AI-Assisted Species Recognition from Camera Traps

Abstract

Camera traps have become the dominant tool for non-invasive wildlife monitoring globally, generating tens of millions of images annually from monitoring networks spanning protected areas, connectivity corridors, and human-wildlife interface zones. The principal bottleneck limiting the ecological value of camera trap data has shifted from image acquisition to image classification: manually reviewing and annotating millions of images for species, count, and behaviour is a task requiring thousands of person-hours per study, delaying data delivery from months to years after collection. Deep learning-based automated species recognition has emerged as the most promising solution, but published models vary enormously in training dataset size, geographic scope, species coverage, and classification accuracy -- making it difficult for practitioners to select or deploy appropriate models for their monitoring contexts. This study presents the most comprehensive multi-model, multi-region benchmark of AI species recognition from camera traps yet conducted, evaluating 8 published deep learning models against a standardised test dataset of 247,000 images from 84 species across 18 countries. The best-performing model (MegaDetector + EfficientNet-B4 transfer-learned; 847,000-image training set) achieved mean species-level F1 = 0.91 across 84 species -- above the estimated human expert accuracy baseline (F1 = 0.89 from blind expert review of the same test set). Transfer learning with as few as 200 images per species improved F1 by 47.4% over zero-shot deployment of general models for the 24 regional specialist species not in the general training datasets.