09.02.24 - Curing the Blind Spots in Sea Surveillance with Neural Networks

How neural networks unravel the mysteries of human activities at sea 🌊

In a recent article led by the Global Fishing Watch, researchers sought to create a mapping of large boat traffic, providing a more accurate view of resource exploitation at sea.

Traditionally, such research relied on data derived from the publicly broadcasted locations of ships. However, a considerable amount of vessels do not share their location or navigate through areas where signal strength is uneven, which creates a blind spot in the data.

In order to take those ‘dark fleet’ into account, the researchers used satellite radar imagery, allowing for detections in lightless conditions. With a staggering volume of data (2000 terabytes!), manual inspection was out of the equation, and machine learning techniques had to be used.

Detection of ships employed a conventional algorithm, identifying irregularities in pixel brightness relative to the sea. Estimation of vessel size utilized a convolutional neural network trained on ground truth data from vessels actively broadcasting their position. Subsequently, another neural network categorized the ships’ purpose, distinguishing between fishing and non-fishing activities. This was achieved by combining the boat length with 11 environmental attributes dependent on the vessel’s location, such as distance from port or average surface temperature.

All these tools were combined to obtain a mapping of boats traffic, whether their locations were broadcasted or not. The results are astonishing: around 75% of global industrial fishing vessels are absent from public tracking, which showcases the power of using deep learning methods to overcome blind spots in available data.

Hashtag#fishing Hashtag#ocean

Image source: Global Fishing Watch