Polyp Activity Estimation and Monitoring for Cold Water Corals with a Deep Learning Approach

Abstract

Fixed underwater observatories (FUOs) equipped with a variety of sensors including cameras, allow long-term monitoring with a high temporal resolution of a limited area of interest. FUOs equipped with HD cameras enable in situ monitoring of biological activity, such as live cold-water corals on a level of detail down to individual polyps. We present a workflow which allows monitoring the activity of cold water coral polyps automatically from photos recorded at the FUO LoVe (Lofoten - Vesterålen). The workflow consists of three steps: First the manual polyp activity-level identification, carried out by three observers on a region of interest in 13 images to generate a gold standard. Second, the training of a convolutional neural network (CNN) on the gold standard to automate the polyp activity classification. Third, the computational activity classification is integrated into an algorithmic estimation of polyp activity in a region of interest. We present results obtained for an image series from April to November 2015 that shows interesting temporal behavior patterns correlating with other posterior measurements.