Use of machine-learning algorithms for the automated detection of cold-water coral habitats: a pilot study

Cold-water coral reefs are recognised as important biodiversity hotspots on the continental margin. The location of terrain features likely to be associated with living reef has been made easier by recent developments in acoustic sensing technology. For accurate assessment and fine-scale mapping of these newly identified coral habitats, analysis of video data is still required. In the present study we explore the potential of manual and automatic abundance estimation of cold-water corals and sponges from still image frames extracted from video footage from Tisler Reef (Skagerrak, Norway). The results and processing times from 3 standard visual assessment methods (15-point quadrat, 100-point quadrat and frame mapping) are compared with those produced by a new computer vision system. This system uses machine-learning algorithms to detect species within frames automatically. Cold-water coral density estimates obtained from the automated method were similar to those gained by the other methods. The automated method slightly underestimated (by 10 to 20%) coral coverage in frames which lacked a uniform seabed illumination. However, it did much better in the detection of small live coral fragments than the 15-point method. For assessing sponge coverage, the automated system did not perform as satisfactorily. It mistook a percentage of the seabed for sponge (0.1 to 2% of most frames) and underestimated sponge coverage in frames that contained many sponges. Results indicate that the machine-learning approach is appropriate for estimating live cold-water coral density, but further work is required before the system can be applied to sponges within the reef environment.