Building the Foundation for Maritime AI: SAR-Based Vessel Detection

About the customer

A global leader in the Earth and space observation industry, utilizing AI and satellite data to deliver strategic insights for businesses and government sectors worldwide. Their AI-powered platform transforms complex data from Synthetic Aperture Radar (SAR) satellites, Automatic Identification System (AIS) signals, and proprietary maritime sources into real-time situational awareness. Their solutions enable critical outcomes in vessel tracking, offshore asset monitoring, illegal fishing detection, port surveillance, and maritime domain awareness.

Goals

The objective is to create a high-quality, large-scale ground truth dataset using thousands of Synthetic Aperture Radar (SAR) images. This dataset will aid in developing a high-performance, multiclass vessel object detection model. Our goal is to facilitate the creation and implementation of geospatial AI models by ensuring rapid turnaround, strict quality control, and cost-effective labeling. This approach aims to minimize the time needed for production while enhancing model accuracy at scale.

Challenge

SAR-based vessel detection poses unique challenges due to the characteristics of radar-based imaging. This includes issues like speckle noise, variations in reflectivity signatures, and lower visual clarity than electro-optical (EO) imagery. To accurately identify and differentiate between vessel types, such as tankers, cargo ships, fishing vessels, and patrol boats, it is essential to have a deep understanding of radar signatures and contextual geospatial interpretation. The complexity increases with densely clustered vessels, varying sea conditions, wake patterns, and ambiguities at the land-sea boundary. Additionally, traditional annotation teams often lack the specialized expertise to differentiate vessels from radar clutter or interpret unclear reflections.

Solution

Labelbees has developed a scalable, high-precision data collection pipeline designed explicitly for Synthetic Aperture Radar (SAR) data. By utilizing in-house SAR experts, proprietary workflows, and ontology research tailored for radar imagery, Labelbees has successfully delivered thousands of annotated vessel instances with pixel-level accuracy. Our subject matter experts worked closely with data scientists to ensure the precise interpretation of radar-specific patterns and vessel characteristics.

Our extensive global vessel ontology, which is based on over 100,000 real-world vessel profiles, supports the nuances of SAR-specific labeling. This includes intensity-based bounding, wake interpretation, and the orientation of objects under radar tilt. To manage project complexities, our in-house labeling team utilized subject matter experts, machine learning data specialists, and data scientists to ensure high-quality outcomes throughout the pipeline. By following a rigorous, custom-built data labeling workflow, Labelbees provided a tailored solution that included:

• SAR-Specific Ontology Adaptation
• Identification and Curation of Edge Cases (e.g., backscatter)
• A Data Solutions Team Led by Domain Experts
• Custom Metadata and Environmental Context Tagging
• Detailed Labeling Guidelines and Documentation
• A Robust Quality Assurance Process
• Feedback-Driven Iteration Cycles
• Cross-Modal Validation
• Dataset Insights and Reporting

Result

Labelbees facilitated the rapid development of an advanced SAR-based multiclass vessel detection model. Our high-quality ground truth data enabled the client to create robust models with enhanced detection accuracy across various maritime scenarios, ranging from open ocean to congested near-shore areas.

We delivered large batches of SAR annotations with near real-time turnaround, improved annotation precision by 18% compared to previous benchmarks, and reduced the time required for training data preparation. Additionally, we achieved first-pass annotation acceptance with zero rework cycles.

• Accelerated deployment of AI-driven maritime intelligence features.
• Facilitated cross-domain fusion (SAR + AIS + EO) for enhanced analytics.
• Avoided expensive rework and delays, resulting in the achievement of first-pass approval.
• Lowered model retraining costs and operational latency.
• Boosted competitiveness in the defense, logistics, and maritime security sectors.