Automated zooplankton detection from in situ imagery for forward scattering predictions

Abstract

Ocean midwaters—the vast region between the sunlit surface layers and seafloor—comprise the largest habitat on Earth but are among the least understood marine environments. This project aims to combine concurrently collected imaging and acoustic measurements in the epi and mesopelagic environment to interpret zooplankton scattering from mixed assemblages and determine in situ zooplankton distributions within their local environments. We have trained a machine learning model for the automated detection of 13 zooplankton functional groups from a 1000m rated towed shadowgraph imaging system. The zooplankton detection model currently achieves 80% F1 scores on our validation image set and was trained using adversarial methods. We have derived biometric measurements from the zooplankton image data necessary to generate forward scattering predictions. We compare the distributions of scattering layers detected acoustically with zooplankton distributions from the imagery. We will compare forward scattering predictions derived from sparse geometric representations of the zooplankton and full 3D model volumes using the distance transform of the zooplankton images. This work will further enable the use of optical techniques for midwater surveys and the interpretation of acoustic scattering returns from mixed zooplankton populations.