is a growing demand for intelligent supervision systems capable of providing earlier warnings and more reliable risk interpretation [10,11,12,13,14].

In reality, a fundamental difficulty in collision risk assessment lies in the dynamic and context-dependent nature of multi-ship encounters. In dense traffic, encounter configurations may rapidly shift among crossing, head-on, and overtaking situations, while being further constrained by channel geometry and local navigational regulations [15,16,17,18]. As a result, the risk level of a traffic scene cannot be adequately characterized by an instantaneous snapshot. Instead, it emerges from the temporal evolution of relative motion and the constraints of maneuvering space [19,20]. It is therefore essential to develop methods that explicitly incorporate spatiotemporal history and reflect maritime operational semantics [21,22].

Current researches on collision risk assessment are mainly based on pairwise geometric indicators derived from relative motion, such as Distance at Closest Point of Approach (DCPA) and Time to Closest Point of Approach (TCPA) [23,24]. It is important to note that these traditional geometric metrics are highly reliable and fundamental for assessing risk in open-ocean and widely used in practical decision support systems (e.g., digital maritime regulatory platforms, automated Vessel Traffic Management Information System (VTMIS) early-warning modules). However, their reactive and purely kinematic nature presents specific limitations in restricted waterways. Due to narrow fairways and spatial constraints, the geometrically projected Closest Point of Approach (CPA) may sometimes unrealistically fall on shorelines or unnavigable areas. Furthermore, their applicability is structurally limited in complex traffic scenes. Firstly, DCPA and TCPA intrinsically describe pairwise relationships and do not capture system-level interaction topology [15,25]. In practice, a vessel pair that appears low-risk in isolation may still contribute to a hazardous situation by constraining maneuvering options [26,27,28]. Secondly, pairwise indicators treat neighboring vessels independently, making it difficult to represent global traffic structure [29,30]. Furthermore, in dense waters, consideration must be given to the systemic evolution of risk, which denotes a cascading effect wherein local evasive maneuvers dynamically restrict surrounding navigable space, triggering a topological chain reaction of risk [31,32]. This suggests that pairwise risk does not necessarily translate into system-level risk in busy waterways.

With the widespread availability of Automatic Identification System (AIS) data, data-driven approaches have received increasing attention in maritime risk modeling [33,34]. Various time-series learning methods, such as Recurrent Neural Networks (RNN) and Temporal Convolutional Networks (TCN), have been developed to capture non-linear vessel motion patterns [35,36,37,38]. However, when multiple vessels interact simultaneously, the central modeling challenge extends to the representation of interaction structure. From this perspective, multi-ship traffic scenes are naturally formulated as graphs [39,40,41]. Graph Convolutional Networks (GCNs) and Spatio-Temporal Graph Convolutional Networks (ST-GCNs) provide a principled framework to aggregate neighborhood information [42,43,44].

Nevertheless, directly applying standard GCN architectures reveals a critical mismatch between generic graph construction and dynamic collision risk. Existing studies commonly construct interaction graphs using Euclidean distance thresholds [45,46]. However, spatial proximity is not equivalent to navigational threat. A relatively distant vessel on a converging course may pose a higher risk than a closer vessel moving in parallel. Consequently, purely distance-driven graphs may allocate modeling capacity to low-threat neighbors while under-representing ship encounters. Therefore, there is a critical need to develop rule-integrated and risk-aware graph constructions enabling learned representations to align more closely with practical navigational rules.

To overcome these shortcomings, this study aims to propose an Improved Spatio-Temporal Graph Convolutional Network (IST-GCN) framework for the dynamic collision risk assessment in complex waterways. The proposed framework leverages historical AIS trajectories to predict near-future scene-level collision risk. Specifically, the IST-GCN extracts non-linear spatiotemporal features, models encounters as dynamically evolving graphs with rule-integrated weighting, and aggregates node-level representations into a graph-level assessment. The objective of this research is to support VTS operators in maintaining the overall safety level of the monitored area. The proposed method aims to identify high-risk multi-ship encounters at the scene level, thereby enabling timely intervention to prevent the escalation of localized conflicts. The contribution of this study is summarized as follows:

• Firstly, a predictive collision risk assessment framework named IST-GCN is proposed to jointly model spatiotemporal vessel dynamics and multi-ship interaction topology, structurally embedding maritime operational semantics into the graph learning process.
• Secondly, a rule-integrated, dynamically weighted graph construction strategy is developed. Unlike traditional data-driven models that construct interaction graphs relying solely on Euclidean distance thresholds, this strategy adaptively modulates interaction strength based on kinematic threat cues and COLREGs obligations. Crucially, rather than merely pushing alarms earlier in time, it forces the network to accurately allocate attention to functionally threatening vessels, effectively filtering out nuisance alarms and mitigating VTS operator fatigue in complex scenes.
• Thirdly, moving beyond isolated and reactive pairwise inference, a hierarchical graph-to-scene aggregation mechanism is designed. By capturing the cascading effects of multi-ship interactions, it converts node-level state representations into a macroscopic risk diagnosis, enabling the proactive forecasting of latent risk accumulation well before critical geometric conflicts manifest.
• Finally, comprehensive case studies are conducted using real-world AIS data from the core waters of Zhoushan Port (and an extended generalization scenario in Rizhao Port). The results validate the superior capability of the proposed approach in pinpointing genuine collision threats and its spatial robustness against conventional geometric methods and representative learning-based baselines.

The remainder of this paper is organized as follows. Section 2 elaborates on the proposed IST-GCN framework, detailing the rule-integrated graph construction strategy and the hierarchical spatio-temporal learning architecture. Section 3 presents the experimental setup and conducts a comprehensive performance evaluation using real-world AIS data from Ningbo–Zhoushan Port, including comparative benchmarking and ablation studies. Section 4 discusses the methodological implications, highlighting the shift towards proactive forecasting and analyzing current limitations. Finally, Section 5 summarizes the main contributions of this study and outlines avenues for future research.

2. Materials and Methods

This section elucidates the proposed methodology, specifically focusing on the systematic translation of rule-aware multi-ship interaction modeling into a computable predictive framework. To accommodate the highly dynamic interaction topologies and inherent cascading effects within congested maritime traffic, we formulate a spatio-temporal graph model designed to characterize the evolutionary patterns of multi-vessel encounter sequences. Departing from conventional static screening approaches, the developed framework exploits historical motion trajectories to proactively forecast risk levels over a subsequent short-term horizon. This process culminates in a robust, interpretable, and scene-level classification of collision risk. The comprehensive schematic of the proposed methodological architecture is depicted in Figure 1.

