Camera traps, image analysis and wild mammal monitoring

Camera traps could deliver data sets of high ecological importance through widespread monitoring of whole mammal communities. Nevertheless, widespread community monitoring is associated with a range of challenges related to the resources required to process and analyse large volumes of image data. There is an urgent need for techniques to automate aspects of camera trap image processing and analysis. This project will use image processing to improve the engagement of citizen scientists, to enable community monitoring and abundance estimation for mammals, and to answer fundamental questions about animal movement. The project will use a combination of statistical image processing, ecological theory, and survey methodological development to address 3 aims:

1. Develop image recognition tools to increase engagement of citizen scientists in data processing. Citizen science presents enormous potential for tackling the data processing challenge. Data from a citizen science project show that citizen scientists involved in image classification can be deterred by large volumes of images containing no animals. We will use techniques, potentially including image subtraction, change detection and object recognition, to (i) identify images containing humans (which should not be made available on crowdsourcing websites); and (ii) discriminate between images containing objects of interest and those that do not.

2. Track digital objects and calibrate camera trap capture surfaces. A recent focus of the project partner has been the development of techniques to estimate the abundance and movement speeds of unmarked species from camera trap images. These innovations can revolutionise the way that camera traps are used, making typical study designs applicable to any species and thereby making community-level monitoring commonplace. This requires a calibration process to model the ground surface positions of animals, based on their positions within images, requiring manual digitisation of animal paths. However, this process is labour-intensive, and has great potential for further development. We will use object recognition, geometry and optical physics to automate animal tracking within images, and to simplify the ground surface calibration process while maintaining or improving accuracy.

3. Determine the environmental drivers of travel speed. Animal movement rates are fundamental to a wide range of ecological applications, including many monitoring approaches, as well as predicting encounter rates for use in predator-prey and disease modelling. Nevertheless, recent work with camera traps has cast doubt on existing estimates of travel speed determined using other methods. Using the innovations described above, we will work with available data sets (many of which have been collected by our CASE and project partners and their network) to assess the body mass scaling of travel speed. The student will use automated techniques to re-analyse georeferenced image data, collating it with covariates such as estimated species body mass, trophic level, local productivity (indexed by NDVI) and, where possible, population density. Statistical model selection will be used to determine support for the influence of additional covariates (beyond body mass) in explaining travel speeds.