Essential tools but overlooked bias: artificial intelligence and citizen science
classification affect camera trap data
Simone Santoro, Santiago Gutiérrez-Zapata, Javier Calzada, Nuria Selva, Diego
Marín-Santos, Sara Beery, Kate Brandis, Paul Meek, Alessio Mortelliti, Eloy Revilla,
Jon Paul Rodríguez, Simone Tenan, Manuel Emilio Gegúndez. Methods in Ecology
and Evolution, Submitted

Keywords: artificial intelligence, deep learning, convolutional neuronal networks,
citizen science, camera trap, computer vision

Abstract: 1. Camera trapping, widely employed for wildlife monitoring, generates avast volume of images which must be classified to become analysable data. Yet, the factors influencing classification performance and their effects on ecological interpretation are often overlooked. The variability in classification performance may stem from different image classification methods, animal species, camera specifications, and environmental covariates.

2. This study assesses the performance of a citizen science project hosted in Zooniverse (CS-Zoo) and two artificial intelligence (AI) approaches in identifying diurnal and nocturnal images across six classes: empty, human, cervid, leporid, wild boar, and red fox. Using an out-of-sample (not used for AI training and validation) dataset of 51,588 images classified by experts, we evaluated for each class the accuracy of positive predictions (precision) and the ability to detect all positive instances (recall). Additionally, we conducted a single-season occupancy analysis to compare the
ecological inference derived from each classification system and class.

3. Citizen scientists demonstrated exceptional precision, though recall was more variable. AI outperformed citizen science in recall for some classes, such as leporid. However, both encountered challenges with nocturnal photos and experienced a significant loss of precision for empty images at night. AI data analyses produced positively biased occupancy predictions, whereas CS-Zoo data generated a slight negative bias, with both biases being noticeable only for taxa with low-to-moderate occupancy estimates, such as red fox and leporid. Notably, our AI model processed up to 1,150,000 images daily, over 500 times faster than CS-Zoo.

4. Our findings suggest that while CS-Zoo is preferable for minimizing classification errors, AI’s remarkable processing capacity is indispensable for studies needing rapid and large-scale data processing. However, the significant drop in performance for night photos and less common species in AI and CS warrants careful consideration, particularly regarding missed detections of nocturnal or conservation concern species. A regular comprehensive evaluation is crucial for understanding how classification performance impacts data and ecological inferences, including demographic and behavioural parameters like occupancy, abundance, community composition, and activity patterns. These inferences rely on the inherent classification process in camera
trap data studies.