2.1. Problem Formulation

We consider a bounded region of interest (ROI) within a congested waterway. At any discrete time step t, the traffic scene is populated by a set of $N_t$ vessels, denoted by ${\mathcal{V}}_t=\{1,2,\dots ,N_t\}$. In contrast to conventional instantaneous risk screening, our objective is to formulate a predictive model capable of anticipating the near-future scene-level risk within a specified horizon, conditioned on historical spatiotemporal observations and the underlying interaction topology.

Node Representation. Each vessel $i\in {\mathcal{V}}_t$ is represented as a graph node characterized by a multidimensional feature vector ${\mathbf{X}}_{i,t}\in {\mathbb{R}}^F$ derived from Automatic Identification System (AIS) reports. This vector encapsulates fundamental kinematics, including projected Cartesian coordinates $(x,y)$, speed over ground (SOG), and course over ground (COG), alongside temporal dynamics such as short-term derivatives $\Delta \mathrm{SOG}$ and $\Delta \mathrm{COG}$. These derivative features are estimated from successive AIS messages to effectively capture maneuvering trends. By concatenating the individual node features, we obtain the scene-level feature matrix ${\mathbf{X}}_t\in {\mathbb{R}}^{N_t\times F}$. To model the evolutionary trajectory of encounters, the framework utilizes a sliding historical window of length H as its input:

$${\mathcal{X}}_t^{\mathrm{hist}}=\{{\mathbf{X}}_{t-H+1},{\mathbf{X}}_{t-H+2},\dots ,{\mathbf{X}}_t\}.$$

(1)

Rule-Aware Multi-Ship Interaction Graph. The complex topological dependencies among vessels are operationalized as a time-varying directed weighted graph ${\mathcal{G}}_t=({\mathcal{V}}_t,{\mathcal{E}}_t)$, governed by an adjacency matrix ${\mathbf{A}}_t\in {\mathbb{R}}^{N_t\times N_t}$. Here, the edge weight $a_{ij,t}\ge 0$ quantifies the interaction intensity exerted from vessel j upon vessel i at time t. Departing from simplistic proximity-based graph constructions, the weights $a_{ij,t}$ are rigorously conditioned on collision-risk descriptors—such as DCPA/TCPA-derived cues—and COLREGs-inspired encounter semantics. Consequently, the graph emphasizes functionally threatening neighbors over those that are merely spatially proximal.

Predictive Learning Objective. Let $\tau$ denote the prediction horizon. The primary task is to forecast the maximum collective risk level manifested within the forthcoming interval $(t,t+\tau ]$, utilizing the historical sequence $\left({\mathcal{X}}_t^{\mathrm{hist}},{\mathcal{G}}_{t-H+1:t}\right)$. We formalize the output as an ordinal three-class label $y_{t+\tau }\in \{0,1,2\}$, corresponding to Low, Medium, and High risk levels, respectively. To establish objective ground-truth supervision, we first define a pairwise collision risk index (CRI) for vessels i and j at any time $k\in (t,t+\tau ]$ as a bounded score derived via a monotonic mapping $\varphi (·)$ of DCPA and TCPA:

$${\mathrm{CRI}}_{ij,k}=\varphi ({\mathrm{DCPA}}_{ij,k},{\mathrm{TCPA}}_{ij,k})\in \left[0,1\right].$$

(2)

The scene-level risk $R_{t+\tau }$ is then determined by aggregating the maximum CRI observed across all internal vessel pairs throughout the prediction window:

$$R_{t+\tau }=\underset{k\in (t,t+\tau ]}{max}\underset{i\ne j,i,j\in {\mathcal{V}}_k}{max}{\mathrm{CRI}}_{ij,k}.$$

(3)

Finally, the discrete risk label $y_{t+\tau }$ is assigned based on domain-informed thresholds ${\gamma }_{\mathrm{mid}}$ and ${\gamma }_{\mathrm{high}}$:

$$y_{t+\tau }=\left\{\begin{array}{cc}2\left(\mathrm{High}\right),\hfill & \mathrm{if}{R}_{t+\tau }\ge {\gamma }_{\mathrm{high}},\hfill \\ 1\left(\mathrm{Medium}\right),\hfill & \mathrm{if}{\gamma }_{\mathrm{mid}}\le R_{t+\tau }<{\gamma }_{\mathrm{high}},\hfill \\ 0\left(\mathrm{Low}\right),\hfill & \mathrm{otherwise}.\hfill \end{array}\right.$$

(4)

Accordingly, the proposed IST-GCN aims to optimize a mapping function $f_{\theta }$ such that ${\widehat{y}}_{t+\tau }=f_{\theta }({\mathcal{X}}_t^{\mathrm{hist}},{\mathcal{G}}_{t-H+1:t})$.

2.2. Rule-Integrated Graph Construction

Traditional proximity-based adjacency metrics, such as fixed-radius or k-nearest neighbors, often prove suboptimal as they fail to account for relative motion trends and navigational obligations. We therefore formulate a risk-aware, rule-integrated, and time-varying weighted adjacency matrix ${\mathbf{A}}_t=\left[a_{ij,t}\right]\in {\mathbb{R}}^{N_t\times N_t}$, where the edge weights quantify the functional threat relevance between vessels rather than simple geometric proximity.

Pairwise relative motion descriptors. For any vessel pair $(i,j)$ at time t, let their respective positions and velocities in a local Cartesian frame be denoted by ${\mathbf{p}}_{i,t},{\mathbf{p}}_{j,t}\in {\mathbb{R}}^2$ and ${\mathbf{v}}_{i,t},{\mathbf{v}}_{j,t}\in {\mathbb{R}}^2$. The relative position vector is defined as ${\mathbf{r}}_{ij,t}={\mathbf{p}}_{j,t}-{\mathbf{p}}_{i,t}$, and the relative velocity vector is ${\mathbf{u}}_{ij,t}={\mathbf{v}}_{i,t}-{\mathbf{v}}_{j,t}$. Under the assumption of short-term constant velocity, the time to closest point of approach (TCPA) is derived by projecting the relative position onto the relative velocity vector:

$${\mathrm{TCPA}}_{ij,t}={\displaystyle \frac{{\mathbf{r}}_{ij,t}·{\mathbf{u}}_{ij,t}}{\parallel {\mathbf{u}}_{ij,t}{\parallel }^2}},$$

(5)

where a positive value indicates that the vessels are converging. The distance at closest point of approach (DCPA) then represents the magnitude of the relative position vector at the time of closest approach:

$${\mathrm{DCPA}}_{ij,t}=\sqrt{\parallel {\mathbf{r}}_{ij,t}{\parallel }^2-({\mathrm{TCPA}}_{ij,t}·\parallel {\mathbf{u}}_{ij,t}{\parallel )}^2}.$$

(6)

To ensure numerical robustness, we apply ${\mathrm{TCPA}}_{ij,t}=max(0,{\mathrm{TCPA}}_{ij,t})$ to focus on future collision potential and incorporate a small epsilon to avoid division by zero.

