Summary

In November, 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 occurances like the below screen shot where the PROFESSOR LOGACHEV is shown as being involved in Fishing activity and a Dark Rendezvous at exactly the same time on the same track segment, can no longer occur.

Open

Note, there will be times when the fishing and dark rendezvous events are detected at different times; in these instances both a fishing and dark rendezvous event will be generated.

Open

In addition to preventing exact overlapping fishing and dark rendezvous events, this new model improves the confidence in dark rendezvous events since 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 significantly fewer dark activity events are predicted globally, but there are fewer false positive dark rendezvous events predicted than 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.

  • Training data includes a combination of expert annotated data and for fishing events, ground-truthed data from fisheries observers. 

  • The model was trained on four 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 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. 

“Accuracy”/Performance of Events 

• 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).

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

Fishing events in Skylight v. Global Fishing Watch (GFW)

• GFW also generates fishing events using AIS-based tracks globally, but their model was trained on different data, with different methods, and with different AIS providers. 

  • It is still of most benefit to the user to use both Skylight and GFW when analyzing fishing activities. 

• GFW’s Public Map provides fishing events with a minimum delay of 48 hours (compared to Skylight Fishing events in near real-time ).

• Historical fishing events in Skylight are available back to October 2021. GFW fishing events go back longer. 

Appendix A: Segmentation, features, model, and input duration

Track Segmentation

A key aspect to the fishing training data is track segmentation. The purpose of segmentation is to transform a vessel track from a somewhat randomly sampled collection of AIS position points to a series of sections of internally-similar behavior. The benefits of this are primarily that a track may be represented in terms of operator maneuvering decisions which has clear implications for behavior analysis, as well as a lighter-weight representation of vessels allowing for faster analysis and visualization. The engine behind the segmentation process is a version of the Ramer-Douglas-Peucker algorithm modified to account for speed changes and to adapt parametrically to vessel behavior characteristics. Segmentation effectively capturing fishing behavior (most challengingly, purse seining) was a guiding design principle of the process.

Features & Model Input

The specific model features used are the following measures on segments and their neighbors:

• segment duration in seconds

• segment end to end speed

• segment average travel speed

• local hour (approximate)

• reporting frequency during track segment

• delta course over ground from last segment

• delta speed over ground from last segment

• total distance traveled in segment

• average depth of the segment

• change in depth of the segment start to end

For a twenty-segment window of vessel activity (nineteen + most recent). This can encompass anywhere from approximately one hour to over 40 hours of real time vessel behavior history, depending on how actively the vessel is maneuvering. 

While using twenty segments (as opposed to fewer) for activity recognition has substantial benefits to model performance since it provides the model with a bigger window into vessel operations, one downside is that if a vessel doesn’t have sufficient segment history within three days previous to a particular fishing event, the fishing model cannot generate an output for this event and our system will not be able to flag it. Given the complex maneuvering and thus relatively fine-grained segments associated with fishing behavior, analysis shows that the overwhelming majority of vessel activity falling into this category of not being able to generate an output represents nonfishing activity. However, if you have an example of a fishing event not being generated/vessel not being surfaced that you would expect to see, we would be interested to hear about it. 

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.

Appendix B: Labeled data 

Observer 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).

Expert Annotations

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

Trawlers

Vessel overview and physical characteristics

Trawlers typically target midwater and bottomfish by towing a large, cone-shaped net. Trawl nets can be set and hauled over the side of the boat (known as side trawlers), or over the stern (stern trawlers). When identifying trawlers, look for mechanical trawling winches used to haul the net. 

Behaviors and track analysis

Transit. Trawlers typically transit at relatively high speeds (approximately 7-8 kts) along a relatively consistent heading.  

Fishing. Trawlers will typically fish at speeds between 2-4 kts, though may drift at lower speeds when deploying and recovering nets.

