Increasing citizen scientist accuracy with artificial intelligence on <scp>UK</scp> camera‐trap data

Abstract As camera traps have become more widely used, extracting information from images at the pace they are acquired has become challenging, resulting in backlogs that delay the communication of results and the use of data for conservation and management. To ameliorate this, artificial intelligence (AI), crowdsourcing to citizen scientists and combined approaches have surfaced as solutions. Using data from the UK mammal monitoring initiative MammalWeb, we assess the accuracies of classifications from registered citizen scientists, anonymous participants and a convolutional neural network (CNN). The engagement of anonymous volunteers was facilitated by the strategic placement of MammalWeb interfaces in a natural history museum with high footfall related to the ‘Dippy on Tour’ exhibition. The accuracy of anonymous volunteer classifications gathered through public interfaces has not been reported previously, and here we consider this form of citizen science in the context of alternative forms of data acquisition. While AI models have performed well at species identification in bespoke settings, here we report model performance on a dataset for which the model in question was not explicitly trained. We also consider combining AI output with that of human volunteers to demonstrate combined workflows that produce high accuracy predictions. We find the consensus of registered users has greater overall accuracy (97%) than the consensus from anonymous contributors (71%); AI accuracy lies in between (78%). A combined approach between registered citizen scientists and AI output provides an overall accuracy of 96%. Further, when the contributions of anonymous citizen scientists are concordant with AI output, 98% accuracy can be achieved. The generality of this last finding merits further investigation, given the potential to gather classifications much more rapidly if public displays are placed in areas of high footfall. We suggest that combined approaches to image classification are optimal when the minimisation of classification errors is desired.