Moving Towards Machine Learning for the Analysis of Deep-Sea Imagery Collected by Autonomous Underwater Vehicle.

Imagery captured by underwater vehicles provides valuable information on the distributions of deep-sea marine organisms and their habitats. This is particularly useful in areas that are not easily surveyed with traditional methods such as trawling. The NOAA Northwest Fisheries Science Center has used an Autonomous Underwater Vehicle (AUV) to collect still images of fish and benthic habitats off the US West Coast since 2005. This data has provided valuable ecological insights such as the distribution and abundance of corals, sponges and commercial groundfish species on rocky offshore banks. Currently, our standard analysis is carried out by experts who identify selected fish and invertebrates and measure fish lengths using custom built annotation software. However, the large number of images presents a number of challenges including analysis time. As a result, we have been trialing the use machine learning based methods using an open source toolkit, Video and Image Analytics for Marine Environments (VIAME). In 2017, there was an unusual pyrosome bloom in the NE Pacific and we observed large numbers of pyrosomes on the seafloor in AUV imagery collected off the Olympic Peninsula, Washington. We tested the rapid model generation tools in VIAME to detect pyrosomes in the AUV imagery in order to estimate densities during this event. An advantage of these tools is that they do not require a large number of pre-annotated images for model training. Initially, we encountered a large number of false positive detections but when we applied VIAME thresholding tools, automated estimates of pyrosomes numbers were much closer to the manual count. We also compared these results to the output of deep-learning models on the same images. Although there are still a number of limitations we found that using the machine learning tools had a number of advantages including, increasing our capacity to count additional target fauna and reducing manual annotation time.