Author(s):
Mir Ihrar Ali

Abstract:
The conservation of marine biodiversity has become. increasingly critical in the face of climate change and environ- mental degradation. Identifying individual animals is essential for understanding species behavior, migration patterns, and population dynamics, yet traditional methods remain time-intensive and reliant on expert observation. This study leverages deep learning to automate individual whale and dolphin identification using data from the Kaggle Happywhale challenge. The performance of several Convolutional Neural Network (CNN) architectures is evaluated, including ResNet, DenseNet, EfficientNet, and In- ceptionV3, and explore the combination of Softmax classification with a semi-hard triplet-loss approach. The results reveal that InceptionV3 achieves superior accuracy, while the hybrid Softmax and triplet-loss method offers limited benefits, hindered by dataset challenges such as the prevalence of hard triplets. These findings emphasize the need for refined loss strategies to handle unbalanced datasets in wildlife identification. This work highlights the potential of machine learning to revolutionize marine conservation efforts by enabling scalable and efficient individual recognition systems.

Pages: 964-976

Read Full Article