High-resolution imaging in marine habitats using Autonomous Underwater Vehicles
Professor Stefan B Williams, University of Sydney, Australia
This talk will describe insights gained from a decade of autonomous marine systems development at the University of Sydney’s Australian Centre for Field Robotics. Over the course of this time, this group has developed and deployed numerous underwater vehicles and imaging platforms in support of applications in engineering science, marine ecology, archaeology and geoscience. The group has operated an Australia-wide benthic observing program designed to deliver precisely navigated, repeat imagery of the seafloor. This initiative makes extensive use of Autonomous Underwater Vehicles (AUVs) to collect high-resolution stereo imagery, multibeam sonar and water column measurements on an annual or semi-annual basis at sites around Australia, spanning the full latitudinal range of the continent from tropical reefs in the north to temperate regions in the south. The program has been very successful over the past decade, collecting millions of images of the seafloor around Australia and making these available to the scientific community through online data portals developed by the facility and affiliated groups. These observations are providing important insights into the dynamics of key ecological sites and their responses to changes in oceanographic conditions through time. The group has also contributed to expeditions to document coral bleaching, cyclone recovery, submerged neolithic settlement sites, ancient shipwrecks, methane seeps and deepwater hydrothermal vents. The talk will also consider how automated tools for working with this imagery have facilitated the resulting science outcomes and will explore opportunities to extend these techniques to the study of deep-sea science and exploration.
Whole site multi-resolution photogrammetric surveys of deep-sea vents and cold seeps
Dr Blair Thornton, University of Southampton, UK
There are many applications in marine science and monitoring that require high-resolution images of the seafloor to be obtained. However, the resolution of underwater observations are often at a trade-off with the extent over which they can be made, and this limits their usefulness in non-uniform seafloor environments where the distribution of features varies over spatial scales much larger than the footprint that can be observed, for example in a single image frame. This talk will describe recent efforts to address scale relevance in seafloor imaging applications by using autonomous underwater vehicles instrumented with systems that can image the seafloor from different altitudes, and build multi-hectare 3D visual reconstructions of the seafloor with resolutions with sufficiently high-resolution where needed. This allows continuous wide-area, multi-resolution 3D reconstructions of the seafloor to be generated, allowing patterns to be explored and interpreted over a large range of spatial scales that would not otherwise be possible. This approach will be described giving examples of data recently obtained in deep-sea hydrothermal vent and gas hydrate fields.
An overview of recent developments in artificial intelligence and its potential applications to deep-sea exploration
Professor Jeremy Wyatt, University of Birmingham, UK
Dr Mohan Sridharan, University of Birmingham, UK
AI in general and machine learning in particular have created a great deal of recent interest, not to mention hyperbole. In this talk, Jeremy Wyatt and Mohan Sridharan will give a whistle-stop tour of the current state of AI. The speakers shall describe fundamental algorithmic advances in perception, reasoning, manipulation and learning. The emphasis will be on the nature and properties of algorithms that may have utility in deep-sea exploration. Although the speakers are not currently working on AUVs, they will discuss some case studies from their own work and that of others. The talk will also cover the obvious challenges in practical deployment and some ways that these challenges may be overcome.
Sampling our oceans and issues of taxonomic identification
Dr Philip Culverhouse, Plymouth University, UK
A severe bottleneck exists in the ecological study of our oceans and seas. This has arisen because net-sampling, having been the mainstay of biological oceanographers for over a century, has not provided sufficient data of the distributions, dynamics and populations of important taxa to answer questions posed by researchers today and in the future. In the last decade or so, in-situ, towed or ship-based imaging instruments have been developed that potentially may be able to address the shortfall. These instruments can, and will, provide denser spatial and temporal sampling that is more cost-effective than is possible with net sampling. However, there is problem. We do not have enough skilled people to identify the video and photographic hauls from the water.
Artificial Intelligence has been mooted as the solution, automating identification using the latest ideas in Deep Learning. However, to be successful in using computers to automate taxonomic identification of organisms, the machines will have to learn from the vast literature that describes our natural world. Deep learning is beginning to revolutionise computer-based visual recognition, but it must be tied to natural language and the descriptions that people make in the scientific literature to be able make the transition from laboratory-sized machine classifiers to globally useful tools. We must integrate the existing taxonomic knowledge that is written and drawn with emerging AI tools. Culverhouse will explore this issue.