A preprint reports that a machine-learning approach classified changes in electronic navigational charts as critical or non-critical with 90% and 94% accuracy across two operational datasets. The paper describes this as a reported 5-7% improvement over default models that did not include spatial context or attribute embeddings.
That improvement range needs caution: the supplied analysis does not specify whether it refers to a relative percentage or a percentage-point difference. The result is a benchmark classification finding, not evidence that the system improves maritime safety or reduces review workload.
A richer picture of each chart change
The study asks whether machine learning can automate a binary decision: identifying an electronic navigational chart change as critical or non-critical using structured information about the change.
The proposed representation encoded an object’s class, its change type and relationships with other chart objects. It also added neighboring objects within a 500-metre centroid buffer, while object attributes were converted to strings and embedded with DistilBERT.
Two datasets, two kinds of labels
The evaluation used approximately 140,000 changes in a Critical/Non-Critical dataset labeled by a programmed rule-based system, and approximately 9,000 changes in an Eyes-On dataset reviewed and labeled by human analysts. Together, the data came from 356 ENC cells and 1,308 old-to-new chart pairs covering the United States, Canada, Australia, New Zealand, Norway, the Netherlands and Germany.
Models were assessed with overall accuracy and macro F1-score using five-fold cross-validation. Each dataset was divided into five train-validation splits, and the comparison included XGBoost, logistic regression, random forest, multilayer perceptron and ResNet models.
The strongest signal depended on the dataset
In the detailed XGBoost comparison, the reported Eyes-On accuracy/F1 pairs were 89.2%/87.9% for the default model and 90.3%/89.1% for the tuned model. On the Critical/Non-Critical dataset, the corresponding pairs were 93.5%/93.5% and 94.2%/94.2%.
Feature tests showed a different pattern across the datasets. On Eyes-On, spatial context alone produced 90.1% accuracy and an 88.9% F1-score, ahead of attributes alone at 85.7%/83.7% and the combined encoding at 89.2%/87.9%. On Critical/Non-Critical, the combined encoding led at 93.5%/93.5%, compared with 92.2%/92.2% for attributes alone and 90.0%/90.0% for spatial context.
Across the complete encoding, the trained baseline model families all exceeded 85% accuracy on both datasets and performed better than the two naive rule-based classifiers.
A benchmark, not a deployment result
One figure requires particular care. The Rules system achieved 100% accuracy on the Critical/Non-Critical dataset because it generated that dataset’s ground-truth labels. The paper treats that result as a reference, not as independent evidence of predictive performance.
The authors interpret the findings as support for machine learning in operational ENC maintenance and maritime-safety pipelines, with simple location and spatial aggregation as a basis for more advanced spatial representation learning. The supplied analysis describes the evidence as a cross-validated classification benchmark, not a causal demonstration of safety or efficiency gains.
The analysis leaves open whether the findings would generalize to other ENC products or geographies, or whether a prospective system would reduce review burden or improve maritime-safety outcomes.
Paper data and sources
Original title: Electronic Navigational Chart Change Classification
Authors: Jacob Arndt, Abhishek Potnis, Alexandre Sorokine
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text