Accurate individual identification is vital in field studies. Since traditional marking techniques, though effective, can be intrusive and potentially disrupt natural behaviours, identification using natural markings has gained popularity across various taxa as a non-invasive alternative. Here, we report on a Painted Stork (Mycteria leucocephala) with a distinctive neck injury mark, observed at the National Zoological Park (Delhi Zoo) over three consecutive breeding seasons (2022–2024). To verify its identity and assess nest-site fidelity, we employed a non-invasive approach combining morphometric measurements and Deep Transfer Learning-based image analysis. High-resolution photographs were used to extract linear measurements and assess repeatability, while a Deep Transfer Learning classifier further validated the individual’s identity with 98% accuracy. Image-based morphometric measurements were particularly reliable for longer morphological features, confirming that the scar-marked stork observed over three consecutive years is indeed the same individual. The repeated sightings of the scar-marked stork on the same patch support evidence of nest-site fidelity. Our findings highlight the potential of Deep Transfer Learning and pattern-based recognition as powerful, non-intrusive tools for long-term monitoring of colonial waterbirds.

1.     Morphological measurements of Ringo

We used 82 photographs of Ringo from a collection of 100 that had proper orientation. The mean of the morphological variables of Ringo taken during 2024 was compared with randomly selected images of 2022 and 2023. With tps DIG we carefully measured morphological features in pixel units, namely (1) Upper mandible, (2) Lower mandible, (3) the distance between the mouth corner to the inside corner of the eye, (4) the distance between the inside corner of the eye to the top end of the right nostril, (5) The distance from the base of the pectoral wing to the tip of the wing covering black and white, (6) distance from the feather region to the tarsus and tibia junction. All the above morphological measurements are available in the Excel format.

2.     Deep Transfer Learning

Using the pre-trained model (ResNet-50), we fine-tuned our model with Ringo datasets collected during 2022, 2023, and 2024 to train, validate, and test the accuracy. Our input data size is 224 by 224 pixels, and output with multiclass layers (n=3). Three classes, namely 1) the right side of the Ringo (n = 824 images), 2) the left side of the Ringo (n = 678 images), and 3) non-marked or unknown Painted Storks (n = 1755 images) were created. Using a Transfer Learning approach, we developed a model trained with the digital images of Ringo and other Painted Storks.

We have provided the Deep Transfer Learning model and its MATLAB code along with few sample images to cross check and validate the model. To run this model, one requires a MATLAB environment.

We also provided the layer details of the model.