Encounter type and rule obligation indicators. Each vessel pair is categorized into a specific encounter type $c_{ij,t}\in \left\{\mathrm{head}\text{-}\mathrm{on},\mathrm{crossing},\mathrm{overtaking},\mathrm{other}\right\}$ based on the relative bearing and course difference. To incorporate navigational semantics into the interaction topology, we introduce a simplified rule obligation indicator $o_{ij,t}\in \{0,1\}$, where $o_{ij,t}=1$ identifies vessel i as the give-way vessel relative to vessel j under COLREGs-inspired logic. This indicator enables the graph to encapsulate asymmetric interaction responsibilities, which are vital for characterizing real-world collision avoidance behaviors. In instances where rule information is unavailable or ambiguous, $o_{ij,t}$ is conservatively set to zero.

Risk-aware threat scoring. The aforementioned geometric and semantic descriptors are fused into a differentiable pairwise threat score $s_{ij,t}$, serving as a soft measure of interaction significance. We first define continuous gating functions to capture temporal and spatial proximity: $g_{\mathrm{t}}\left({\mathrm{TCPA}}_{ij,t}\right)=exp(-{\mathrm{TCPA}}_{ij,t}/\tau )$ and $g_{\mathrm{d}}\left({\mathrm{DCPA}}_{ij,t}\right)=exp(-{\mathrm{DCPA}}_{ij,t}/d_0)$, where $\tau$ and $d_0$ calibrate the sensitivity to temporal imminence and spatial separation. An encounter-type modulation term $m\left(c_{ij,t}\right)\in {\mathbb{R}}_{+}$ is then applied to reflect navigational criticality, typically following the priority order $m\left(\mathrm{head}\text{-}\mathrm{on}\right)>m\left(\mathrm{crossing}\right)>m\left(\mathrm{overtaking}\right)>m\left(\mathrm{other}\right)$. The final threat score is operationalized as:

$$s_{ij,t}=m\left(c_{ij,t}\right)g_{\mathrm{t}}\left({\mathrm{TCPA}}_{ij,t}\right)g_{\mathrm{d}}\left({\mathrm{DCPA}}_{ij,t}\right)(1+\lambda o_{ij,t}),$$

(7)

where $\lambda \ge 0$ amplifies the influence of rule-based obligations. Crucially, $s_{ij,t}$ acts as a relative interaction weight that steers the focus of subsequent message passing rather than functioning as a deterministic alarm.

Dynamic adjacency, sparsification, and normalization. To mitigate the noise associated with dense connectivity in high-traffic scenes, the graph is sparsified by retaining only the M most threatening neighbors for each target vessel i: ${\mathcal{N}}_{i,t}=TopM\left(\{s_{ij,t}:j\in {\mathcal{V}}_t,j\ne i\}\right)$. The adjacency weights are then normalized using a softmax-based approach:

$$a_{ij,t}=\left\{\begin{array}{cc}{\displaystyle \frac{exp\left(s_{ij,t}\right)}{{\sum }_{k\in {\mathcal{N}}_{i,t}}exp\left(s_{ik,t}\right)}},\hfill & j\in {\mathcal{N}}_{i,t},\hfill \\ 0,\hfill & \mathrm{otherwise},\hfill \end{array}\right.$$

(8)

yielding a row-stochastic, dynamically weighted matrix that emphasizes high-threat interactions. Self-loops can be incorporated by adding an identity matrix during graph convolution normalization. Through this iterative construction, the resulting graph topology evolves adaptively, focusing model capacity on vessels most likely to influence near-future collision risk and providing a physically grounded foundation for spatio-temporal graph learning.

2.3. IST-GCN Model Encoder

The IST-GCN architecture is operationalized through a dual-encoder framework designed to extract and integrate multi-dimensional representations of ship encounters. The comprehensive schematic of the proposed methodological architecture is depicted in Figure 2. This section elucidates the structural components of the spatial and temporal encoders, which jointly transform raw kinematic sequences into actionable risk assessments.

2.3.1. Spatial Interaction Encoder

The spatial interaction encoder is engineered to characterize the instantaneous topological dependencies and functional threats among multiple vessels within the region of interest. At each discrete time step t, the node embeddings ${\mathbf{H}}_t^{\left(l\right)}$ undergo iterative updates through a rule-integrated message-passing mechanism. This process is designed to map raw kinematic features into a high-dimensional latent space that reflects the complex interaction structure of the maritime traffic scene. To stabilize feature propagation and mitigate the numerical instability inherent in dynamically evolving graphs, we apply a symmetric normalization scheme to the self-loop augmented adjacency matrix ${\overline{\mathbf{A}}}_t={\mathbf{A}}_t+\mathbf{I}$. The resulting normalized matrix ${\tilde{\mathbf{A}}}_t={\mathbf{D}}_t^{-{\displaystyle \frac{1}{2}}}{\overline{\mathbf{A}}}_t{\mathbf{D}}_t^{-{\displaystyle \frac{1}{2}}}$ ensures that the scale of node embeddings remains consistent across varying traffic densities.

The core of this encoder lies in its multi-branch (multi-head) spatial convolution architecture, which facilitates the simultaneous extraction of diverse interaction semantics:

$${\mathbf{H}}_t^{(l+1)}=\sigma \left(\sum _{m=1}^{M_h}{\tilde{\mathbf{A}}}_t^{\left(m\right)}{\mathbf{H}}_t^{\left(l\right)}{\mathbf{W}}^{(l,m)}\right),$$

(9)

where $M_h$ denotes the number of parallel attention branches, and each branch ${\tilde{\mathbf{A}}}_t^{\left(m\right)}$ is parameterized by distinct rule-gating or threat-weighting configurations.

