At depths below the ocean’s surface, reaching hundreds of meters, a fascinating array of marine creatures thrive amidst the absence of sunlight. These remarkable deep-sea coral reefs provide a habitat for an astonishing diversity of bottom-dwelling species, including squat lobsters, sea spiders, crabs, brittlestars, and soft corals, among numerous others.

Deep-sea coral reefs possess a paradoxical nature, as they are both sturdy and delicate, susceptible to damage from the trawling nets dragged across the seabed by industrial fishing vessels. Once these vulnerable marine ecosystems are harmed, their recovery is a slow process. Consequently, the mapping and monitoring of these reefs have become of utmost importance to scientists, conservationists, and sustainable fisheries, aiming to safeguard and preserve these precious habitats.

The task of observing the seabed presents a formidable and expensive challenge. However, one solution that has emerged is the use of image-based surveys. This method entails trailing a setup of cameras behind a vessel to capture high-resolution images of the seabed. While the process of capturing these images is relatively straightforward, the subsequent step of analysing and interpreting each frame poses a significant undertaking for scientists.

During this review process, scientists meticulously examine each image frame by frame, aiming to discern various features of the seafloor. Their primary objective is to differentiate established coral from other elements such as coral debris, rubble originating from other living organisms, sand, gravel, and rocks. This intricate task requires careful scrutiny and expertise to accurately identify and classify the distinct components captured in the images.

Deep learning aids deep sea monitoring

The aforementioned time-consuming and labour-intensive process presents a significant bottleneck, constraining the available data for making informed decisions regarding the effectiveness of strategies aimed at protecting deep-sea coral reefs.

Artificial intelligence (AI) plays a crucial role in addressing this challenge. Researchers have developed a deep learning system that possesses the capability to analyse images and swiftly identify and measure deep-sea coral reefs, surpassing human efficiency.

Comprising a multi-layered network of artificial neurons, the system emulates the functioning of the human brain by acquiring the ability to discern intricate patterns within datasets. By leveraging this AI technology, the identification and measurement of deep-sea coral reefs can be accomplished in a fraction of the time required by humans.

According to Chris Jackett, a research scientist at CSIRO and one of the authors of the study, the deep learning system has demonstrated remarkable speed and accuracy. Jackett highlighted that the trained model successfully classified over 2,300 images in under 20 minutes, a task that would typically require more than three months for a person to accomplish.

The deep learning system was trained to distinguish between six seafloor features, including established coral, coral rubble, rubble from other living organisms, sand or mud, pebble or gravel, and rock, achieving an impressive accuracy rate of 98.19%. Remarkably, in certain cases, the model exhibited more consistent performance than a human observer.

The training process utilized images collected from the RV Investigator, which undertook a voyage off the southern coast of Tasmania in 2018. A deep tow camera system was employed, capturing continuous video and taking photos every five seconds at depths ranging from 600 to 1,800 meters below the ocean’s surface.

Researchers manually reviewed approximately 6,000 photos, which were then utilized to train the deep learning system. The initial dataset comprised a massive collection of 140,000 data points or “snips.”

The data used for training the deep learning system turned out to be “noisy,” as a significant number of snips contained a mix of various features, had issues with lighting (being too dark or too light), or were affected by photobombing from local fauna like seastars or urchins.

The researchers undertook the task of cleaning the data, reducing it to approximately 70,000 snips that met the necessary criteria for the model to learn effectively. They conducted experiments with different training methodologies and network structures until they discovered a combination that achieved an impressive accuracy rate of 98.18%.

During the testing phase, the final deep learning model exhibited comparable performance to a human observer when confronted with uncleaned data that it had never encountered before. The model consistently demonstrated accuracy, even in scenarios involving complex images with multiple features that can pose challenges for human classification.

AI accelerates the process, but human involvement remains vital. Chris Jackett suggests combining human input with the deep learning model for efficient analysis of new images. AI aids in protecting fragile ecosystems like deep-sea coral reefs, benefiting deep-sea fisheries and addressing conservation challenges in our oceans.