Machine Learning Model: Sentinel-1 Imagery and Vessel Detections in Sentinel-1

Sentinel-1 is a satellite managed by the European Space Agency. Read about the use of imagery from Sentinel-1 in Skylight here.

Model

The Sentinel-1 Model is based upon the classic object detection model, Faster R-CNN, which is a two stage detector made up of a Region Proposal Network (RPN) and a classification head. The RPN’s job is to generate proposals, where there is likely an object of interest, and the classification head grabs the most confident of these proposals and predicts scores (such as is_vessel) for each proposal.

The model detects vessels and predicts vessel attributes from Sentinel-1 SAR images. In particular, it uses the dual polarization mode (VV + VH) of the Interferometric Wide swath (IW) acquisition mode of Sentinel-11, and produces point detections of vessels, cropped outputs surrounding those detections, and attributes associated with the detected vessel; currently estimated length is displayed in Skylight).

Training Data

The current version of the model has been trained on Sentinel-1 scenes from several geographic areas, mostly near coast, that were annotated by hand by subject matter experts. A total of 55,499 point labels were used. These areas are distributed globally and shown in the image below.

The point labels comprising the training set used to train the detection model

The model predicts the positions of vessels, and assigns each prediction a confidence score. Detections with a confidence score > .9 are displayed in Skylight. The other attribute that the model predicts is vessel length.

To distinguish moving ships from static objects like islands, platforms. and other static non-vessel structures, the model is trained on data that includes overlapping images captured at different times. The model compares these images and learns to disregard static images.

Validation Data

The validation set is data held back from training specifically so it can be used to validate the model. For the Skylight Sentinel-1 model a total of 6156 point labels were used for validation. The global distribution for the validation set is roughly the same as the training set.

The point labels comprising the validation set used to evaluate the detection model

 

Open Source

In support of Skylight’s mission to deliver meaningful impact towards reducing IUU fishing, we have made the Sentinel-1 imagery detection model available via open source here: https://github.com/allenai/vessel-detection-sentinels/