Deep Learning Analysis Using CNN For Bird Species Classification System

For the past few years, the rapid extinction of many bird species due to climate change and habitat destruction caused by human activities is a growing concern. Accurately tracking the distribution of these species and understanding the factors that contribute to a region's biodiversity is essential for creating effective conservation plans. However, the task of identifying bird species from images is challenging and requires a significant amount of time and effort. Deep learning-based methods are superior at detecting and localizing birds in images because they can handle situations when the birds usually portray of various forms, sizes, and backgrounds. This project proposes an implemented training and validation model system for bird species classification based on performance of the four Keras deep learning CNN model on two different datasets based on the evaluation in term of accuracy, loss function, and confusion matrix. The proposed model is trained using the two set of datasets of bird images which consist of 1,494 and 2,353 data training sets for 15 rare bird species which can be found in Malaysia and worldwide. The process of training the model was carried out using ResNet152V2, InceptionV3, DenseNet121 and DenseNet169 models. Several repeated tests were conducted repeatedly using Python on Google Collab with Tesla T4 GPU. Results showed that DenseNet169 had the best performance for both datasets, with 95% accuracy for dataset 2 and 85% training and 86% validation accuracy for dataset 1. This system will be beneficial for ornithologists and researchers in tracking rare bird species for ecosystem preservation.