Deep Learning in the Ocean’s Depths

A sensor-packed biologger provides insights about underwater animal behavior with the help of deep learning methods.

Nick Bild
21 days agoMachine Learning & AI
Video footage from tag (📷: L. Brewster et al.)

Long-term studies of underwater animals present many challenges to researchers, due to the difficulty of physically observing these creatures for any meaningful duration of time in their natural habitats. Fortunately, with the reduction in cost, size, and power consumption of sensors and processing units in recent years, many insights can be gleaned through the use of autonomous electronic devices. By tagging animals with such devices, it is possible to gain an understanding of their biomechanics, activity patterns, energy expenditure, diving and mating behaviors, among other physiological and behavioral data points.

Despite their massive size, growing up to eight feet long and sometimes weighing in at over eight hundred pounds, and the fact that they are commonly found in waters off the coast of Florida, the Atlantic goliath grouper’s behaviors have been poorly characterized. A team of biologists and engineers at Florida Atlantic University built a multi-sensor tag that can be attached to these giant fish that will passively collect data to help scientists better understand their normal behaviors.

The tags include a three axis accelerometer, gyroscope, and magnetometer, in addition to temperature, pressure, and light sensors. A video camera and hydrophone (to record underwater sounds) were also included in the tag. Sensor data is continuously logged for a period of three days, after which time the tag is designed to fall off and float to the surface of the water. From there, it broadcasts a signal via onboard VHF and satellite transmitters such that it can be located and collected by the research team.

To make sense of this mountain of sensor data, the team employed a number of machine learning algorithms, both conventional (random forest, support vector machine) and deep learning (convolutional neural network). Classifiers were developed by initially training models with sensor data that had been manually paired with a predefined set of behaviors by using video data for confirmation. Once trained, the models were able to recognize nine different behaviors with a high degree of accuracy, taking the burden of activity classification off of humans, while simultaneously greatly improving the speed at which insights can be gained. In most cases, the deep learning approaches performed with a higher degree of accuracy.

The team noted that specific patterns of behavior, that had previously been noted through direct observation of goliath grouper, were detected in the data retrieved from the tags. This served as an important validation of the method, indicating that behaviors are being faithfully recorded. In this initial small study, a great deal of behavioral variety was observed, indicating that further studies of larger populations will be needed to fully understand what makes the fish tick.

While goliath grouper were the target of this particular study, the same tags and methods could be employed to study a much wider range of underwater animals around the world. This work has shined a light on the potential that deep learning has to help researchers better understand the depths of the oceans.

Nick Bild
R&D, creativity, and building the next big thing you never knew you wanted are my specialties.
Latest articles
Sponsored articles
Related articles
Latest articles
Read more
Related articles