Using remote imaging and machine learning tools to quantify ecological impacts of red deer (Cervus elaphus)


The spatial distribution and movement of animal populations shapes a variety of ecological processes, including resource-consumer interactions, inter-specific competition and parasite/disease dynamics. However, quantifying animal occurrence over large landscapes can be difficult, particularly in systems where animals move heterogeneously with respect to habitat. Thus, a key limitation in understanding the drivers of ecological interactions involving mobile species is our ability to reliably measure animal occurrence across time and space. This issue is particularly relevant to Scottish red deer populations, which occur in increasingly high densities in many areas. High deer abundance has a range of ecological consequences, including increased abundance of ticks, which commonly use deer as reproduction hosts (Millins et al, 2017). Standard census methods for deer are either extremely costly (e.g. helicopter surveys), and thus are difficult to repeat multiple times within a season, or are dependent on indirect indices (e.g. dung surveys) which are plagued by methodological biases and detection issues. When linking distribution and movement patterns of deer to highly dynamic ecological processes, such as tick abundance and distribution, much finer temporal resolution, high-quality data are needed.

Capitalising on ongoing studies of deer, livestock and ticks in the Uist Islands, and building on partnerships with Scottish Natural Heritage (SNH, Case Partner) and deer management stakeholders, this PhD project will use novel imaging technologies to quantify the distribution and movement of red deer and interactions with livestock and human presence, which are key to understanding exposure to parasites (ticks) and tick-borne diseases (e.g. Lyme borreliosis). Wildlife survey techniques developed at the University of Glasgow (Torney et al. 2019) will be employed to enhance methods for counting deer from aerial drone video and camera traps across two areas (the Uists and an area of the Scottish mainland with well-monitored deer abundances). These sites are ideal for the development of these methods because deer occur in relatively open vegetation and populations are relatively small and isolated. The student will deploy machine learning (ML) algorithms to efficiently classify and census animals from imagery. Information generated through this project will directly inform management actions aimed at reducing tick prevalence and other detrimental impacts of deer. For instance, both sheep and deer are important hosts for ticks, yet there are often conflicting views about which host should be the target of management interventions, such as lethal control, fencing, reduction of stocking densities and habitat manipulation. Thus, one broader aim of this PhD project is to shed new light on the relative roles of deer versus sheep in driving the spatiotemporal dynamics of tick populations in rural Scotland. Control of deer densities is also essential for improving sapling survival in regenerating forests, reducing vehicle collisions on roadways and optimizing harvest quotas. Accurate estimates of deer population densities are needed, and this project, with the support of Case Partners at SNH, has the potential to transform the tools and our ability to make ecological inferences related to deer distribution and movement in a variety of habitats (e.g. farmland, heath, bog, woodland) in Scotland.

Click on an image to expand

Image Captions

iapetus1.png – Example of applying machine learning algorithms to classify animals from survey image Adapted with permission from Stacey Fulford (U Glasgow)
iapetus2.png – Red deer stag. Privately owned image (L Gibert, U Glasgow)


Development of deer survey methods to estimate abundance. An important first step of automating counts of objects in images is to develop a sufficiently large dataset to train machine learning algorithms. This requires manually locating animals in a wide range of images with different backgrounds and delineating animals using annotation software. Larger training datasets result in more accurate identifications (i.e. fewer false negative/positive errors). Thus, the first objective will be to collect a large number of images of deer from drones and camera traps, and manually annotate these for use with training ML models. A subset of images will be kept aside for validation and accuracy assessment. All wildlife surveys, even so-called ‘total counts’, suffer from a failure to detect every individual present in a population, and detection rates can depend strongly on habitat (Morrison et al. 2018). To evaluate the effect of detection accuracy on survey estimates of aerial transect surveys and camera traps, the student will run simulations of survey data across a range of key parameters (encounter rate, degree of clustering and accuracy). To validate survey accuracy, surveys will be flown intensively across the Uists and Highland sites on a 3-4 week cycle to estimate abundance and distribution over time. Estimates from drone surveys will be compared to those from other datasets collected by the Scottish local Deer Management Groups.

