Computer-assisted identification of small Australian mammals in camera trap imagery

Abstract

We investigated the feasibility of identifying Australian native small mammals via semi-automated analysis of camera trap imagery. A portion of the animal's pelage ('texture patches') was extracted from the upper body of three known small mammal species in colour camera trap images and used to identify two target rodent species (the bush rat 'Rattus fuscipe's and the Hastings River mouse 'Pseudomys oralis') and dasyurids ('Antechinus' spp.), which are challenging to distinguish from one another in camera trap images. Numerical descriptors based on the differences of the means and standard deviations of each of the colour-space normalised texture patches (blocks of image pixels) were used to distinguish the species. A support vector machine classification model was developed using the numerical descriptors (from the original training data used to estimate the model), which correctly labelled species with 97.78 ± 0.45% accuracy. Model robustness was also assessed using 10-fold cross-validation, which estimated model accuracy in general as 91.11 ± 0.45%. We demonstrate therefore that computer-assisted small mammal identification is feasible and that the algorithm developed might be useful in assisting with recognition of small mammals in situations where it is difficult for the user to achieve this visually.