PaperlayerView exact revision on arXiv

Source-linked AI summary

Evaluating Graph Neural Networks for Change-Criticality Classification in Maritime Navigation Charts

Abhishek Potnis, Jacob Arndt

arXiv:2609.02996v1cs.LGcs.AIstat.ML

TL;DR

Classifying whether ENC changes pose navigational-safety risks is challenging, and the best graph representation and GNN configuration for the task are unclear. The paper represents old and new ENCs as paired graphs and evaluates GNNs for criticality classification. Graph-based approaches improve classification, with GraphSAGE leading the tested architectures on accuracy and macro-F1.

  • Problem

    It is unclear which message-passing operations, GNN configurations, and graph representations best classify ENC changes by navigational-safety criticality, a consequential chart-maintenance task.

    Source: §I · p. 1

  • Method

    The authors encode old and new ENC datasets as graphs with spatial and semantic relationships as edges, then compare GNN architectures for change-criticality classification.

    Source: §I · p. 1, B. Graph Dataset Construction · p. 2

  • Results

    GraphSAGE achieves the highest accuracy (up to 0.8766 ± 0.0050) and macro-F1 (up to 0.8632 ± 0.0046), outperforming the MLP baseline and other tested GNN variants.

    Source: B. Experimental Results · p. 3, §V · p. 4

  • Takeaways & Limitations

    The findings indicate that ENC change criticality depends on spatial and relational context as well as object class, and that graph structure is valuable for this classification task.

    Source: §IV · p. 3, §V · p. 4

Abstract

from arXiv · show

Graph neural networks (GNNs) are a class of neural networks suitable for learning on graph-structured data. Their application to spatial data is a natural extension, however its relatively unclear which message-passing operations, architectural configurations, and graph representation is best suited for classifying changes to objects in electronic navigational charts (ENCs)--geospatial vector datasets used for marine navigation. Maintaining these datasets is a challenge, and categorizing changes to objects in the ENC based on their significance to navigational safety is of particular importance. Here, we propose to represent these vector navigation datasets as a graph structure where the spatial objects serve as nodes and their spatial and semantic relationships form edges. We encode both the old ENC dataset and new ENC dataset into a pair of graphs and frame the task as a graph-pair classification problem. Building on this representation, we investigate the use of GNN architectures to classify whether the encoded graphs constitutes a critical or non-critical risk to navigational safety. We train and evaluate several GNN architectures and model configurations on ENC changes reviewed by maritime experts. Our results demonstrate that graph-based representations improve the classification of ENC updates, providing a scalable approach for automating or improving ENC maintenance workflows.

I. INTRODUCTION

ENCs are complex marine-navigation datasets, and deciding whether their changes pose critical safety risks is challenging and time-consuming. The paper proposes graph-based classification and evaluates GNN architectures for this task.

I. INTRODUCTION

  • ENC changes must be judged for critical or non-critical navigational-safety risk, a time-consuming task requiring expertise in marine navigation, cartography, and the ENC specification.ENCs contain numerous records representing real-world entities relevant to safe navigation.Source: §I · p. 1
  • The authors represent ENC objects as graph nodes and spatial and ENC-standard relationships as edges, then use GNNs to classify changes.The study compares architectures on a fixed ENC graph dataset under a consistent experimental setup.Source: §I · p. 1
  • The paper systematically evaluates multiple GNN architectures, depth, and connectivity for graph-based ENC change-criticality classification.Source: §I · p. 2

II. METHODS

The task is to determine whether ENC changes threaten navigational safety. It is framed as predicting criticality from old and new ENC graphs alongside the identified changes.

A. Problem Formulation

  • The prediction target is whether an ENC change is a critical or non-critical risk to navigational safety at sea.Source: A. Problem Formulation · p. 2
  • Inputs are an old ENC dataset, a new ENC dataset, and a change identifying objects that were added, deleted, or modified.Source: A. Problem Formulation · p. 2
  • The task is framed as graph-pair classification, using old and new graph representations to capture spatial and relational context.Source: A. Problem Formulation · p. 2

B. Graph Dataset Construction

ENC datasets and changes are represented as paired graphs, with ENC objects as nodes and spatial and standard-defined relationships as edges. The node features are deliberately minimal, and the edge types are unified for this study.

B. Graph Dataset Construction

  • Each old and new ENC dataset becomes a graph whose nodes are ENC objects and whose typed edges represent spatial and ENC-standard relationships.The graph is defined as G = (V, E, R), with nodes V, edges E, and edge types R.Source: B. Graph Dataset Construction · p. 2
  • Each ENC object is a node, represented only by its categorical object-class identifier to isolate graph structure from richer attributes.Source: B. Graph Dataset Construction · p. 2
  • The edge set includes four S-57 semantic relation types and three spatial topological types, merged into one unified set rather than relation-specific message passing.The spatial types are overlaps, touches, and equals.Source: B. Graph Dataset Construction · p. 2

C. Architectures

The experiments compare a non-graph MLP baseline with several GNN architectures. The baseline omits spatial context and ENC relationships, while the graph models aggregate neighborhood information in different ways.

C. Architectures

  • The MLP baseline uses four 256-dimensional linear layers and learned embeddings for object class and change type, without spatial or ENC-relationship information.Source: C. Architectures · p. 2
  • The evaluated graph models include GCN, which uses normalized-adjacency aggregation, and GAT, which learns edge-attention weights for neighboring nodes.Source: C. Architectures · p. 2
  • GraphSAGE aggregates neighborhood information by mean or pooling, while TransformerConv applies multi-head attention over local graph neighborhoods.Source: C. Architectures · p. 2

D. GNN Architecture Considerations

The study varies GNN depth and layer-wise residual connections while keeping graph inputs consistent, to isolate their effects on predictive performance.

