A Fishing Event is generated when fishing behavior is detected from the AIS signal by a Machine Learning model.

Overview

The machine learning model will create a fishing event for each unique fishing instance it recognizes. 

The model can detect several types of fishing, including long lining, purse seining, trawling and squid jigging.

Machine Learning Model

To create the fishing model, AI engineers at Ai2 started by creating a tool to label vessel track data (AIS).

In the images below, you can see how experts labeled parts of a vessels tracks. This labeled part of the track is the data used to teach the model how to recognize fishing.

The team labeled over 10,000 months worth of AIS vessel track data to create the model.

Event Details

Standard information

Vessels' information - name, MMSI, Type based on AIS data, IMO, Length based on AIS data

Date and time of the event

Location (lat/lon)

Event history - number of times Skylight has detected this behavior from this vessel

A duration is not specified. Analysts may analyze the tracks to more closely estimate the duration of fishing effort.

Latency

The latency (delay) for fishing events can be as little as 15-20 minutes depending on the AIS signal and time required to analyze the vessel behavior to label it as Fishing.

Performance

The last evaluation was run on October 2024 on a sample of 497 Fishing Events. 

The precision for detecting fishing on AIS was at 71%. If we break down the ability to detect fishing activity by the vessel type of the vessel, the precision is 90% for vessels known to be a fishing vessel type, and 50% for vessels whose type is unknown.