Measuring the temporal evolution of seagrass Posidonia oceanica coverage using autonomous marine robots and Deep Learning

Abstract

This paper describes an advanced methodology to monitor and assess, in temporal series, meadows of the seagrass Posidonia oceanica. The process includes, the following steps: (a) exploring marine regions of certain biological interest with Autonomous Underwater Vehicles equipped with cameras looking downwards, (b) taking images continuously during missions of preprogrammed trajectories, (c) processing the images off-line to build colour photomosaics, (d) segmenting seagrass from the background in every image using a pretrained Neural Network, (e) computing the same photomosaics but using the segmented images, and (f) computing automatically the bottom coverage of the seagrass counting the proportion of pixels labelled positively. This procedure avoids the involvement of divers, allows increasing depth, extension and duration of missions and offers 2D maps of the whole inspected areas in a single view, which allows us to get more accurate coverage measurements than those obtained with traditional techniques.

Experiments have been performed with datasets collected in areas of the Balearic Islands colonized with P. oceanica seagrass and subject to low and high touristic and anchoring pressure during high season, repeating the same transects in consecutive years in order to obtain interannual results. Data obtained with this methodology permit a direct biological qualitative, quantitative and temporal analysis and interpretation, such as the percentage of temporal decline of seagrass coverage in some of the surveyed areas and the annual increase of the meadows extension in others.