Unlike standard graph convolutional networks that aggregate information based on uniform spatial decay, this multi-branch design enables the network to adaptively prioritize information flow from vessels posing higher functional threats—defined by the interplay between COLREGs obligations and the CRI. Specifically, each branch can be optimized to focus on different navigational sub-contexts, such as give-way responsibilities or imminent geometric conflicts. By aggregating these refined message-passing outputs, the final spatial layer generates a comprehensive node embedding that encapsulates not only the immediate encounter semantics but also the local interaction constraints imposed by the surrounding traffic. This interaction-aware representation provides a robust spatial foundation for subsequent temporal evolution modeling.

2.3.2. Temporal Evolution Encoder

The temporal evolution encoder is responsible for modeling the dynamic progression of multi-ship encounters over the historical observation window $\left[t-T+1,t\right]$. It operates on the sequence of interaction-aware embeddings generated by the spatial encoder to identify latent risk precursors and characterize the temporal dependencies inherent in vessel maneuvers. To effectively capture non-linear motion trends and cascading maneuver responses, we employ a gated temporal convolution (TCN) architecture. This mechanism utilizes causal convolutions to ensure that the risk assessment at time t is strictly conditioned on past observations, thereby preserving the temporal causality of maritime traffic evolution. The gated activation units are operationalized using one-dimensional causal convolutions. To concisely capture both feature extraction and noise filtration, the final temporal representation ${\mathbf{Z}}_t$ is formulated as:

$${\mathbf{Z}}_t=tanh({\mathbf{W}}_f\ast {\mathbf{H}}_t^{\left(L_s\right)})\odot \sigma ({\mathbf{W}}_g\ast {\mathbf{H}}_t^{\left(L_s\right)}),$$

(10)

where ∗ denotes the temporal convolution operator, and ${\mathbf{W}}_f,{\mathbf{W}}_g$ represent the trainable weight matrices for the feature and gate branches, respectively. Here, ⊙ is the element-wise Hadamard product. The tanh function extracts complex motion features, while the sigmoid function acts as a temporal gate to suppress noise from irregular AIS reporting intervals.

To synthesize a unified scene-level assessment from these heterogeneous temporal representations, a permutation-invariant aggregation mechanism is required. We implement an attention-based pooling strategy to prioritize risk-dominant vessels, effectively mirroring the cognitive heuristic where maritime operators focus on the most threatening ships in a complex scene. The aggregation is succinctly formulated as:

$${\mathbf{g}}_t^{\mathrm{att}}=\sum _{i\in {\mathcal{V}}_t}{\alpha }_{i,t}{\mathbf{z}}_{i,t},{\alpha }_{i,t}=softmax\left(e_{i,t}\right),$$

(11)

where ${\alpha }_{i,t}$ reflects the learned significance of vessel i in driving the collective risk, derived from its latent attention score $e_{i,t}$. This hierarchical encoding process culminates in a MLP prediction head followed by a softmax layer, which outputs the probability distribution ${\widehat{\mathbf{p}}}_{t+\tau }$ over discrete risk levels (Low, Medium, and High). By coupling the rule-integrated spatial interaction topology with long-term temporal evolution, the IST-GCN encoder facilitates proactive collision risk forecasting that is both physically grounded and computationally efficient for real-time VTS applications.

2.4. IST-GCN Model Decoder

The functionality of the scene-level risk decoder extends beyond simple feature aggregation; it serves as a semantic bridge that distills microscopic, node-level interaction states into a macroscopic, actionable risk diagnosis. Given the heterogeneous node embeddings ${\mathbf{z}}_{i,t}\in {\mathbb{R}}^D$ produced by the spatio-temporal encoders, the decoder faces two fundamental modeling challenges: the time-varying cardinality of the vessel set $N_t$ (due to ships entering or leaving the ROI) and the absence of a canonical node ordering. To address these constraints, we engineer a permutation-invariant decoding architecture that synthesizes a fixed-length scene representation, ensuring that the risk assessment remains structurally robust regardless of traffic density fluctuations.

2.4.1. Risk-Focused Attention Aggregation

Conventional pooling operators, such as global average pooling, often dilute critical risk signals by treating all vessels uniformly, effectively “averaging out” the threat posed by a single dangerous target in a sea of safe vessels. Conversely, global max pooling may ignore secondary but significant threats. To overcome these limitations and mirror the cognitive heuristics of VTS operators—who selectively focus their attention on high-threat targets amidst complex backgrounds—we implement a parameterized, risk-focused attention mechanism.

Each vessel node is first projected into a latent scalar score representing its contribution to the global risk context. This is operationalized through a learned alignment function:

$$e_{i,t}={\mathbf{q}}^{T}tanh({\mathbf{W}}_a{\mathbf{z}}_{i,t}+{\mathbf{b}}_a),$$

(12)

where ${\mathbf{W}}_a\in {\mathbb{R}}^{D^{\prime }\times D}$ and ${\mathbf{b}}_a\in {\mathbb{R}}^{D^{\prime }}$ are trainable projection parameters, and $\mathbf{q}\in {\mathbb{R}}^{D^{\prime }}$ serves as a learnable “context query vector” that encapsulates the global pattern of a high-risk scenario. The scalar scores are subsequently normalized via a softmax function to produce the attention distribution ${\alpha }_{i,t}$, ensuring ${\sum }_i{\alpha }_{i,t}=1$:

$${\alpha }_{i,t}={\displaystyle \frac{exp\left(e_{i,t}\right)}{{\sum }_{k\in {\mathcal{V}}_t}exp\left(e_{k,t}\right)}}.$$

(13)

The final scene-level embedding ${\mathbf{g}}_t^{\mathrm{att}}$ is then synthesized as a weighted sum ${\mathbf{g}}_t^{\mathrm{att}}={\sum }_{i\in {\mathcal{V}}_t}{\alpha }_{i,t}{\mathbf{z}}_{i,t}$. This mechanism allows the model to perform a “soft selection” of salient features, effectively amplifying signals from vessels involved in rule-violating or geometrically converging encounters while suppressing noise from non-interacting traffic.

2.4.2. Probabilistic Prediction and Interpretability

The synthesized scene embedding ${\mathbf{g}}_t^{\mathrm{att}}$, which now encapsulates the most salient spatiotemporal risk characteristics, is fed into a prediction head. This component consists of a Multi-Layer Perceptron (MLP) designed to project the high-dimensional graph representation into the target class space. The final output is generated via a softmax activation layer, yielding a normalized probability distribution over the discrete risk categories (Low, Medium, High):

$${\widehat{\mathbf{p}}}_{t+\tau }=softmax({\mathbf{W}}_c·ReLU({\mathbf{W}}_p{\mathbf{g}}_t^{\mathrm{att}}+{\mathbf{b}}_p)+{\mathbf{b}}_c),$$

(14)

where ${\mathbf{W}}_p,{\mathbf{b}}_p$ and ${\mathbf{W}}_c,{\mathbf{b}}_c$ denote the weights and biases of the projection and classification layers, respectively.

