Enhancing Bird Species Classification: A Weighted Average Transfer Learning Ensemble Approach

With the escalating threat to avian life due to environmental degradation, the need for an automatic bird image recognition system has become paramount. The diversity among the estimated 11,000 bird species worldwide presents a challenge, demanding advanced technological solutions for their conservation. The proposal utilizes a custom dataset comprising 525 species to formulate a distinctive ensemble deep-learning model for categorizing bird species. The weighted average ensemble (WAE) technique, specifically combining EfficientNetB5 and VGG19, selected from a pool of five deep learning models, including MobileNetV2, ResNet152, and InceptionV3, attains an impressive 97.18 % accuracy. Using data augmentation approaches, the training data becomes more diverse, strengthening the model. The evaluation metrics extend beyond accuracy, including F1 Scores, comprehensively assessing the model's output quality.