Résumé
1. Accurate estimation of individual ages is crucial for studies in ecology, behaviour and conservation. However, when birth dates are unknown, estimating chronological ages often relies on post-mortem morphological analyses or invasive and cumbersome techniques. Here we investigate the potential of deep learning applied to photographic portraits for non-invasive chronological age prediction. 2. Comparing the predictive capabilities of several recent deep learning models with 25,500 portraits of wild mandrills collected on 284 individuals of known ages in situ, we show that the foundational transformer models DINOv2 largely outperformed convolutional networks (notably ResNext, ConvNeXt, EfficientNetv2) and the other popular transformer model VOLO.
3. To gain insight into the model's predictions, we first examine the influence of the background. Although the model relied on background information for its predictions, this did not lead to a significant improvement in overall accuracy: there was no meaningful difference between predictions when age estimates were from images with or without background. Second, we show that inter-individual variation in prediction errors is partly explained by biological factors. At the individual scale, the prediction error was consistent through time: when individuals appeared older than their chronological age when young, they also consistently appeared older throughout their life. In addition, we found that offspring of older mothers appeared older compared to those of younger mothers, consistent with previous findings on the link between offspring development and maternal age in this species. 4. Altogether, these results indicate that the most modern artificial intelligence methods offer a simple, low-cost and non-invasive approach for chronological age estimation and that the difference between chronological and estimated ages could be used by behavioural ecologists to study individual growth, pace of development and biological aging processes.