Unsupervised Port Berth Identification from Automatic Identification System Data

TL;DR

The paper addresses incomplete port berth documentation by proposing an unsupervised, data-driven method to localize berthing sites from freely available AIS data. It combines DBSCAN-based filtering, spatial data augmentation guided by vessel dimensions and heading, geohash encoding, Gaussian Mixture Models, and MDL-based model selection, with hyperparameters tuned via KL-divergence across data splits and evaluated using Bhattacharyya distance with Monte Carlo integration. The approach yields substantial improvements over prior work, producing precise berth boundaries across ports of varying sizes and enabling reliable berth localization for AIS geofencing and port optimization. The method demonstrates strong generalizability, robustness to data splits, and qualitative agreement with satellite imagery and existing berth labels, contributing a scalable tool for open-data berth localization and maritime analytics.

Abstract

Port berthing sites are regions of high interest for monitoring and optimizing port operations. Data sourced from the Automatic Identification System (AIS) can be superimposed on berths enabling their real-time monitoring and revealing long-term utilization patterns. Ultimately, insights from multiple berths can uncover bottlenecks, and lead to the optimization of the underlying supply chain of the port and beyond. However, publicly available documentation of port berths, even when available, is frequently incomplete - e.g. there may be missing berths or inaccuracies such as incorrect boundary boxes - necessitating a more robust, data-driven approach to port berth localization. In this context, we propose an unsupervised spatial modeling method that leverages AIS data clustering and hyperparameter optimization to identify berthing sites. Trained on one month of freely available AIS data and evaluated across ports of varying sizes, our models significantly outperform competing methods, achieving a mean Bhattacharyya distance of 0.85 when comparing Gaussian Mixture Models (GMMs) trained on separate data splits, compared to 13.56 for the best existing method. Qualitative comparison with satellite images and existing berth labels further supports the superiority of our method, revealing more precise berth boundaries and improved spatial resolution across diverse port environments.

Figures (14)

• Figure 1: An example of a port berth in the port of Limassol used for loading and unloading container ships. Source: GoogleMapsImage
• Figure 2: Visualization of the AIS messages (black dots) collected within a month for the port of Cape Town and its surroundings.
• Figure 3: Visualization of the AIS messages (blue dots) collected within a month for the port of Limassol (red pentagon).
• Figure 4: An example of ShipNext labels for the port of Limassol (left) and the port of Auckland (right). The red rotated rectangles are the port berth labels, and the orange areas are the terminal labels.
• Figure 5: Illustration of DBSCAN algorithm and GMM applied to AIS Data in the port of Cape Town. Left --- The red points indicate AIS data points that are classified as part of clusters, having passed the DBSCAN filtering, while the black points represent those identified as outliers by the algorithm. Right --- The trained GMM with optimal hyperparameters where each Gaussian component represents a potential berth (each Gaussian component is truncated at 3 standard deviations for visualization purposes).