ShipSense

ShipSense has created a novel way of augmenting data with synthetic images, using a fine-tuned type of AI called Stable Diffusion. This enables them to detect and locate ships from satellite imagery more accurately by using AI-generated images. They also built a dashboard for stakeholders to monitor overfishing, and identified hotspots of dark vessel activity by digging into 55,000+ AIS radar records.

It can be difficult to detect overfishing using AI models due to a high class imbalance. There are fewer positive examples of ships on water compared to the vast amounts of negative examples of patches of water without ships. Generative adversarial networks (GANs) have been used to generate synthetic data, but require around 50,000 sample images to train a decent model. ShipSense was able to use Stable Diffusion to generate realistic synthetic images with only 68 original images.