This hierarchical formulation aligns the model’s predictive output directly with the operational requirements of maritime supervision, providing a probabilistic measure of confidence for each risk level. Crucially, the learned attention weights $\left\{{\alpha }_{i,t}\right\}$ offer intrinsic interpretability to the decision support system. By visualizing these weights as a “risk saliency map,” VTS authorities can immediately identify which specific vessels or interaction clusters are driving the predicted alarm, thereby transforming the model from a “black box” into a transparent tool for proactive safety intervention.

3. A Graph Learning-Driven Method for Multi-Ship Collision Risk Prediction in Complex Waterways: Results from Case Studies

3.1. Dataset and Experimental Setup

3.1.1. Data Acquisition and Preprocessing

The empirical validation of the proposed IST-GCN framework is conducted using a large-scale Automatic Identification System (AIS) dataset acquired from the core navigational waters of Ningbo–Zhoushan Port, recognized as one of the world’s premier maritime hubs. This study area encompasses major fairways and high-density convergence zones characterized by frequent crossing, overtaking, and merging encounters among heterogeneous vessel types. Such complex traffic conditions render this dataset particularly advantageous for benchmarking multi-ship collision risk detection methodologies that prioritize interaction topology and collective risk evolution. The data span a continuous observation period from June to August 2023, capturing a diverse range of traffic states from off-peak to peak congestion. To ensure modeling fidelity, a rigorous data cleaning protocol was applied to the raw AIS messages—which contain timestamps, vessel identifiers, coordinates, speed over ground (SOG), and course over ground (COG)—to excise anomalies such as missing coordinates, duplicate timestamps, and kinematic outliers. The dataset and code are publicly available at https://anonymous.4open.science/r/risk_detection-74D5 (accessed on 26 March 2026).

3.1.2. Spatiotemporal Sample Construction

Subsequent to preprocessing, the continuous vessel trajectories were structured into fixed-length spatiotemporal sequences via a sliding-window mechanism. Specifically, each sample comprises a 60-s historical observation window (H) serving as the model input, coupled with a subsequent 30-s prediction horizon ($\tau$) for near-future risk labeling. Within each temporal window, all vessels simultaneously present in the Region of Interest (ROI) are modeled as nodes within a dynamic interaction graph, enabling the IST-GCN to extract features from historical encounter evolution rather than relying solely on instantaneous states. A statistical summary of the dataset and the spatiotemporal sample construction is presented in

Table 1. Summary of AIS Dataset and Spatiotemporal Sample Construction.

Item	Value
Study area	Ningbo–Zhoushan Port (core waters)
Data period	June–August 2023
Total duration	92 days
Number of vessels	2846
Total AIS records	18,372,591
Average reporting interval	10.4 s
Main vessel types	Container ship, bulk carrier, tanker, others
Time window length	60 s
Prediction horizon	30 s
Number of spatiotemporal samples	123,608
Average vessels per sample	8.7

, with the study area visualized in Figure 3.

3.1.3. Ground Truth Labeling and Implementation

To establish objective ground truth for supervised learning, scene-level risk labels are synthesized using a CRI derived from standard DCPA and TCPA metrics. To integrate these continuous collision risks into the discrete test dataset, the raw AIS trajectories are first segmented into interaction scenarios based on spatiotemporal proximity. For every spatiotemporal sample, the CRI is computed at each timestamp for all vessel pairs within the prediction interval $\left[t+1,t+\tau \right]$, forming a time-series representation of objective navigation risk. The scene-level ground truth is then defined by the maximum CRI value observed among all internal interactions during this horizon. Accordingly, a three-tier ordinal risk schema is adopted: (i) High risk ($y=2$) corresponds to scenarios where $max\left(\mathrm{CRI}\right)\ge 0.6$, indicating imminent collision threats; (ii) Medium risk ($y=1$) denotes intervals where $0.4\le max\left(\mathrm{CRI}\right)<0.6$, reflecting hazardous situations necessitating vigilance; and (iii) Low risk ($y=0$) covers normal navigation states with $max\left(\mathrm{CRI}\right)<0.4$. This labeling strategy bridges physical motion parameters with high-level supervisory categories.

Consistent with real-world maritime traffic distributions, the dataset exhibits significant class imbalance, with approximately 82% of samples categorized as Low risk, 13% as Medium risk, and 5% as High risk. To mitigate potential bias towards the majority class, a weighted cross-entropy loss function was employed during training, assigning higher penalty weights to the minority High- and Medium-risk samples. The proposed framework was implemented using the PyTorch (version 1.13.1) deep learning library and trained via the Adam optimizer with an initial learning rate of 0.001 and a batch size of 64. All computational experiments were executed on a high-performance workstation equipped with an NVIDIA RTX 4090 GPU and 64 GB of system memory.

3.2. Qualitative Analysis: Risk Evolution in Multi-Ship Encounters

To empirically validate the interpretability and proactive capabilities of the IST-GCN framework, we performed a qualitative analysis on representative multi-ship encounter scenarios extracted from the test dataset. Unlike conventional geometric methods (e.g., DCPA/TCPA) that rely on instantaneous snapshots and rigid thresholds, the proposed model leverages historical spatiotemporal dependencies and rule-integrated topology to perceive emerging risk trends within the future prediction window $\tau$. This section examines the model’s performance in two distinct yet critical encounter types: complex crossing and cascading merging situations.

3.2.1. Scenario I: Preemptive Warning in Rule-Constrained Crossing

We first analyze a four-vessel crossing scenario to demonstrate the model’s sensitivity to navigational rules. As visualized in the trajectory plot of Figure 4a, the interaction scenario involves four vessels (Ship A–D) with intersecting paths. Initially, the interaction graph exhibited weak topological coupling. However, as the vessels converged, the rule-integrated adjacency matrix dynamically modulated the edge weights, progressively shifting the model’s attention toward the vessel pairs with clear COLREGs-based conflicts.

The quantitative progression of the risk assessment is detailed in

Table 2. Quantitative Risk Evolution Metrics in a Typical Crossing Scenario.

