Synthetic Imagery Improves Ecological Classification When Real Data Are Scarce, Not Direct Substitutes

Abstract

ABSTRACT The growing availability of ecological imagery creates new opportunities to monitor biodiversity and ecosystem change at broad scales. Yet image classification models remain constrained by uneven sampling, limited labeled data, and differences between training and deployment imagery. Synthetic images generated with generative methods may help address these limitations, but their value across image domains remains poorly tested. We evaluated whether synthetic imagery improves genus‐level classification for 20 North American tree genera using iNaturalist community‐science photographs and Auto Arborist street‐view images. We generated domain‐specific synthetic images with Stable Diffusion v1.5 and trained ResNet‐50 classifiers on controlled mixtures of real and synthetic data. We tested synthetic replacement, supplementation under data scarcity, conventional augmentation alternatives, and transfer between image domains. Replacing real images with synthetic images reduced performance on held‐out real test sets, although declines differed between domains. Adding synthetic images improved classification when real classifier‐training data were scarce, but gains diminished as real data increased. In cross‐domain tests, synthetic supplementation improved some scarce‐data baselines but did not close the transfer gap or replace real‐image training sets. Synthetic images were therefore most useful as supplements rather than direct substitutes. For ecological applications, generated images may support model development for poorly sampled taxa, but representative real images remain essential for robust model generalization.