Data Study Group Final Report: Centre for Environment, Fisheries and Aquaculture Science (CEFAS)

Sea pens are colonial animals closely related to the corals that live attached to the sea floor. They generally form a feather-shaped colony with a horny skeleton. They respond to multiple human interventions (contamination and human waste, fishing, boats, etc.) and therefore they are an indicator of the health of their ecosystem. Sea pen communities are Vulnerable Marine Ecosystems (VMEs), which have been proposed as indicators of good conservation status for mud habitats. The goal of this challenge is to study their distribution and abundance, so their presence can be used to evaluate whether the seabed is exposed to good water quality and is undamaged by fishing gear.

The Centre for Environment, Fisheries and Aquaculture Science (Cefas) studies the distribution of sea pens to obtain marine environmental insights about the health of mud ecosystems. From previously acquired video footage, they were interested in classifying two types of sea pens which are distributed at very different densities in two different survey locations: the slender sea pen (Virgularia mirabilis) and the phosphorescent sea pen (Pennatula phosphorea). Both sea pens live in fine sediments ranging from sheltered inshore waters to deeper water offshore (∼ 10-400 m depth). The slender sea pen has a central stem only a few millimetres thick lined by small tentacled polyps arranged in two opposing lateral rows on the central stem. The colony varies in colour from white to creamy yellow and can grow up to 60 cm long. Colonies of the phosphorescent sea pen on the other hand are stout and fleshy and up to 40 cm long. The triangular leaf-like branches formed of fused polyps range from translucent to a deep reddish-pink colour.

Prior to this DSG, Cefas had created an analysis pipeline using preliminary machine learning algorithms which would classify sea pens from seafloor video footage with varying accuracy. A key challenge identified was that of standardising all video data, which showed different levels of brightness, contrast and other characteristics. This was rather challenging, as different years use different lighting systems, and it can be hard to spot sea pens under LED lighting conditions.

In summary, the preliminary models did not generalise well over different years, since the characteristics of the cameras and the lightning conditions changed over time. The aim of this DSG project and report is to attempt to find a machine learning pipeline based on computer vision techniques that can accurately classify and detect these two types of seapens across footage from several years, as well as gain some insights into which video modification techniques can be applied to homogenise the different video conditions over the years.