To keep up and identify precisely where and when a vessel is engaging in potentially illegal or non-compliant behaviors, fisheries managers need to do more with less. To help, Skylight has developed a machine learning algorithm designed to automatically detect where a vessel has likely been fishing.
By many measures our seas are running out of fish.
The grim state of the world’s major marine fish stocks are such that the Food and Agriculture Organization of the United Nations (FAO) classifies 93 percent as fully fished or overfished, with more than a third taken at unsustainable levels. While poor or ineffective fisheries management is partially at fault, one of the biggest culprits is illegal, unreported, and unregulated (IUU) fishing. This activity undermines the sustainable management of these precious resources which nearly half of the world’s population, particularly those most vulnerable in small island developing states, rely on as a valuable protein source. But, there are reasons to be optimistic. There is growing evidence that when fisheries are properly managed, stocks can rebound and serve as the catalyst for restoring others around them.
With hundreds of thousands of vessels and boats plying our planet’s seas every day, the role technology must play today and in the future is central to restoring the vitality of our ocean. In the most egregious cases, distant water fleets of more than 3,000 factory trawlers stretch illegal drift nets for miles strip mining our ocean. To keep up and identify precisely where and when a vessel is engaging in potentially illegal or non-compliant behaviors, fisheries managers need to do more with less. To help, Skylight has developed a machine learning algorithm designed to automatically detect where a vessel has likely been fishing.
An example of Skylight's fishing model in the platform. Skylight’s fishing model detects fishing activity by analyzing data from vessels transmitting on the Automatic Identification System (AIS) in near real-time. The model was trained with annotated track examples from subject matter experts as well as fisheries observer data provided by the National Oceanic and Atmospheric Administration’ (NOAA) Alaska Fisheries Science Center. Using anywhere between one to 40 plus hours of past movement data, the model considers how a vessel’s speed is changing, the way it is turning and maneuvering and the depth of the ocean to evaluate whether its current behavior is fishing. NOAA’s data in particular has been a crucial source of ground truth for the model and greatly improved the way Skylight can help authorities and organizations detect potential illegal or non-compliant fishing activity. As a result, the platform is able to identify many types of fishing behaviors including trawling, seining, longlining and squid jigging regardless of whether the vessel is transmitting itself as a fishing vessel.
"Historically, NOAA has lacked the capacity to develop models as complicated or with as large of datasets as those implemented by Skylight,” said Jordan Watson, Mathematical Statistician for NOAA’s Alaska Fisheries Science Center. “This is an example of the value and importance of private-public partnerships for supporting a sustainable blue economy while building our capacity to meet our mission."
Pinpointing vessels that are likely fishing with a machine learning algorithm comes with its own challenges. One of the biggest challenges Skylight’s model faces is where fishing events are harder to distinguish. These places are where navigation is restrained and results in heavy vessel traffic or where vessels have similar movements from port operations or miscellaneous activities such as whale watching and law enforcement. Skylight’s model is not a guarantee that fishing happened. It is also a model based on AIS which many fishing vessels do not carry. But despite this, it does serve an important role of flagging a massive amount of potential fishing activity automatically, under 24 hours, and without manual analysis – something maritime and fisheries analysts didn’t have before.
Today, Skylight’s fishing model is being used across the globe by governments, organizations, and the international community. For example in May this year, the Madagascar Centre de Surveillance des Pêches (Fisheries Surveillance Center) issued a warning letter to the ZOVA 5 for illegally trawling within their two nautical mile zone because of a Fishing Event generated by Skylight. ZOVA 5 was warned that if it is caught fishing inside the boundary again, it will be sanctioned. Going forward, the Skylight team continues to improve the performance of the model to better serve its users fighting illegal fishing.
“We know that a lot of fishing is not visible on AIS, so we are spending a lot of time this year and in the coming year investing in a larger variety of data sources that can pick up more fishing vessels,” said Namrata Kolla, Product Manager for Skylight. “While most of these sources like satellite radar and optical imagery have been around for many years, they have not always been available to enforcement agencies, particularly in developing states, at no cost, timely, and easy-to-act-on manner. That is Skylight’s main focus when we develop all of our events, including the Fishing events.”
But it can’t be overstated the role NOAA’s data played in developing Skylight’s fishing model. A model is only as good as its data. And NOAA’s contributions gave the Skylight team a high quality data set that required substantially less effort to develop a model that could provide actionable insights. It also set a precedent that high quality data sets such as NOAA’s could be used to train the platform’s machine learning models.
If you have fisheries observer data or other data that would help Skylight detect important vessel activity for your agency or partner organizations, please email us at support@skylight.global .