Pixel-level quantification of biofouling taxa associated with salmon aquaculture infrastructure using deep learning

Biofouling communities readily develop on infrastructure used in marine-based salmon. Knowing the composition of biofouling assemblages at specific locations provides valuable information to aid decision-making on what management practices should be prioritized. Analyzing large quantities of in situ footage of nets or processing taxonomic samples is time-consuming and a practical bottleneck for industry. This work developed deep learning methodologies to automatically segment batches of in situ images and visualize the relative abundance of important biofoulers, with the objective of providing industry with rapid-assessment tools to quantify the structure of in situ biofouling communities. The main tool developed in this paper, an Ensemble Soft Voting System that segments twelve biofouling classes along with water and unfouled netting, obtained a global accuracy of 86.7% and a mean accuracy of 67.7% across the test dataset. Consistent misclassification of stalked Pyura species as unstalked solitary ascidians and clumps of bryozoans as sponges led to discrepancy between global and mean accuracy. The “matrix” class considered by the Ensemble Soft Voting System was then further divided into two subclasses that pose different risks to farmed salmon depending on the abundance of hydrozoan taxa, using a single Convolutional Neural Network with a global accuracy of 81.8% and a mean accuracy of 84.1% across the test dataset. These tools provide industry with accurate, rapid data on the relative abundance of biofouling taxa relevant for fish health outcomes and allow industry to make operational decisions in near real-time with minimal manual input and disruption to operations. • An Ensemble Learning System was developed to segment common biofouling classes. • Robust distinction between harmful and benign biofouling communities was achieved. • Interpretable abundance information can be generated for each class. • Analysis of batches of in situ images in real-time is now possible. • These tools provide economic and environmental benefit to the aquaculture industry.