AI for Sustainable Wind Energy: Enhancing Bird Classification with Supervised Contrastive Learning

Wind energy is a fundamental component in the shift towards clean and renewable energy sources and is essential for achieving Sustainable Development Goal 7 [1]. Its sustainable and responsible implementation still requires research, in particular its impact on birds, which are most affected by wind power plants. Birds play a big role in different ecosystems as they are very mobile species, and studying their migration behaviors helps track ecosystems and climate change dynamics. This thesis investigates the potential Artificial Intelligence techniques for bird species classification. We compare the performance of Supervised Contrastive Learning against models trained with Self-Supervised Contrastive Learning and fully supervised models. Our experiments tested the impact of different sets of augmentations, training durations and temperatures on the performance of the models in both pre-training and transfer tasks. Our results show that Supervised Contrastive Learning performs significantly better than Self-Supervised Contrastive Learning, in both pre-training and transfer tasks. For transfer tasks a careful balance between temperature and augmentations is crucial for better performance. Although larger models are needed to fully harness the potential of Supervised Contrastive Learning, the results obtained with ResNet-18 models promise better results with larger architectures.