Inspiration

The four of us came together for a school ocean fundraiser, untied by a shared desire to protect our oceans. We wanted to do more than address the damage we could see. Some of the greatest threats happen far beyond our shores, where monitoring is difficult. We saw an opportunity to use machine learning and satellite technology to help uncover unusual marine time activity and combat threats to the communities that depend on healthy and safe oceans.

What it does

Ghost Fleet combines an interactive globe, historical activity heatmaps, vessel identity searches, and risk screening. Users can select an area, review associated vessels, and investigate warning signs. Our machine learning model also forecasts historical fishing activity.

How we built it

We built the frontend with React and TypeScript and the backend with Python and FastAPI. We connected Global Fishing Watch APIs for maritime activity, vessel information, and risk evidence. We also integrated a trained decision-tree model to forecast fishing activity from historical patterns.

Challenges we ran into

We initially built a LiDAR-based ship-detection model, but realized it did not integrate properly with the rest of our system. Abandoning that approach cost us valuable development time and forced us to rethink our solution.

We also faced API limitations, incompatible model data, slow vessel screening, and challenges combining teammates’ work. Making sure missing or conflicting information did not produce misleading conclusions required careful testing.

One of the funniest moments that was equally frustrating was when specific data related to the Ghana fishing territories and rules were mixed incorrectly within our Main file, causing a lot of time debugging by hand and looking through the files to find reference photos of Ghana's fishing territories. We call it the Ghost of Ghana...

Accomplishments that we're proud of

  • Pivoting from an integration setback to a working maritime investigation tool.
  • Connecting real historical data, vessel reports, and machine learning in one application.
  • Making complex evidence easier to explore and understand.
  • Handling uncertainty without automatically labeling vessels illegal.

What we learned

We learned how to research unfamiliar technologies, read API documentation, and evaluate whether a data source actually fits our goals. We gained experience using GitHub branches, pull requests, and merges to collaborate. We also learned how to prepare training data, train machine learning models, evaluate predictions on unseen data, and connect a saved model to a website.

What's next for Ghost Fleets

We want to improve large-area performance, expand validated data coverage, complete marine habitat reference setup across our team’s machines, and evaluate our investigation tools against independently verified cases.

Specifically, improving the algorithm of our investigations tab can increase the speed and efficiency of each check, as well as finding a way to lower the overall time checking by changing it from checking sequentially to bilaterally.

Built With

Share this project:

Updates

Submission history