Spectral features of dual-frequency multibeam echosounder data for benthic habitat mapping

Automatic methods of seafloor mapping are still in their early stage of development, despite the technical progress made in recent years. A serious imperfection is the limited types of predictor features available for seabed classification. It is therefore desirable to introduce new class of spectral features to benthic habitat mapping. In this study, we introduced eight spectral features of a rough seafloor surface that were indicative of better seabed classification. We compared them with traditional secondary features, like terrain variables and textural features. The suitability of 48 variables was tested, and the most important features were identified. The selected variables were used to perform a supervised object-based image analysis using four machine learning algorithms. We found that backscatter was the strongest predictor, followed by several spectral features from bathymetry that appeared more predictive than bathymetry itself. The highest overall accuracy of predictive model reached approximately 86% using the support vector machine classifier. The innovative results of this study suggest further application of the spectral features for predictive benthic habitat mapping, including research based on multi-frequency multibeam echosounder datasets. The utilisation of spectral features derived from bathymetry provide an important step towards more accurate maps of benthic habitats and seabed sediments composition.