<span lang="EN-ID">Application of convolutional neural network architectures for coral <span lang="EN-ID">reef health assessment

BACKGROUND AND OBJECTIVES: Coral reefs provide vital ecological and economic benefits but are increasingly threatened by climate change, pollution, and anthropogenic stressors that drive bleaching and degradation. Effective monitoring is essential, yet conventional field assessments remain labor-intensive, time-consuming, and expert-dependent. The study objectives were to develop a convolutional neural network-based framework for automated classification of reef coral morphotypes and coral health status from underwater photographs; evaluate three transfer learning architectures, dense convolutional network with ۱۶۹ layers, residual network with ۱۵۲ layers, and visual geometry group network with ۱۹ layers, under limited and imbalanced training data; and deploy the best-performing model for automated reef health monitoring across five stations at Kelapa Dua Island, Indonesia.METHODS: Reef coral morphotypes were categorized as boulder, branching, and table, while health status was classified as healthy, bleached, or partially bleached using the Coral Health Chart as a standardized reference. Three convolutional neural networks architectures pre-trained on ImageNet were fine-tuned on datasets of ۶۶۳ morphotype images and ۶۲۴ health status images, with class-weighted loss applied to address pronounced imbalance in the health dataset. FINDINGS: dense convolutional network with ۱۶۹ layers achieved the highest overall accuracy, reaching ۹۳ percent for morphotype classification and ۸۲ percent for health status assessment, substantially outperforming residual network with ۱۵۲ layers ۱۵۲ (۶۶ percent and ۶۴ percent) and Visual geometry group network with ۱۹ layers (۴۲ percent and ۶۴ percent). Per-class analysis revealed that the health classification task is significantly affected by class imbalance, with the partially bleached category dominating the dataset (۵۵ percent of health samples); the F۱-score for the healthy class reached only ۰.۴۴, highlighting systematic model bias. CONCLUSION: These findings demonstrate both the promise and the current limitations of deep learning for automated coral reef monitoring, and underscore the need for larger, class-balanced, and station-independent datasets for future model development. The dense convolutional network with ۱۶۹ model was deployed successfully at five monitoring stations at Kelapa Dua Island, where partially bleached coral conditions predominated, consistent with elevated sea surface temperatures and suppressed salinity at the time of the survey.

کلیدواژه ها:

Convolutional neural network (CNN) ، Coral health status ، DenseNet۱۶۹ ، Kelapa Dua Island ، Reef coral morphotype ، Transfer learning.