Publication Abstract

Habitat maps derived from multibeam acoustic data using Random Forest and Object-Based Image Analysis

Markus Diesing*, David Stephens*, Matthew Curtis*, Bill Meadows*, Neil Golding

The United Kingdom’s (UK) Marine and Coastal Access Act 2009 provides the legal basis to designate and protect Marine Conservation Zones (MCZs). These are a type of marine protected area, which will exist alongside European Marine Sites (Special Areas of Conservation and Special Protected Areas); Sites of Special Scientific Interest, Ramsar Sites and other national designations to form an ecologically coherent network of marine protected areas. Whilst recommendations for a suite of MCZs have recently been made, it became apparent that further work needs to be undertaken to strengthen the evidence base for some of the recommended MCZs. The primary aim of this study was to improve the understanding of seabed substrates and habitats in selected areas of the UK’s offshore waters where recommended MCZs overlap with existing multibeam acoustic data.

We applied semi-automated approaches to map seabed substrate type and composition. The Random Forest machine learning algorithm belongs to the family of decision tree learning and was employed to predict sediment composition (content of mud, sand and gravel). It proved to be effective in handling large amounts of input data layers derived from multibeam acoustic data.

Additionally, Object-Based Image Analysis was employed to map exposed bedrock. For input image layers, we used multibeam backscatter and the bathymetric position index (BPI) calculated from the multibeam bathymetry data. Experimentally determined threshold conditions for backscatter strength and BPI were employed to delineate exposed bedrock.

Overall, the predictions produced plausible results in agreement with general knowledge of the site. Statistical indices of accuracy did, however, indicate lower levels of agreement. The most likely reason for this are positioning inaccuracies of the legacy sampling data, which were collected prior to the introduction of GPS positioning and are subject to an unknown but potentially significant positional error.