The project
The goal was to turn a spectacular idea into a demonstrable system: start from satellite imagery, look for visual signatures of vessels, then expose a trained detector in a usable public interface.
The important part is not only the model. The value is the full chain: image acquisition, area selection, tiling, annotation, training, checkpoint comparison, inference and deployment.
Technical angle
The public model is a YOLOv8s detector fine-tuned on RGB satellite imagery. The public README documents the checkpoint train-20260417T124314Z-fad9d3ed_best.pt, run infer-b88a2887, training name super-visible-y8s-newlabels-focuslite-e45, dataset sentinel-2-rgb, and local index mAP50 0.7912.
What it proves
This case study shows the ability to handle difficult data, train a specialized computer-vision model and ship a real demo around it.