Detecting Unexpected Marine Species with Underwater Cameras and Deep Learning

Integrating machine learning with audio and video monitoring for automating marine ecosystem monitoring is in its early stages but advancing rapidly. Typically, these models exhibit robust performance on marine classes and environments on which they have been trained; however, they often misclassify previously unseen marine classes with high confidence, erroneously assigning them to known categories. This study addresses the challenge of detecting previously unseen marine categories in underwater video data, while accurately classifying the examples of seen classes. We propose a system that leverages the features extracted from machine learning classifiers, and unseen marine classes from the publicly available OzFish dataset to accurately identify unobserved species in a low-lit, low-resolution underwater video dataset obtained by the Department of Fisheries and Oceans, Canada. We report accuracy scores of 83.0% and 85.2% for ViT and ResNet18 architectures, respectively, on an extensive test set of 46,641 images with both seen and unseen classes. This methodology marks a significant advancement toward the development of automated, scalable marine monitoring systems capable of adapting to varying species distributions and environmental changes.