...

In October, 2022 a new machine learning model for fishing activity and dark rendezvous was released in Skylight. This single model classifies AIS track segments into one of three exclusive classes: fishing activity, dark rendezvous, or neither. This is a change from the previous binary models that predicted fishing and dark rendezvous separately. As a result of the new three-class model, there are no longer instances where track segments are identified as both dark rendezvous and fishing. So instances like the below screen shot where the Xidi PROFESSOR LOGACHEV is shown as being involved in Fishing activity and a Dark Rendezvous simultaneously can no longer occur.

...

In addition to preventing overlapping fishing and dark rendezvous events, this new model improves our the confidence in dark rendezvous events since it the model can detect the difference between fishing and dark rendezvous. The previous binary machine learning model for dark rendezvous had not been trained with fishing data, so it would mistakenly classify fishing activity as dark rendezvous. The result is fewer dark activity events are predicted globally, but they are more likely to have occurred than dark rendezvous predicted with the previous model.

Model Training:

• How it was trained: 

  • The model was trained via supervised learning on both fishing and dark rendezvous labels in order to improve model accuracy over the previous binary models.

  • The model was trained on a combination of expert annotated data and ground-truthed data from fisheries observers. 

  • The model was trained on many types of fishing behavior, including trawling, seining, long-lining, and squid jigging. Notably not on this list, though they may be picked up for their similarities with other fishing methods: gill netting, spear fishing, pole & line. 

  • See Appendix B for additional information.

• The main characteristics of the vessel movement the model considers are:

  • How the vessel’s speed is changing

  • How the vessel is turning and maneuvering 

  • Depth of the ocean / bathymetry 

• The model needs a minimum amount of historical movement in order to determine the vessel is fishing right now. Depending on the movement of a vessel at that particular time, this could mean anywhere from 1 to 40+ hours of past movement to evaluate whether its current behavior is fishing (see Appendix A).

• The model does not take vessel type into account specifically so that vessels that display fishing activity though they are not transmitting themselves as fishing vessels can be found.

• The event icon, whether fishing or dark rendezvous, is not placed on every event, just on the first behavior detected in that 24 hr period for that vessel. 

...

• This is a model based on AIS positions alone - it will never be a guarantee that fishing or a dark rendezvous happened. Consider this a flag for potential activity.

• Analysts need to continue using vessel event tracks and other tools and data to make judgement calls. 

• As always, we would love your feedback and especially contact us if you have information that ground truths whether the fishing or dark rendezvous event actually happened or not. 

• A common place where there will still be some potential false positives / where fishing events are harder to distinguish are places where navigation is restrained because of the shape of the coastline (e.g. the Inside Passage in the Pacific Northwest), high vessel traffic (e.g. Strait of Malacca), more generally complex navigation (tugs or port operations more generally), and some miscellaneous other activities (dredging, law enforcement, crew or material transfer / rendezvous activity, whale watching, etc). Addressing these false positive generating behaviors is a focus of next steps in development of this feature.

• Determining whether a vessel is actually a buoy uses a separate model from this model. 

...

The model architecture used for fishing classification is a Long-Short Term Memory network (LSTM), a type of model designed for operating on time series data such as this. The model outputs binary fishing/not-fishing classifications at a segment level, though in future releases we are discussing a model that classifies fishing gear used.

Appendix B: Labeled data 

...

One source of training data comes from fisheries observers. For this data source, we matched tracks from our system with reported fishing activity from observer records to automatically label fishing behavior sections in tracks. These tracks were then checked by an expert analyst to verify this process, and the labeled tracks were used in model training. Usage of this observer data substantially improved model performance (+50%) specifically on vessels fishing in the region that provided the observer data (as well as generalizable global improvements). So if you have fisheries observer data and would like to help us improve performance on your area of responsibility, definitely get in touch.

Expert Annotations

Our other major (and currently largest) source of model training data is AIS tracks (raw position points) hand-labeled by expert analysts. 

...