Designing a validation ground truth dataset for habitat mapping using Multibeam sonar data, Object-Based Image Analysis and AUV video
Abstract
Effective marine environment monitoring programs require accurate high-resolution maps of benthic habitats. The past decade has witnessed tremendous improvements in the remote-sensing technologies and data processing... [ view full abstract ]
Effective marine environment monitoring programs require accurate high-resolution maps of benthic habitats. The past decade has witnessed tremendous improvements in the remote-sensing technologies and data processing methodologies used in this purpose. Multibeam echosounders (MBES) are now routinely used to provide the high-resolution, full-coverage datasets characterising the seafloor depth and structure while Autonomous Underwater Vehicles (AUV) fitted with high-resolution video cameras provide the ground-truth data to accurately identify and locate benthic habitats and organisms. A large number of methodologies to process these datasets and integrate them into a benthic habitat map are now available. However, despite these advances, the accuracy of the final map and hence its suitability for monitoring and management purposes remains highly dependent on the ground-truth data used for model validation: it needs to (1) cover all habitats present, (2) target each habitat type in a number of instances that is large enough to overcome the stochastic variability in the MBES data and (3) samples need to be spread wide enough from one another to overcome spatial autocorrelation and ensure statistic independence.
Using MBES and AUV video datasets acquired in 2013 in Refuge Cove, in the Wilsons Promontory Marine Park (Victoria, Australia), we developed a methodology to create a validation ground-truth sampling design that satisfies these conditions. The AUV survey was designed based on prior knowledge, observations and expertise of the site’s habitats. Benthic habitat classes were defined from this video dataset. The minimum distance between samples to ensure spatial independence was calculated using Moran’s I. An Object-Based Image Analysis was run on the MBES data products and followed by a clustering algorithm to map the broad spatial variability of the seafloor structure. This process resulted in a sampling scheme composed of locations from the original AUV video data and complementary locations that were later sampled with a drop video camera survey. The ground-truthing dataset thus obtained was used for validation of a habitat map obtained from integrating MBES and AUV data into a supervised QUEST decision tree, resulting in an overall accuracy of 78%.
Authors
-
Alexandre Schimel
(Centre for Integrative Ecology, School of Life & Environmental Sciences, Deakin University)
-
Daniel Ierodiaconou
(Centre for Integrative Ecology, School of Life & Environmental Sciences, Deakin University)
-
Mary Young
(Centre for Integrative Ecology, School of Life & Environmental Sciences, Deakin University)
-
Markus Diesing
(Centre for Environment, Fisheries and Aquaculture Science)
-
Anna Downie
(Centre for Environment, Fisheries and Aquaculture Science)
-
Matt Edmunds
(Australian Marine Ecology)
-
Sean Blake
(Deakin University)
-
Brett Mitchell
(Parks Victoria)
-
Grace Gaylard
(Deakin University)
Topic Area
11 - Using Monitoring to Map the Marine World
Session
OS-11D » Monitoring to map the marine world (13:40 - Thursday, 9th July, Lecture Theatre D2.211)
Presentation Files
The presenter has not uploaded any presentation files.