Charting new waters: Skylight’s work to revolutionize maritime data annotation

In the world of machine learning, the spotlight often shines on the outcomes of models that seem to read minds and predict futures. Yet, behind these models lies an unsung hero: the meticulous process of crafting the datasets that fuel these innovations. A large, high-quality dataset is a prerequisite for any high-performance machine learning-based model. However, creating these datasets is typically labor-intensive, time-consuming, expensive, and challenging without robust tools and resources.

Now, to identify vessel behavior, the task becomes even more intricate. Imagine a world where machines understand the nuances of maritime behavior, from the strategic maneuvers of fishing vessels to clandestine meetings and transshipments. This vision begins with data annotation - a critical and often underemphasized step where subject matter experts teach machines to do what comes naturally to humans: learning by example to recognize patterns and make predictions.

At Skylight, an AI-powered maritime monitoring platform from the Allen Institute for AI (Ai2), we confronted these challenges head-on while trying to classify the complexities of vessel behavior in AIS. Previously, we depended on having multiple annotators work with ad-hoc tooling to construct training datasets. This allowed us to produce models quickly but made annotating efficiently and iterating on existing models challenging. Last year, we embarked on a journey to develop a first-of-its-kind tool tailored to make annotations easy and keep these models up-to-date more quickly and efficiently.

"This platform doesn't just make our work faster; it makes it better."  - Rachel Ratner, Senior Software Engineer for Skylight

This in-house annotation platform, designed for speed and simplicity, transforms how we harness the power of Big Data. Specifically, the way we leverage the 150 million Automatic Identification System (AIS) messages Skylight receives daily. It represents a pioneering effort in the field, aiming to streamline the annotation process required to empower subject matter experts with the ability to visualize, evaluate, and label AIS positional data very efficiently.

"It's about turning a complex task into something intuitive," said Rachel Ratner, Senior Software Engineer. "This platform doesn't just make our work faster; it makes it better."

Henry Herzog, Software Engineer, emphasized the broader impact: "By scaling our annotation efforts, we're not just refining our models. We're setting a new standard for the entire field."

In just a few months, Skylight has amassed over 2 million annotated AIS positions with only a handful of annotators, thanks to this tool. Beyond its initial achievements, the platform's ongoing contribution lies in efficiently assembling indispensable datasets for constructing sturdy models. Equally as important, the platform has opened up new avenues for Skylight to tap into the wealth of firsthand experience and domain-specific knowledge of subject matter experts. In our quest to revolutionize maritime vessel data annotation, Skylight's new innovative tool not only helps accelerate AI's understanding of vessel behavior but also strives to usher in a new era of ocean protection through advanced machine learning capabilities.