4:15 PM - 4:30 PM
[U02-03] Deep learning-based multi-label classification for monitoring multi-temporal coral conditions in the Indo-Pacific
Keywords:Coral conservation, Ecological monitoring, Ensemble learning, Neural networks, Swin transformer, Underwater photogrammetry
Coral reef ecosystems are crucial as they support diverse marine life and provide ecosystem services to humans. However, they are currently under threat from anthropogenic activities and climate change. Citizen science-based conservation activities are being conducted globally to protect and restore coral reefs. Regular monitoring of coral conditions is essential for understanding the status of coral and directing conservation actions. Automated classification using deep learning has shown great potential in assisting ecologists and conservationists in tracking and identifying threatened coral colonies. However, this classification task is highly challenging due to the ecological diversity of coral stressors, seasonal dynamics in coral reef ecosystems, the complexity of underwater imaging conditions, inaccurate annotations, and the use of outdated algorithms and datasets.
To address these challenges in coral image classification and contribute to solving real-world problems, we conducted this study in collaboration with a non-profit organization in Koh Tao, Thailand, which has been engaged in continuous monitoring and coral conservation activities for 16 years. This study comprises four major components: (1) Construct an all-season, updated, and open-source dataset using underwater photogrammetry, containing over 30,000 high-resolution samples representing common stressors and Indo-Pacific coral conditions. (2) Compare and evaluate the performance of seven off-the-shelf representative deep learning architectures, including the recently emerging Transformer architectures, for this multi-label classification task. (3) Introduce a new multi-label classification method for detecting coral conditions and extracting ecological information, utilizing the ensemble learning approach. (4) Test the model’s generalizability using wet season data. Apply transfer learning to enhance the model's capability in classifying multi-seasonal images. This proposed method can accurately classify coral conditions across all seasons as healthy, compromised, dead, and rubble; it also identifies corresponding stressors, including competition, disease, predation, and physical damage. The classification results demonstrate that our approach, which integrates Swin Transformer and EfficientNet through the ensemble learning strategy, has achieved state-of-the-art performance compared to other methods tested on the dataset. These findings can support conservation activities and provide references for decision-making for reef managers and stakeholders.
To address these challenges in coral image classification and contribute to solving real-world problems, we conducted this study in collaboration with a non-profit organization in Koh Tao, Thailand, which has been engaged in continuous monitoring and coral conservation activities for 16 years. This study comprises four major components: (1) Construct an all-season, updated, and open-source dataset using underwater photogrammetry, containing over 30,000 high-resolution samples representing common stressors and Indo-Pacific coral conditions. (2) Compare and evaluate the performance of seven off-the-shelf representative deep learning architectures, including the recently emerging Transformer architectures, for this multi-label classification task. (3) Introduce a new multi-label classification method for detecting coral conditions and extracting ecological information, utilizing the ensemble learning approach. (4) Test the model’s generalizability using wet season data. Apply transfer learning to enhance the model's capability in classifying multi-seasonal images. This proposed method can accurately classify coral conditions across all seasons as healthy, compromised, dead, and rubble; it also identifies corresponding stressors, including competition, disease, predation, and physical damage. The classification results demonstrate that our approach, which integrates Swin Transformer and EfficientNet through the ensemble learning strategy, has achieved state-of-the-art performance compared to other methods tested on the dataset. These findings can support conservation activities and provide references for decision-making for reef managers and stakeholders.