Figure 5. One year trackline of a trawler fishing between Denmark and Sweden, by activity

Figure 6. Twenty hour snapshot of a trawler’s trackline by speed; speed noted in knots

Figure 7. Twenty hour snapshot of a trawler’s trackline by activity; speed noted in knots

Longliners

Vessel overview and physical characteristics

Longliners set free floating (pelagic) or weighted (fixed) lines, sometimes many miles in length with thousands of baited hooks. Lines are set at sea, and recovered after a number of hours, requiring the longliner to return to the same location or find the line floats if it has drifted with the currents. When identifying longliners from images, look for vessels with wenches, weather covers on the stern to protect the crew, and numerous buoys and flags tied along the rails when not fishing. 

Behaviors and track analysis

Transit. Longliners typically transit at relatively high speeds along a relatively consistent heading.  

Fishing. Longliners will typically fish in one area for days or months, resulting in large clusters of AIS positions at varying speeds and headings. The fishing lines are often laid out at higher speeds (7-10kts) and recovered at slower speeds (1-5kts). Lines can drift from a few hours to a few days before being recovered (Figure 10). Revisiting a track in parallel lines is distinctive.

Figure 8. One year trackline of a longliner fishing off the coast of Hawaii, by activity

Figure 9. One year trackline of a longliner fishing off the coast of Hawaii, by speed

Figure 10. Seven day snapshot of a longliner’s trackline by activity; speed noted in knots

Purse Seiners

Vessel overview and physical characteristics

Purse seiners catch fish by encircling them with a long net and drawing (pursing) the bottom closed to capture the fish. The net is paid into the water while the boat travels in a large circle around the fish and retrieved through a hydraulic power block or winch. When identifying purse seiners, look for long clean decks, a boom or pursing davit or crane with the hydraulic power block, and nets stacked on the back. Purse seiners also often have crows nest/lookout or helicopters onboard to help spot desirable schools of fish from a distance. 

Behaviors and track analysis

Transit. Purse seiners typically transit at relatively high speeds (approximately 6-9kts) along a relatively consistent heading. 

Fishing. When fish are detected, the vessel will deploy the net and transit in a circle. Often a skiff (small boat) is used to encircle the fish. Once the net is closed, the vessel will then drift at a low speed while it brings the catch onboard and sorts and processes the fish in a process known as brailing.  If AIS is frequently captured and the purse seiner circles the fish, a clear circular pattern may be observable in the vessel’s trackline (see Figure 12). If AIS is intermittent or a small skiff without AIS is used to encircle the fish, the trackline may not capture the net deployment. Instead, the vessel’s drifting behavior during brailing is likely the best indicator of fishing activity (see Figure 13). 

Figure 11. One year trackline of purse seiner fishing off the coast of Spain, by activity

Figure 12. Three and a half hour snapshot of purse seiner fishing and brailing by activity; speed noted in knots

Figure 13. Two hour snapshot of purse seiner fishing and brailing by activity, speed noted in knots

Squid Jiggers

Vessel overview and physical characteristics

Squid jiggers typically fish at night in continental-shelf waters between 60-120m deep. Overhead lights are used to attract the squid, which are then caught on barbless lures on lines jigged up and down in the water by mechanically-powered jigging machines. Each jigging machine has a roller extending from the side of the boat to allow the line to be jigged up and down without causing abrasion to the line.  

Behaviors and track analysis

Transiting. Squid jiggers typically transit at relatively high speeds along a relatively consistent heading.

Fishing. Squid jiggers typically fish at night at slow, drifting speeds (0-2 kts). Observing the vessel’s speed, depth of the water and proximity to the continental shelf, and local time can be used to indicate fishing activity. 

Transhipping. Because transshipment requires close quarters coordination between two or more vessels, it typically occurs at a constant speed in a straight line for several hours. Speed is usually slower than transiting speeds, but faster than fishing speeds (see Figure 17). 

Figure 14. One year trackline of a squid jigger fishing off the coast of Peru, by speedFigure 15. One year trackline of a squid jigger fishing off the coast of Peru, by activity

Figure 16. Eleven days snapshot of squid jigger fishing, by activity; speed noted in knots

Figure 17. One day snapshot of squid jigger fishing and transshipment, by activity; speed noted in knots