A Scalable AI Approach to Bird Species Identification for Conservation and Ecological Planning
Souad Hassanie, Aiman Gohar, Raja Hashim Ali, Talha Ali Khan, Iftikhar Ahmad, Shan Faiz
Abstract
Birds represent an important component of ecosystems, holding substantial ecological and human interest, and significantly contributing to ecological equilibrium via their roles as pollinators, seed dispensers, and predators. If identified correctly, the wild life departments, farmers, ecologists, and various other experts can ensure saving endangered bird species, improving crop yield and ensure seed safety, and plan safe sanctuaries for birds, etc. Due to the vast array of bird species and the often subtle distinctions among them, human-driven manual classification poses a formidable challenge. The manual classification of bird species is a labor-intensive and error-prone process, predominantly reliant on domain experts. However recent algorithmic developments and progress in the deep-learning based object detection and classification models has made this difficult and time-consuming task easy and accurate. Nevertheless, numerous existing CNN models have been trained or fine-tuned on limited datasets, rendering them ill-equipped to handle the classification of a substantial volume of diverse bird species. To address this challenge, this study introduces a novel approach based on transfer learning and the ResNet-50 model to ensure accurate identification of endangered species. Our methodology involved fine-tuning the ResNet-50 model on the 100 Bird Species dataset, with the training data consisting of 85,000 images, representing 525 distinct bird species with each species comprising of 130 or more diverse pre-processed images, including 25 endangered and 12 critically endangered species. The transfer learning based mechanism applied on the large dataset achieved a state-of-the-art accuracy of 87% in bird species classification through deep learning. Our results showed that out of 12 critically endangered species like California condor, Ashy storm petrel, California quail and other quails, none is misclassified by our system. Similarly, none of the 25 endangered species is misclassified by our system, indicating a solid performance for use in identification of endangered and critically endangered bird species. This study shows that deep learning and transfer learning can be successfully deployed to identify birds, and ensure the safety of endangered and critically endangered species of birds, making it a powerful and adaptable approach for the future.