The fishing classifier model is applied to the track of every vessel with an AIS vessel class of fishing or unknown. Portions of the track where the activity is determined by the model to be fishing, are coloured pink. Start and stop times for each continuous fishing event are listed in the track history of fishing vessels.

Our fishing classifier is based on a version of Global Fishing Watch’s (GFW) model described in Kroodsma et. al. 2018¹.

The model uses a machine learning algorithm known as a convolutional neural network to classify a time series of movement related features for each vessel. The features include information such as vessel speed and course, local time and month, distance from the nearest shore, and time between AIS messages.

When developing the model, GFW worked with fishing experts to classify fishing activity². They evaluated 247,000 hours of AIS tracks from 624 vessels, capturing a range of fishing methods. The model was trained to distinguish between fishing and non-fishing activity using 503 of these vessels’ tracks. The tracks from the remaining 121 vessels were used to evaluate the model’s accuracy. Regardless of the type of fishing being undertaken (e.g. longlining, trawling, purse seine), the model was able to correctly classify a vessel as fishing or not fishing more than 90% of the time.

The Starboard fishing classifier has an additional post-processing step—for any two hour period where fishing is detected we classify the entire two hour period as fishing. This is done to merge any short fishing events that occur near each other into one fishing event, providing a more realistic representation of fishing behaviour. However, it will also have the effect of expanding the fishing event past the start and end times.

The fishing classifier is processed hourly and potential fishing events that have occurred as recently as four and a half hours ago can be classified.