Research overview: training-data bias and domain shift in camera-trap machine learning

Machine-learning classifiers trained on camera-trap images often perform poorly when applied to new locations, a failure rooted in training-data bias and domain shift, the mismatch between the backgrounds, species mixes, and conditions of training versus deployment. This overview surveys why models that excel on familiar cameras stumble on unfamiliar ones, learning spurious correlations with specific backgrounds rather than the animals themselves, and how imbalanced datasets dominated by common species and empty frames compound the problem. It discusses diagnosis through location-stratified evaluation and mitigation via diverse training data, object detection that crops animals from backgrounds, data augmentation, and domain-adaptation techniques. The overview underscores a central lesson for camera-trap AI: reported accuracy on held-out images from the same project overstates real-world performance, and honest evaluation must test generalization to genuinely new sites.