An integrated deep learning framework for automated coral health assessment using a custom-annotated database.

Coral reefs face accelerating threats from climate-induced bleaching events, underscoring the need for scalable and automated health monitoring tools. This paper presents a comprehensive deep learning framework integrating object detection and image classification with a rigorously evaluated preprocessing pipeline for automated coral health assessment. Building on the custom annotated coral bleaching dataset introduced in Weraikat et al. [1], comprising 2,640 underwater coral images, four preprocessing strategies, RGB normalization, HSV conversion, histogram equalization, and color jittering, were comparatively evaluated using both image-fidelity metrics and downstream deep learning performance. Experimental results demonstrated that HSV preprocessing achieved the strongest downstream detection and classification performance while preserving high structural image fidelity, enhancing the discrimination between healthy and bleached coral tissue. For object detection, YOLOv8 substantially outperformed YOLOv4, achieving a mAP@50 of 0.986 and recall of 0.938. For coral health classification, ResNet50 achieved the highest test accuracy (92.51%) and F1-score (0.93), outperforming VGG16, VGG19, and EfficientNet architectures. The proposed two-stage YOLOv8-ResNet50 framework provides a reproducible and scalable solution for automated coral health monitoring with strong potential for integration into marine conservation and environmental monitoring systems.