Camera traps to estimate diurnal movements. Because drone surveys will only be conducted during daylight hours, there is a need to understand where and when deer move throughout the day and night. In collaboration with SNH scientists, a network of camera traps (n=25) will be deployed along the movement paths of deer in each site throughout the year. Paths will be identified through a combination of input from local farmers and from preliminary aerial survey data. We will employ similar machine learning techniques for identifying and classifying deer from camera trap images to enhance efficiency of detecting and counting deer. Camera traps will also provide behavioural data (e.g. walking versus foraging) which can be detected using ML techniques (Norouzzadeh et al. 2018) and population age structure. Statistical analysis of camera trap data will be supported by Park and Morrison.

Deer-habitat suitability to predict tick abundance. Presence-absence data from aerial surveys will be used to predict deer habitat suitability. Predictions of deer use across a given landscape will be based on resource selection models that incorporate environmental covariates (habitat type, proximity to humans, terrain ruggedness, aspect, hydrological flow). These suitability models will provide prediction of tick abundances over large areas. Notably, suitability estimates will be compared with actual tick abundances from surveys conducted as part of the larger Uist tick project by Gilbert, Millins and Biek.

Project Timeline

Year 1

Literature review; Development of machine learning and survey methodology; training courses in statistics, programming (Python), Spatial Ecology, GIS, Single Species Modelling

Year 2

Fieldwork: drone surveys and camera trap deployment at first site; analysis of preliminary data and refinement of methodologies; Attendance at British Ecological Society (BES) Movement Ecology Special Interest Group

Year 3

Fieldwork: drone surveys and camera trap deployment at second site; Analysis of data; writing chapters; presentation at BES annual meeting.

Year 3.5

Finalise thesis and conduct Viva. Funding to cease after 3.5 years.

& Skills

The student will join a diverse and productive multi-disciplinary research body at Glasgow’s Institute of Biodiversity Animal Health and Comparative Medicine (IBAHCM), home to leading experts on spatial statistical techniques and quantitative analysis of animal movement. The IBAHCM brings many years of collective experience conducting ecological and epidemiological research in various settings. IBAHCM offers courses in (1) Spatial Ecology, (2) Bayesian Statistics, (3) Human Dimensions of Conservation, (4) Single Species population models, among many others. There will also be access to a unique set of Special Interest Groups in the IBAHCM, which will give the student an added support network beyond their direct supervision. For example, the Spatial Ecology Group focuses on measuring species distribution, and applying novel statistical approaches to predict how organisms move and spread; the Disease Ecology Group emphasises ecological interactions relevant to the transmission of parasites and pathogens. The student would be encouraged to interact with an active group of ecologists, epidemiologists and statisticians (e.g. Torney, Babayan) using machine learning for various applications in IBAHCM and The Maths and Statistics Department.

References & further reading

Millins, C., Gilbert, L., Medlock, J., Hansford, K., Thompson, D. B., & Biek, R. 2017. Effects of conservation management of landscapes and vertebrate communities on Lyme borreliosis risk in the United Kingdom. Philosophical Transactions of the Royal Society of London Series B-Biological Sciences, 372(1722), 20160123.

Morrison, T. A., A. B. Estes, S. A. R. Mduma, H. T. Maliti, H. L. Frederick, H. Kija, M. Mwita, A. R. E. Sinclair, and E. M. Kohi. 2018. Informing aerial total counts with demographic models: population growth of Serengeti elephants not explained purely by demography. Conservation Letters 11:1-8.

Norouzzadeh, M. S., A. Nguyen, M. Kosmala, A. Swanson, M. S. Palmer, C. Packer, and J. Clune. 2018. Automatically identifying, counting, and describing wild animals in camera-trap images with deep learning. Proceedings of the National Academy of Sciences of the United States of America 115:E5716-E5725.

Torney, C. J., D. J. Lloyd-Jones, M. Chevallier, D. C. Moyer, H. T. Maliti, M. Mwita, E. M. Kohi, and G. C. Hopcraft. 2019. A comparison of deep learning and citizen science techniques for counting wildlife in aerial survey images. Methods in Ecology and Evolution 10:779-787.

Further Information

Applications: to apply for this PhD please use the url:


Uist Deer Management Plan, 2017. Accessible at:

Apply Now