D. GNN Architecture Considerations

  • The experiments vary GNN layer count and add layer-wise residual connections to study architectural effects.The comparison uses the same node features, unified edge set, and global features across architectures.Source: D. GNN Architecture Considerations · p. 2
  • Changing layer count alters nodes’ receptive fields and information propagation, affecting performance and the risk of over-smoothing.Source: D. GNN Architecture Considerations · p. 2
  • Layer-wise residual connections are used to support stable training in deeper GNNs by improving gradient flow and mitigating performance degradation.Source: D. GNN Architecture Considerations · p. 2

E. Graph-Pair Classification Model for ENC Change Classification

A Siamese-style model independently encodes the old and new ENC graphs with shared GNN weights, then compares their embeddings alongside global change metadata to predict change criticality.

E. Graph-Pair Classification Model for ENC Change Classification

  • The Siamese-style model sends each old or new ENC graph through the same GNN encoder to produce graph-level embeddings.The evaluated encoders are GCN, GAT, GraphSAGE, and TransformerConv; global change metadata is mapped to learnable embeddings.Source: E. Graph-Pair Classification Model for… · pp. 2–3
  • Subtracting the old-graph embedding from the new-graph embedding forms a difference representation, with zero-padding used for mismatched node counts.Source: E. Graph-Pair Classification Model for… · p. 3
  • The model concatenates the graph difference with global feature embeddings and uses a fully connected network to predict critical versus non-critical change.Shared GNN weights provide a consistent encoding for comparing old and new graph structures across architecture, depth, and residual configurations.Source: E. Graph-Pair Classification Model for… · p. 3

III. EXPERIMENTS

Experiments compare GNN architectures on a consistent graph dataset of expert-reviewed ENC changes, using stratified five-fold cross-validation and reporting multiple performance metrics.

III. EXPERIMENTS

  • All architectures use the same graph dataset, node features, edge structure, and edge types so performance differences reflect architectural variations rather than input representation.Source: §III · p. 3
  • The dataset contains 9,281 expert-reviewed ENC changes: 6,069 were labeled critical and 3,212 non-critical.Source: §III · p. 3
  • Five-fold stratified cross-validation uses 80% of records for training and 20% for validation in each fold, preserving review-label distributions.Source: §III · p. 3
  • The study reports mean and standard deviation across folds for macro F1 score, precision, recall, and overall accuracy.Models train for 100 epochs with Adam at a fixed learning rate of 1e−3 and weighted binary cross-entropy to address class imbalance.Source: §III · p. 3

A. Ablations

The ablation evaluates GNN architecture, depth, and residual connectivity in a grid designed to separate graph-propagation effects from depth effects.

A. Ablations

  • The ablation compares GCN, GAT, TransformerConv (GT), and GraphSAGE at depths of 3, 4, and 5 layers.Source: A. Ablations · p. 3
  • Each architecture-depth variant is trained both with and without layer-wise residual connections to probe over-smoothing and vanishing gradients in deeper networks.Source: A. Ablations · p. 3
  • The evaluation grid is designed to isolate the contribution of graph propagation mechanism from network depth.Source: A. Ablations · p. 3

B. Experimental Results

GraphSAGE leads the tested models on accuracy and macro-F1, while residual connections have architecture-dependent effects. The discussion links these findings to the role of graph topology and neighborhood aggregation in ENC change classification.

B. Experimental Results

  • GraphSAGE reaches accuracy 0.8766 ± 0.0050 and macro-F1 0.8632 ± 0.0046, exceeding the MLP baseline and all other GNN variants.The MLP baseline reports accuracy 0.8080 ± 0.0055 and macro-F1 0.8483 ± 0.0096; GraphSAGE remains stable across depths.Source: B. Experimental Results · p. 3
  • For GCN, residuals raise 3-layer macro-F1 from 0.8031 to 0.8354 and yield up to 0.8421 at 5 layers, alleviating depth-related degradation.Source: B. Experimental Results · p. 3
  • GAT gains at 3 layers (macro-F1 0.8059 to 0.8269) but is mixed deeper; TransformerConv gains are marginal, while GraphSAGE can decline.For GraphSAGE, 3-layer macro-F1 changes from 0.8609 to 0.8461 with residuals.Source: B. Experimental Results · p. 3

IV. DISCUSSION

  • The authors attribute GraphSAGE’s depth stability to concatenation-based updates preserving node identity; GCN and GAT residual gains suggest signal degradation in this graph topology.They describe residual pathways as redundant or potentially interfering for GraphSAGE, whose native aggregation may mitigate vanishing gradients in this context.Source: §IV · p. 3

V. CONCLUSION

The paper frames ENC change criticality as graph-pair classification and finds graph structure valuable for the task. GraphSAGE leads the tested architectures, and the authors identify broader benchmarking as future work.

  • The approach represents old and updated ENC datasets as graphs, with spatial objects as nodes and semantic and spatial relationships as edges, for graph-pair classification.Source: §V · p. 3
  • Among tested architectures, GraphSAGE consistently achieves the highest accuracy and macro-F1, with its inductive neighborhood aggregation aligning with ENC spatial semantics.Source: §V · p. 4
  • Future work will expand benchmarking across alternative graph constructions, node feature choices, and GNN architectures.Source: §V · p. 4
Loading 2609.02996v1…