Publication Abstract
- Title
-
A multi-model ensemble approach to seabed mapping
- Publication Abstract
-
A multi-model ensemble approach to seabed mapping
Markus Diesing*, David Stephens*
The advent of multibeam echosounder (MBES) technology has revolutionised our ability to map and represent the seabed of oceans and marginal seas in recent years. We are now able to map the seabed at unprecedented spatial resolution and accuracy. At the same time, the data volumes are dramatically increasing. In the United Kingdom (UK), the increase in seabed coverage with MBES data is mainly driven by the Civil Hydrography Programme. It has been estimated that seabed coverage has increased in recent years to 200,000 km2 or approximately 26% of the UK Continental Shelf.
Those large data volumes are not effectively analysed by eye using expert interpretation. We therefore compared the performance of six supervised machine learning techniques (k-Nearest Neighbours, Decision Tree, Random Forest, Support Vector Machines, Artificial Neural Networks and Naïve Bayes) in their ability to map seabed substrate classes. We selected ‘Bayman’s Hole to Dunbar’ off the north-east coast of England as a test site covering a seabed area of 5272 km2. Input features were MBES bathymetry, backscatter and several derivatives thereof. Sample data were derived from a legacy dataset of seabed samples from the British Geological Survey (BGS Legacy Particle Size Analysis uncontrolled data export (2011), British Geological Survey, www.bgs.ac.uk). Sample data were split into training and test sets in a random stratified way, to allow an assessment of model performance. Overall map accuracy, kappa statistic and balanced error rate were calculated. The influence of the choice of input features on prediction performance was also tested.The three best performing models achieved overall map accuracies of 80 – 81%, kappa statistics of 0.45 – 0.50 and balanced error rates of 0.37 – 0.41. The outputs of these models were combined to an ensemble map using a simple voting procedure to determine substrate class. A major advantage of this approach is that it also yields the agreement between model predictions which could be used as a measure of class allocation uncertainty. We found that in 70.3% of seabed area, all three models agreed, in 27% of area at least two models agreed and in only 2.7% of area there was no agreement between models.
- Publication Internet Address of the Data
- Publication Authors
-
Markus Diesing*, David Stephens*
- Publication Date
- September 2013
- Publication Reference
-
MESH Atlantic Final Conference, Aveiro, Portugal, 15-17 September 2013
- Publication DOI: https://doi.org/
