An evaluation of supervised and unsupervised classification techniques for marine benthic habitat mapping using multibeam echosounder data

Calvert, Jay; Strong, James Asa ORCID: https://orcid.org/0000-0001-8603-097X; Service, Matthew; McGonigle, Chris; Quinn, Rory. 2014 An evaluation of supervised and unsupervised classification techniques for marine benthic habitat mapping using multibeam echosounder data. ICES Journal of Marine Science, 72 (5). 1498-1513. 10.1093/icesjms/fsu223

Full text not available from this repository.

Abstract/Summary

Marine habitat mapping provides information on seabed substrata and faunal community structure to users including research scientists, conservation organizations, and policy makers. Full-coverage acoustic data are frequently used for habitat mapping in combination with video ground-truth data in either a supervised or unsupervised classification. In this investigation, video ground-truth data with a camera footprint of 1 m2 were classified to level 4 of the European Nature Information System habitat classification scheme. Acoustic data with a horizontal resolution of 1 m2 were collected over an area of 130 km2 using a multibeam echosounder, and processed to provide bathymetry and backscatter data. Bathymetric derivatives including eastness, northness, slope, topographic roughness index, vector rugosity measure, and two measures of curvature were created. A feature selection process based on Kruskal–Wallis and post hoc pairwise testing was used to select environmental variables able to discriminate ground-truth classes. Subsequently, three datasets were formed: backscatter alone (BS), backscatter combined with bathymetry and derivatives (BSDER), and bathymetry and derivatives alone (DER). Two classifications were performed on each of the datasets to produce habitat maps: maximum likelihood supervised classification (MLC) and ISO Cluster unsupervised classification. Accuracy of the supervised habitat maps was assessed using total agreement, quantity disagreement, and allocation disagreement. Agreement in the unsupervised maps was assessed using the Cramer's V coefficient. Choice of input data produced large differences in the accuracy of the supervised maps, but did not have the same effect on the unsupervised maps. Accuracies were 46, 56, and 49% when calculated using the sample and 52, 65, and 51% when using an unbiased estimate of the population for the BS, BSDER, and DER maps, respectively. Cramer's V was 0.371, 0.417, and 0.366 for the BS, BSDER, and DER maps, respectively.

Item Type:	Publication - Article
Digital Object Identifier (DOI):	10.1093/icesjms/fsu223
ISSN:	1054-3139
Date made live:	15 Nov 2018 16:38 +0 (UTC)
URI:	https://nora.nerc.ac.uk/id/eprint/521576

Actions (login required)

View Item

Document Downloads

More statistics for this item...