Session: 08-05-01 AI‑Enabled Marine Sensing, Monitoring & Digital Twins

Submission Number: 175455

Aquascan: Graph-Based Learning for Distributed Marine Sensing 

Marine monitoring faces unprecedented challenges as climate change and human activity reshape the oceans. Traditional tracking methods are increasingly unable to address the scale and complexity of modern maritime sensing needs, especially in distributed and partially observable environments.

In this work, we investigate the application of Graph Neural Networks (GNNs) for predicting marine life trajectories, conducting a comparative study against Kalman filters in distributed sensor networks. To support this investigation, we introduce the Aquascan framework, a tool for modeling complex, multi-entity marine behavior with distributed sensing, enabling reproducible benchmarking of learning-based prediction methods. This design allows us to model realistic scenarios of marine observation where direct measurements are sparse, and sensing coverage is irregular.

Our experiments span multiple prediction horizons and reveal a consistent performance gap between the two approaches. GNNs maintain more than 95% Area Under the ROC Curve (AUC) across all horizons, while Kalman filters degrade substantially, from 97% AUC in the short term to 69% AUC over longer horizons. The advantage of GNNs becomes even more pronounced under challenging conditions: at a 5 km sensor spacing, GNNs achieve 92.8% AUC compared to 66.9% for Kalman filters. This improvement stems from the ability of GNNs to exploit network topology, propagate information across multiple sensors, and reason about non-detections when inferring entity presence in coverage gaps.

Our results demonstrate that graph-based approaches provide substantial benefits for distributed marine monitoring, opening new possibilities for tracking wildlife, unauthorized vessels, and drifting objects in sparse sensing environments.

Presenting Author: Abel Dantas FEUP

Presenting Author Biography: Abel Dantas is a PhD candidate in Informatics Engineering at the University of Porto (FEUP), supervised by Professor Carlos Baquero, focusing on distributed systems and Byzantine fault tolerance. He serves as Guest Professor at both IPVC and FEUP, bringing extensive industry experience from leading technical teams in the gaming and blockchain sectors. His research interests span distributed systems, Byzantine fault tolerance, AI/ML, and decentralized networks. He has published work on CRDTs in peer-to-peer VR systems at PaPoC 2025 (ACM) and is currently exploring applications of graph neural networks to marine sensing challenges. As founder and lead researcher at Xarp, he develops AI-driven smart mirror technology for behavioral analytics and HCI applications, working at the intersection of ML and distributed systems through the use of CRDTs and GNNs.

Authors:

Abel Dantas FEUP