Sparse Gaussian process for online seagrass semantic mapping

Abstract Over recent decades, we have witnessed a widescale deterioration in seagrass habitats in the Mediterranean Sea. Although the causes underlying this regression are not yet clear, there is a broad consensus that new methods are required to generate high-quality data on seagrass meadows. Specifically, there is a lack of available data in order to assess changes over time in the spatial distribution of seagrass. This article introduces a new methodology to ease data-gathering operations in Posidonia oceanica meadows for benthic mapping, by using the latest developments in lightweight autonomous vehicles and image processing. The proposed methodology has been designed to build a series of online maps, by employing an input data feed based on an encoder–decoder convolutional neural network (CNN) that automatically segments the images recorded by the vehicle. In turn, it uses a sparse Gaussian Process (GP) to map uncertainty linked to the spatial distribution of seagrass. This article presents the most suitable CNN and GP, as well as performance validation results for the method, based on data acquired in three field tests.