Non-invasive individual identification of adult Pleurodeles waltl using Deep Learning.

Reliable individual identification is essential for long-term tracking and reproducible laboratory animal studies. In amphibians, invasive marking methods like tags and dye injections can cause welfare concerns and may be unreliable because of tag loss or migration. We developed and evaluated a non-invasive identification system for 25 adult Pleurodeles waltl using smartphone-captured images and the pre-trained convolutional neural network EfficientNetV2. To determine the most informative imaging region, separate models were trained and tested using head and whole-body images. The head-image model achieved 95.3% accuracy on the independent test dataset (macro F1-score = 0.946; Cohen's kappa = 0.951), markedly outperforming the whole-body model (56.6% accuracy). Grad-CAM visualization showed that the model primarily focused on dorsal head spot patterns, indicating that these markings are more informative for individual recognition than whole-body patterns. Because this method requires only a smartphone and a trained model, it can be implemented without specialized marking equipment. This approach enables accurate individual identification, while avoiding invasive marking and therefore supports both animal welfare and research reproducibility. It may provide a practical basis for standardizing individual identification in future amphibian research.