This post is only a short example, be creative in your post and share images, evaluations, ideas, and more ✨

OceanScout – Ship Detection with TerraMind TiM

Benjamin B. & Team OceanScout
contact → bluesky.demo@placeholder . org (placeholder)

Why it matters

Illegal fishing, unchecked cargo routes and pollution make the ocean one of the most data-hungry frontiers. 75 % of global industrial fishing vessels and 25 % of transport/energy vessels sail untracked. Satellite imagery promises a synoptic view, yet cloud cover and the sheer size of the sea still hide a surprising share of vessels.

How we used TerraMind

Our detector ingests nothing but Sentinel-2 optical scenes. Inside TerraMind uses Thinking-in-Modalities (TiM) to generate a pseudo-SAR representation that mimics the bright back-scatter signature of metal hulls and a Land-Use/Land-Cover (LULC) map whose “water” class acts as a dynamic mask, steering attention away from land. All modalities are input to the fine-tuned encoder, which detects more ships with the additional TiM modalities. Additionally, the generated LULC map is quite accurate and can be used for postprocessing steps like filtering out predictions on land, in case the model gets confused.

A quick tour of the demo

TerraMind spots individual vessels in New York Harbor. Green highlights ships detected by the base model, blue the ones by the TiM model.

Next, we aggregate three months of detections into a heat map: Warmer cells reveal the busiest lanes and potential emission hotspots.

The model can also be used in areas like the Channel Islands marine-protected area. We detect repeated vessel presence deep inside the reserve.

What we learned so far

TerraMind’s built-in generation capabilities are more than “side outputs”; they can be directly used for processing when speed matters.
The LULC masking feels almost like plug-and-play—no extra labelling round was required to tame false positives on land.
A single foundation model keeps the stack slim: no separate SAR network, no ensemble, fewer moving parts to break.

Try it yourself

Code & demo space – https:// github.com/placeholder-org/blueseas-tim
Walk-through blog – https:// medium.com/@ocean-ai/placeholder-link
(links are placeholders)

Where we’re heading next

We are experimenting with an improved model that sketches oil-spill films around detected ships, to track pollution by vessels. On the model side, we look into building a truly resilient multimodal model that can process both S-1 and S-2 and automatically generates the missing one using TiM.