Deep learning based individual identification and population estimation of the yellow spotted mountain newt...

The yellow-spotted mountain newt (Neurergus derjugini), an endangered amphibian endemic to the Zagros Mountains, faces critical threats from habitat loss and climate change. Effective conservation requires reliable population monitoring, yet traditional marking methods are invasive and impractical. This study presents a non-invasive, image-based approach combining geometric computer vision and deep learning for individual identification and population estimation. We captured 549 adult N. derjugini in their natural habitat, photographing dorsal patterns under standardized conditions. A geometric pipeline (HSV thresholding, morphological operations) extracted yellow spot features (area, circularity, count), achieving 93% detection accuracy. Three convolutional neural networks (CNNs)-DenseNet121, EfficientNetB0, and InceptionV3-were fine-tuned for phenotypic classification, with DenseNet121 attaining the highest accuracy (99.11%) and AUC (0.98). Region-specific analysis showed optimal performance when combining head and trunk patterns (96.32% accuracy). A mark-recapture framework, applied to two sampling sessions (n = 332 and 217 individuals), identified 65 recaptures, yielding a Lincoln-Petersen population estimate of 1108 individuals. Our results demonstrate that deep learning outperforms traditional methods in robustness and scalability, particularly under variable field conditions. This study advances amphibian conservation by providing a rapid, ethical, and scalable tool for monitoring endangered species. Future directions include expanding datasets for temporal stability validation and deploying mobile applications for real-time field use. By integrating AI with ecological research, this work highlights the transformative potential of automated identification in biodiversity conservation.