[From the INDEEP mailing list]
PhD at Heriot-Watt University: for further information and to apply please contact Dr Dan Harries <D.B.Harries@hw.ac.uk>
Deadline 1 May 2015
The development of novel automated image classification as a tool in marine taxonomy and application to the assessment of marine biodiversity and biogeography
PROJECT SUPERVISOR: DR DAN HARRIES SECOND SUPERVISOR: PROF MURRAY ROBERTS
This project will develop new image classification tools for marine taxonomy. It is an interdisciplinary project that spans marine ecology, computer vision and engineering. The image analysis tools will be developed and refined by direct application to invertebrate biodiversity assessment of benthic samples derived from marine biogenic reef habitats and sea caves. Biogenic reefs are characterised by high structural complexity, high biodiversity and low resilience to anthropogenic damage. These attributes are amongst those that contribute to the high conservation value associated with such habitats. Target habitats will include shallow water reefs (e.g. horse mussels and flame shells) as well as the less easily accessible cold-water coral habitats. Sea caves represent isolated habitat islands that are characterised by a relatively diverse and abundant sponge fauna. The project will assess patterns of species occurrence in these reef and cave habitats by combining the best current practices in benthic biodiversity assessment and developing new image processing algorithms designed to automate species identification. Pre-existing samples from N Atlantic cold-water coral habitats will be utilised so the project is not reliant on further offshore expeditions, although the student would be encouraged to join a deep-water research cruise as part of their training. Additional samples from shallow-water reefs will be collected in conjunction with existing research projects. Existing sponge samples are available from a variety of cave sites and further samples will be generated by scheduled surveys at North Rona, St Kilda and Mousa over the next two years.
The identification of sponges and soft corals is a time-consuming process requiring microscopic (optical or SEM) examination of structures like spicules and sclerites. This project will explore the use of foreground object extraction methods to isolate the objects of interest. Image features will be extracted from the foreground object thereby training examples to develop an automatically-derived representation of the spicules or sclerites. This will be not only distinct with respect to the different classes of spicule or sclerite, but will also take into account their varying appearance and orientation. Much effort in Computer Vision has recently been put into invariant descriptors, but a sound representation of shape is still an open topic for exploration. Thus there are two strands of investigation that this aspect of the project will explore: the first is a semi-supervised machine learning procedure based on image feature extraction where salient features may not have obvious interpretation to the human eye; the second is a top-down process of representing rotationally invariant shapes. Both processes will require the mobilisation of prior expert knowledge which is a key component of this collaboration between Life Sciences and Engineering. It should be noted that despite the constrained nature of this problem this is an area ripe for novel image analysis which has not been thus far tackled seriously.
FURTHER INFORMATION ABOUT THE PROJECT
Over the last 15 years Harries has worked on a range of projects relating to the assessment of ecological factors influencing biodiversity of shallow water biogenic reef and sea cave habitats. Many of these projects were directly related to conservation management and policy development. Accurate and cost-effective species identification and an understanding of ecological interactions are essential pre-requisites for effective conservation management and policy. This project has potential to both facilitate more effective species identification and to advance ecological understanding of habitats of conservation importance.
Roberts has worked on the biodiversity associated with cold-water coral habitats since the late 1990s. His group works on megafaunal distribution analysis using seabed photography and video surveys and on macro- and epifaunal diversity from recovered samples (Henry & Roberts 2007; Roberts et al. 2008). Project advisor Henry has particular expertise in the statistical modelling of factors controlling epifaunal biodiversity on cold-water coral reefs (Henry et al. 2010). We would aim to incorporate these approaches including the analysis of geophysical data (in collaboration with a second ‘habitat mapping’ Watt PhD project, see Synergies below) to generate terrain variables (slope, aspect, rugosity) and hydrographic variables from existing current meter datasets (2012 Changing Oceans and earlier research cruises).
Robertson researches behaviour recognition using multi-modal signals, principally images and video but also other modalities (IR, audio) (Harding & Robertson 2012). His work is interdisciplinary, including computational models for medical physics (Robertson et al 2000), flood risk modelling (Robertson & Chan 2009), biologically-inspired visual saliency representations (Harding & Robertson 2012) and use of image segmentation techniques for barnacle species identification to assess the effects of climate change. He leads work on sensor fusion for human-robot interaction in the EU FP7 project LOCOBOT. Previous funders include the MOD Major Programme, Detection of Difficult targets and various programmes at DERA/QinetiQ. This work, in conjunction with a prestigious 1851 Royal Commission Fellowship at Oxford University, developed new behaviour recognition and explanation algorithms from video data. He has a strong track record of the successful exploitation of his algorithms in applications and systems contributing to live trials. His main interests are behaviour (Baxter et al 2012) and anomaly detection (Robertson & Letham 2012) as well as normal activity explanation (Baxter et al 2010; Robertson & Reid 2011).
