Benjamin Evans

Allan Tucker

Chris Carbone

Recent years have seen an increase in camera-trap survey monitoring by ecological researchers. Camera-trap surveys collect imagery of medium-large mammal species across a region of interest. Dependent on the activity in an areas and number of camera trap days a survey’s conducted, millions of images may be captured. Currently researchers label each image with species and behavioural information or enlist Citizen Science volunteers.
Developments in machine learning, specifically deep learning has provided promising methods for general image recognition and may be utilised to assist in the camera trap labelling. Current models and methods that have been developed do not generalise well past the dataset they’re trained on.
The aim of the following research is to produce a technique that’s able to generalise to new data from around the world. We predict that introducing bio-geography into the classifier will increase the accuracy along with using multiple models trained for similar sub-species with an overarching model guiding which sub-species model should be used to classify.

• Benjamin C. Evans, Allan Tucker, Oliver R. Wearn, Chris Carbone (2021) Reasoning About Neural Network Activations: An Application in Spatial Animal Behaviour from Camera Trap Classifications ECML PKDD 2020 Workshops. ECML PKDD 2020. Communications in Computer and Information Science, vol 1323. Springer, Cham. DOI: https://doi.org/10.1007/978-3-030-65965-3_2