Time Step	Min DCPA (m)	Min TCPA (s)	CRI (GT)	Predicted Risk Score	Predicted Label
$T-30\mathrm{s}$	1240.5	320.4	0.12	0.15	0 (Low)
$T-15\mathrm{s}$	980.2	185.6	0.28	0.46	1 (Medium)
T (current)	750.4	94.2	0.45	0.72	1 (Medium)
$T+15\mathrm{s}$	420.8	45.1	0.82	0.88	2 (High)
$T+30\mathrm{s}$	310.5	10.4	0.91	0.94	2 (High)

. A critical divergence between the proposed method and the baseline geometric indicators is observed. While the minimum DCPA and TCPA values remained within nominally safe ranges at earlier timestamps, the IST-GCN successfully identified a latent risk accumulation. This proactive capability is vividly illustrated in Figure 4b, where the IST-GCN risk curve (solid orange line) peaks significantly earlier and higher than the Traditional CRI curve (dashed black line) over the prediction horizon. This confirms that the spatiotemporal encoder effectively internalizes maneuvering intentions, enabling the detection of hazardous states well before they manifest as critical geometric proximity.

3.2.2. Scenario II: Capturing Cascading Effects in Merging Traffic

The robustness of the model is further examined in a multi-ship merging scenario involving three vessels (Ship E, F, and G), as shown in Figure 5a. This scenario is characterized by constrained navigable space and interdependent maneuvering. In this case, when a primary vessel initiated a course alteration to avoid a close-quarters situation, the IST-GCN captured the resultant cascading effects imposed on neighboring vessels. Through the mechanism of message passing on the dynamically weighted graph, risk dependencies were propagated across the network, enabling the model to recognize coordinated maneuver constraints.

The superiority of the proposed framework is evident in the risk level prediction curves shown in Figure 5b. The IST-GCN risk score rises rapidly, crossing the high-risk threshold (0.6) significantly earlier than the Traditional CRI metric. Specifically, the proposed model identifies the High Risk ($y=2$) state approximately 12 s in advance of the baseline method. This additional lead time is operationally significant, affording maritime authorities and VTS operators a broader reaction window to initiate monitoring or traffic organization measures.

3.2.3. Scenario III: Validation on a Near-Miss Event

To directly validate whether the proposed IST-GCN merely approximates instantaneous geometric indicators or possesses an independent predictive capability, a real-world near-miss encounter was extracted from the testing dataset for quantitative analysis. This scenario involves a critical crossing situation where evasive action was taken at the last possible moment, physically representing a high-risk near-miss event. Specifically, according to the near-miss definition in [47], we identified this encounter within the study period, featuring a 150-m bulk carrier and a 200-m container ship converging at a narrow intersection with a relative bearing of approximately 75°. Both vessels maintained speeds exceeding 12 knots before the give-way vessel executed an abrupt starboard turn to avert a collision.

As detailed in

Table 3. Risk Evolution Comparison in a Real-World Near-Miss Event.

Time Step	Min DCPA (m)	Min TCPA (s)	Traditional CRI	IST-GCN Score	Predicted Label
$T_{nm}-80\mathrm{s}$	1150.2	185.4	0.22	0.38	0 (Low)
$T_{nm}-60\mathrm{s}$	890.5	120.6	0.35	0.65	2 (High)
$T_{nm}-40\mathrm{s}$	640.8	85.2	0.48	0.81	2 (High)
$T_{nm}-20\mathrm{s}$	320.4	45.8	0.72	0.93	2 (High)
$T_{nm}$	180.5	15.2	0.94	0.98	2 (High)

, the traditional geometric metrics (DCPA and TCPA) are inherently reactive. As the encounter progresses toward the near-miss point ($T_{nm}$), the traditional CRI only crosses the critical high-risk threshold (>0.6) at $T_{nm}-20\mathrm{s}$, lagging significantly behind the unfolding situation. In contrast, by leveraging historical maneuvering sequences and rule-integrated topological interactions, the IST-GCN model anticipates the latent collision threat much earlier. The predicted risk score transitions into the High Risk state at $T_{nm}-60\mathrm{s}$, providing an additional 40 s of early-warning lead time compared to the traditional geometric indicator. This empirical evidence confirms that the IST-GCN is capable of proactively forecasting collision risk dynamics rather than merely fitting the mathematical formulation of existing geometric metrics.

3.3. Robustness of Our Proposed Method

To verify the robustness of the proposed method, an additional validation experiment was conducted using an AIS dataset collected from Rizhao Port. Compared to the primary study area, Rizhao Port presents distinct fairway topologies and traffic compositions, notably featuring a higher density of deep-draft bulk carriers and different intersection geometries. The complex traffic trajectories and the corresponding trajectories heatmap in this new environment are visualized in Figure 6. The pre-trained models were evaluated directly on the Rizhao Port testing set to assess their generalization capabilities. As summarized in

Table 4. Performance Metrics on the Rizhao Port Dataset.

Method	Accuracy (%)	Precision (%)	Recall (%)	F1-Score	FAR (%)
ST-GCN	83.1	82.0	80.5	0.812	15.8
IST-GCN (Ours)	90.8	89.7	89.3	0.895	9.2

, the proposed IST-GCN maintains robust predictive performance in the alternative waterway. It achieved an Accuracy of 90.8% and an F1-score of 0.895. Although this represents a slight, expected attenuation compared to its performance in Ningbo-Zhoushan Port (Accuracy of 92.4%), the IST-GCN continues to significantly outperform the baseline Standard ST-GCN, which experienced a more severe performance degradation (F1-score dropping to 0.812). Crucially, the IST-GCN successfully kept the FAR suppressed at 9.2% in the new environment. This demonstrates that the rule-integrated graph construction mechanism successfully extracts fundamental navigational logic—such as COLREGs obligations and relative kinetic trends—that is universally applicable across different maritime regions, thereby confirming the framework’s strong generalizability and potential for broader practical engineering applications.

3.4. Quantitative Analysis: Comparative Benchmarking

To rigorously quantify the predictive efficacy of the proposed IST-GCN framework, a comprehensive benchmarking study was executed against a spectrum of representative baseline models, ranging from traditional geometric heuristics to state-of-the-art deep learning architectures. The evaluation was conducted on the Ningbo–Zhoushan Port testing set, employing a 60-s historical observation window to forecast scene-level risk categories over a subsequent 30-s horizon. The comparative baselines include: (i) Geometric Heuristics: a threshold-based method relying on instantaneous DCPA/TCPA; (ii) Sequence Models: RNN and LSTM, which capture temporal dependencies but ignore spatial interaction; (iii) Spatio-Temporal Baselines: TCN and Standard ST-GCN, which model joint dynamics but lack rule-integrated semantic guidance.

