Automated goat breed identification using CNN-based transfer learning with VGG16 and mobilenetv2

Abstract

Effective management of livestock genetic resources and the implementation of breeding programmes require accurate identification of indigenous breeds. However, under field conditions, goat breed identification is often challenging due to phenotypic similarities among breeds and reliance on subjective visual assessment. Therefore, the present study aimed to develop an automated image-based deep learning framework using transfer learning for the classification of two morphologically related indigenous goat breeds, Barbari and Jamunapari, employing convolutional neural network (CNN) architectures. A curated dataset consisting of 640 images collected from 160 goats (80 animals per breed) was developed, where four images per animal were captured: two frontal and two left-lateral views. The dataset was partitioned at the animal level into training, validation and testing subsets using a 70:10:20 split (training:validation:testing) to ensure independent evaluation. Two transfer learning architectures, VGG16 and MobileNetV2, pretrained on ImageNet, were fine-tuned for binary breed classification. Model performance was evaluated using confusion matrix–derived metrics including accuracy, precision, recall, F1-score and AUC–ROC, along with computational efficiency parameters such as inference time and memory usage. On the independent test set, MobileNetV2 achieved the highest classification accuracy (94.53%), with precision of 93.85%, recall of 95.31% and F1-score of 94.57%, whereas VGG16 achieved an accuracy of 93.75%, with precision of 95.16%, recall of 92.19% and F1-score of 93.65%. ROC analysis indicated strong discriminative capability for both models, with AUC–ROC values of 0.94 for MobileNetV2 and 0.95 for VGG16. Overall, the results demonstrate that deep learning–based image classification can effectively distinguish closely related goat breeds and provide a scalable tool for automated breed identification.