Fieldwork costs are eliminated because the project has access to existing samples and fully funded surveys are already scheduled which will yield further material. Laboratory costs will be minimal because microscope facilities and associated consumables are available through existing research projects. Roberts’ lab also maintains a shared PC cluster for Geographic Information System, Graphics and Video Analysis that would be available to the student.
This project would solidify a developing collaboration between the Schools of Life Sciences and Engineering through the joint work to develop image analysis techniques as an aid to marine species identification. This aspect of the project will be strengthened through the
involvement of the Natural History Museum whose ‘DAISY’ initiative is tackling similar issues for the automated identification of insects (Reed 2010). Although based in Life Sciences, the student would be encouraged to spend time interacting with the Vision, Image, and Signal Processing group and to work for periods at the Natural History Museum, the National museum of Scotland and Ulster Museum. This would provide a valuable wider training experience and broaden the student’s experience beyond working in a single department or institution.
The project has direct synergies with existing complementary projects relating to biogenic reefs and sea caves and will add to our coherent research program thus broadening and enhancing our existing research remit.
Project outcomes will have a significant ‘impact’ through providing a cost effective tool for species identification which will be of direct benefit to marine environmental consultancies and government conservation agencies. It will also broaden understanding of the ecological patterns and connectivity of species and habitats of major conservation importance. The cold-water coral samples are from poorly-understood deep-water habitats where undescribed invertebrate species are often discovered (e.g. Henry & Roberts (2007) reported at least 10 undescribed species in just 10 box core samples from comparable N Atlantic coral carbonate mound samples). The development of automated image processing techniques to identify sponges and octocorals is innovative and if successful would be a very useful development with commercial potential, not just in academic research but with government researchers and marine environmental consultancies.
Baxter R, Lane DM, Robertson NM (2012) Recognising High-Level Agent Behaviour in Data Scarce Domains. IEEE Transactions on Systems, Man and Cybernetics (part B). in press
Baxter R, Robertson NM, Lane DM (2010) Probabilistic Behaviour Features: Feature-Based Behaviour Recognition in Data-Scarce Domains. IET International Conference on Information Fusion, Edinburgh, July 2010.
Harding P, Robertson NM (2012) Visual Saliency from Image Features with Application to Compression. Cognitive Computation (DOI: 10.1007/s12559-012-9150-7)
Henry L-A, Davies AJ, Roberts JM (2010) Beta diversity of cold-water coral reef communities off western Scotland. Coral Reefs 29: 427-436
Henry L-A, Roberts JM (2007) Biodiversity and ecological composition of macrobenthos on cold-water coral mounds and adjacent off-mound habitat in the bathyal Porcupine Seabight, NE Atlantic. Deep-Sea Research Part I 54: 654-672
Reed S (2010) Pushing DAISY. Science 328: 1628-1629
Roberts JM, Henry L-A, Long D, Hartley JP (2008) Cold-water coral reef frameworks, megafaunal communities and evidence for coral carbonate mounds on the Hatton Bank, north east Atlantic. Facies 54: 297-316
Robertson NM, Chan T (2009) Aerial Image Segmentation for Flood Risk Analysis. IEEE International Conference on Image Processing, Cairo
Robertson NM, Diaz-Gomez M, Condon B (2000) Estimation of torque on mechanical heart valves due to MRI including an estimation of the significance of the Lenz Effect using a computational model. Physics in Medicine and Biology (45) 200: 3793-3807
Robertson NM, Letham J (2012) Contextual Person Detection in Outdoor Scenes. Proc. European Conference on Signal Processing (EUSIPCO 2012), Bucharest, Romania
Robertson NM, Reid (2011) Automatic Reasoning about Causal Events in Surveillance Video. EURASIP Journal on Image and Video Processing Special Issue on Advanced Video Based Surveillance (DOI:10.1155/2011/530325)
We invite research leaders and ambitious early career researchers to join us in leading and driving research in key inter-disciplinary themes. Please see www.hw.ac.uk/researchleaders for further information and how to apply.
Heriot-Watt University is a Scottish charity registered under charity number SC000278.