Innovations in camera trapping: Streamlining image classification and adapting camera traps for small mammal monitoring

Abstract

In the past 20 years, camera trapping has emerged as a revolutionary technique for monitoring elusive animals that are otherwise difficult to observe. With growing uptake of camera traps for projects of ever-increasing scale, challenges have emerged in processing the vast quantities of data that camera traps can produce. Additionally, due to the many advantages of camera trapping compared to conventional survey methods, there is increasing appetite for extending the methodology beyond the medium-to-large mammal species that have, until now, been the main focus of camera trapping studies. In this thesis, I aimed to investigate these two frontiers in camera trap research. Firstly, I considered and compared the accuracy of several types of classifier at labelling a camera trap image dataset. The classifiers included members of the public who were registered users of an online citizen science platform, anonymous users who participated via public terminals, and an AI model trained to recognise UK mammal species. In addition, I designed combined approaches relying on concordance between human and AI classifications. The aim was to demonstrate the value of combining human and AI image labels for increasing the efficiency and/or accuracy of classifications. Registered citizen scientists provided more accurate classifications than anonymous participants, but when classifications from either source were considered in combination with AI, image labels were consistently highly accurate. There was however, a trade-off, with high accuracy coming at the expense of increasing numbers of image sequences for which labels from different sources were not concordant. Secondly, I used adapted camera traps to survey small mammals in the Northeast of England. By varying bait treatments between camera traps, I investigated the optimal bait strategy to refine this method and encourage future uptake. I found no significant differences in the probability of detection of small mammal species by bait type, but the inclusion of mealworms was related to higher numbers of captures of shrew species. I found the non-native greater white-toothed shrew Crocidura russula to be present at a much lower proportion of sites than native small mammal species. My findings offer insight into current topics in camera trapping research by highlighting the value of integrating AI and human contributions in classification workflows, and by demonstrating the success of adapted camera traps for expanding the taxonomic breadth of species that can be observed with this method.