3.4.1. Model Performance Evaluation

The quantitative performance metrics for the three-tier risk classification task are systematically summarized in

Table 5. Overall Risk Prediction Performance for Three-Tier Classification (Low, Medium, High).

Method	Accuracy (%)	Precision (%)	Recall (%)	F1-Score
Geometric (DCPA/TCPA)	72.4	68.5	65.4	0.669
RNN	76.8	74.2	73.5	0.738
LSTM	81.2	79.4	78.2	0.788
TCN	84.3	82.7	82.1	0.824
Standard ST-GCN	85.6	84.1	83.3	0.837
IST-GCN (Ours)	92.4	91.5	90.8	0.911

. The proposed IST-GCN demonstrates superior performance across all evaluation indicators, achieving a peak Accuracy of 92.4% and an F1-score of 0.911. This represents a substantial improvement over the traditional Geometric method (F1 = 0.669), which struggles in dense traffic due to its inability to account for maneuvering trends and rule-based priorities.

Furthermore, IST-GCN significantly outperforms the sequence-based models (RNN/LSTM) and the generic ST-GCN. While the Standard ST-GCN (F1 = 0.837) improves upon isolated sequence modeling by incorporating spatial topology, it is constrained by its reliance on Euclidean distance for graph construction. This often leads to the aggregation of noise from spatially proximate but navigationally irrelevant vessels. In contrast, by incorporating the rule-integrated adjacency matrix, IST-GCN effectively suppresses these non-threatening interactions, enabling the model to focus its capacity on genuinely hazardous encounters.

3.4.2. Performance in Safety-Critical Scenarios

In the context of maritime supervision, the cost of misclassifying a high-risk situation (False Negative) is disproportionately higher than a false alarm. Therefore, we specifically analyze the model’s sensitivity to Medium Risk ($y=1$) and High Risk ($y=2$) categories. As visualized in the normalized confusion matrix (Figure 7), the proposed IST-GCN achieves a commendable Recall of 90.8% for High-Risk events (shown as 0.91 in the matrix). The strong diagonal dominance, particularly in the high-risk category, ensures that the system provides reliable early warnings for potentially catastrophic collisions, a capability that is critical for trust-based VTS decision support.

In contrast, while sequence-based baselines often struggle to distinguish developing threats from normal traffic, the IST-GCN demonstrates superior discriminative power. Specifically, the misclassification rate of Medium-Risk scenes as Low Risk is suppressed to just 5% (0.05), minimizing the risk of overlooking hazardous situations. This performance gain can be directly attributed to the rule-aware topology, which preserves faint but critical risk signals that might otherwise be drowned out by dominant low-risk traffic patterns.

3.5. Ablation Study

To isolate and quantify the specific contributions of the core modules within the IST-GCN framework, a rigorous ablation study was conducted by systematically dismantling the rule-integrated adjacency modeling and the CRI-guided interaction weighting mechanisms. The full IST-GCN model was benchmarked against three distinct variants to evaluate their individual impact on risk prediction accuracy and alarm reliability: a Base-GCN that constructs interaction graphs solely based on Euclidean distance thresholds; a CRI-GCN that incorporates risk-based weighting but lacks COLREGs-inspired encounter modulation; and a Rule-GCN that encodes static navigational obligations but excludes time-varying risk magnitude.

The comparative results, systematically summarized in

Table 6. Ablation Analysis of Rule-Aware and CRI-Guided Components.

Configuration	Accuracy (%)	Precision (%)	Recall (%)	F1-Score	FAR (%)
Base-GCN (distance-only)	85.6	84.1	83.3	0.837	14.3
CRI-GCN (w/o COLREGs modulation)	88.4	87.2	86.5	0.868	11.2
Rule-GCN (w/o CRI weighting)	89.1	88.5	87.8	0.881	10.5
IST-GCN (full model)	92.4	91.5	90.8	0.911	8.5

, demonstrate that the full IST-GCN architecture consistently outperforms all ablated variants across every evaluation metric. Most notably, compared to the distance-only Base-GCN, the full model yields a 6.8% improvement in Accuracy and elevates the F1-score from 0.837 to 0.911. This substantial margin confirms that integrating domain knowledge transforms the graph from a simple spatial descriptor into a functional risk encoder. A critical insight further emerges from the False Alarm Rate (FAR) analysis. Removing the rule-aware mechanism (as seen in the CRI-GCN variant) causes the FAR to deteriorate significantly, rising from 8.5% to 11.2%. This empirical evidence underscores that risk cues based purely on geometric approach are insufficient to distinguish between benign passing maneuvers and genuine collision threats. Moreover, while the Rule-GCN variant reduces false alarms compared to the baseline, its inferior Recall indicates that static rule masks fail to capture the temporal urgency of evolving encounters. The CRI-guided weighting bridges this gap, enabling the model to sensitize its focus on imminent threats, thus preserving a high Recall of 90.8%.

The mechanism underlying these performance gains is visually corroborated in Figure 8, which maps the learned attention weights distributed across neighboring vessels. As observed in Figure 8a, the Base-GCN exhibits an “attention diffusion” phenomenon, indiscriminately assigning high weights to all proximate vessels, including those in non-threatening sectors such as vessels trailing astern. This inability to filter navigationally irrelevant noise directly contributes to its elevated false alarm rate. In sharp contrast, Figure 8b demonstrates the semantic selectivity of the IST-GCN, where the model strategically concentrates its attention capacity on the vessel located in the critical “give-way” sector (highlighted in green). By structurally suppressing weights for vessels that do not pose a functional threat despite their spatial proximity, the model achieves a cognitive alignment with maritime operational logic. This interpretability not only explains the superior statistical performance but also verifies that the model has learned physically meaningful interaction patterns rather than spurious correlations.

4. Discussion

The empirical results presented in Section 3 substantiate that the proposed Improved Spatio-Temporal Graph Convolutional Network (IST-GCN) constitutes a robust and physically grounded framework for maritime collision risk assessment. Beyond the quantitative performance gains in accuracy and recall, these findings hold significant implications for intelligent maritime supervision and the future operation of Vessel Traffic Services (VTS). This section critically discusses the methodological advantages of the proposed approach, particularly its role in shifting the paradigm from reactive detection to proactive forecasting, while also addressing current limitations and prospective research directions.

