Development of an image binarization software tool for net occlusion estimations

Abstract

Marine biofouling poses a set of challenges to the salmon aquaculture industry, where an accumulation of biota on pens can lead to significant net mesh occlusion. This can reduce flow rates, risking oxygen depletion. The industry currently manages this challenge through regular cleaning of nets, with sporadic manual visual estimations of net occlusion an important but time-consuming task. This study developed a simple automated desktop application to more regularly and more robustly quantify pen net occlusion caused by biofouling. This software application pre-processes and binarizes images collected from cameras currently used by the industry into water and non-water pixels. The percentage of net occlusion is then calculated from the binary image. Accurate binarization of representative images was achieved by training a deep learning network on images collected in situ. The resulting network attained a validation accuracy of 96.4 % and a mean test accuracy of 93.5 %. From the test images, 98.3 % of pixels annotated as non-water and 88.8 % of pixels annotated as water were correctly classified by the network. This automated tool has the capacity to better inform industry and create a more efficient cleaning framework based on the needs of individual pens, based on data that can be more readily obtained as compared to manual net inspections.