4.1. Proactive Forecasting via Rule-Integrated Topology

A pivotal contribution of this study is the methodological transcendence from reactive, geometry-based risk detection to proactive, rule-integrated risk forecasting. Conventional collision avoidance systems predominantly rely on instantaneous snapshots of DCPA/TCPA metrics, which are inherently reactive—triggering alarms only after a safety domain has been violated or when a hazardous encounter is already geometrically inevitable. Such approaches often fail to account for the large inertia and delayed hydrodynamic response of merchant vessels, leaving insufficient time for effective intervention. In contrast, the IST-GCN leverages a 60-s historical observation window to forecast the risk category over a subsequent 30-s horizon. This look-ahead capability effectively internalizes the temporal trends of vessel maneuverability, enabling VTS operators to identify emerging hazardous states before they manifest as critical spatial conflicts.

However, the reliability of such proactive forecasting is fundamentally contingent on the quality of the underlying interaction graph. Traditional proximity-based graph models tend to allocate uniform attention to all spatially adjacent vessels, resulting in the indiscriminate aggregation of noise from navigationally benign neighbors (e.g., parallel sailing). The proposed IST-GCN addresses this by embedding a rule-integrated semantic filter directly into the topological construction. By synthesizing COLREGs obligations with the CRI, the model selectively amplifies interactions that are kinematically converging or rule-violating while suppressing those that are operationally safe. In addition, in practical VTS deployment, the IST-GCN framework is strategically optimized to balance True Positives and False Positives. It maximizes TP through a weighted loss function to ensure maritime safety, while simultaneously suppressing FP via the rule-integrated graph to prevent operator alarm fatigue. Furthermore, the additional generalization experiment in Rizhao Port confirms that this optimal FP/TP balance is highly robust across different waterway topologies without overfitting.

This synergistic integration of temporal forecasting and rule-based spatial filtering yields a dual advantage: it achieves a high sensitivity to genuine threats (90.8% Recall) while significantly mitigating the false alarm problem endemic to dense traffic monitoring (8.5% FAR). Consequently, the framework not only provides an early warning buffer for timely decision-making but also ensures structural alignment with maritime operational logic, fostering greater trust among human operators who require automated tools that think in accordance with established navigational rules. Furthermore, the practical deployment of the IST-GCN in VTS centers is strongly supported by its real-time computational efficiency. As empirical tests indicate, the model achieves an average inference time of 149.5 ms per scene. Given that standard AIS data update intervals typically range from 2 to 10 s, this rapid processing throughput ensures seamless, zero-latency risk monitoring. In an operational context, this computational lightness allows the framework to be efficiently integrated into existing Vessel Traffic Management Information Systems (VTMIS) as a backend module. It ensures that the proactive risk alerts generated by the model can be delivered to VTS operators synchronously with live traffic feeds, without imposing prohibitive hardware burdens on port authorities.

4.2. Limitations and Future Research Directions

Despite the demonstrated efficacy of the IST-GCN framework in proactive risk forecasting, several intrinsic limitations warrant objective acknowledgment to guide subsequent research efforts. The primary constraint arises from the model’s dependence on unimodal Automatic Identification System (AIS) data streams. While AIS provides essential kinematic baselines, it remains vulnerable to signal packet loss, multipath interference, and intentional spoofing, particularly in congested waterways or satellite-denied environments. In scenarios involving non-cooperative targets where transponders are disabled or malfunctioning, the structural integrity of the interaction graph is compromised, leading to topological fragmentation and potential risk underestimation. To mitigate this vulnerability, future iterations of the framework must pivot towards multi-modal sensor fusion strategies. The integration of heterogeneous data sources—specifically shore-based marine radar for non-cooperative target tracking and CCTV visual feeds for near-field semantic verification—would significantly bolster system robustness against single-source data degradation.

Furthermore, the framework’s applicability to autonomous navigation is currently limited. First, its representation of COLREGs is simplified, relying on basic encounter classifications rather than complex maneuvering behaviors. Second, it focuses solely on risk prediction without generating actionable collision avoidance or route planning strategies. Future work will address this by integrating fine-grained rule representations and coupling the model with decision-making modules for closed-loop autonomous control. Finally, while the current IST-GCN framework focuses on localized, short-term risk forecasting in port areas rather than entire routes, frequent local evasive maneuvers can induce cumulative navigational delays. Therefore, future research could integrate these micro-level risk assessments with macroscopic scheduling systems to optimize Estimated Time of Arrival (ETA) and Requested Time of Arrival (RTA) strategies, thereby enhancing overall port logistical efficiency.

5. Conclusions

In order to achieve proactive identification of emerging hazardous states and provide adequate early-warning lead time, this study makes an attempt to propose an Improved Spatio-Temporal Graph Convolutional Network (IST-GCN) framework for the short-term forecasting of ship collision risk in complex maritime environments. By explicitly integrating maritime domain knowledge (i.e., CRI and COLREGs) into a spatiotemporal graph learning framework, the proposed approach aims to bridge the gap between individual trajectory analysis and system-level interaction topology.

With the AIS data from the core waters of Ningbo–Zhoushan Port, one case study is created to validate the effectiveness of the proposed methodology in this study. Results show that the proposed IST-GCN achieves a peak accuracy of 92.4% and an F1-score of 0.911 in the three-tier risk classification task. Compared to the representative deep learning baselines (e.g., standard ST-GCN), the proposed methodology could contribute to an improvement of approximately 7.5 percentage points. This suggests that the proposed IST-GCN framework could consistently outperform both traditional geometric indicators and generic deep learning models.

Findings of this study reveal that the incorporation of rule-aware dynamic adjacency weighting could substantially reduce unnecessary warnings in complex encounter scenarios. Specifically, the False Alarm Rate (FAR) is reduced to 8.5% where proximity-based models tend to be overly conservative. This shows that the proposed methodology could provide a more reliable and accurate estimate for the vessel collision risk in dense traffic waters. Moreover, results show that the average inference time is 149.5 ms per scene with a processing throughput of 214 frames per second. This indicates that the proposed framework is well suited for real-time deployment in operational maritime supervision systems, such as Vessel Traffic Services (VTS).

Nevertheless, the forecasting performance is currently dependent on the quality of single-source AIS data. Therefore, future work will focus on integrating multi-source sensing data (e.g., radar) to improve robustness. Furthermore, integrating the proposed method with the collision avoidance decision-making functions of MASS would be an interesting direction for future research.