COMPARATIVE EVALUATION OF DEEP LEARNING ARCHITECTURES FOR CORAL BLEACHING DETECTION USING IMAGE RECOGNITION
Abstract
Coral reefs are vital marine ecosystems supporting biodiversity and coastal communities but face threats from climate-induced coral bleaching. Traditional assessment methods rely on manual visual surveys that are labor-intensive, time-consuming, and difficult to scale. This study presents Smart Coral Diagnostics, an AI-powered web system utilizing Convolutional Neural Networks to automate bleaching detection from underwater images. A mixed-method approach integrates development with expert evaluation, comparing architectures like GoogleNet, EfficientNet, MobileNet, ResNet50, and YOLOv8. Results show GoogleNet achieved the highest classification performance at 81.82% accuracy, surpassing EfficientNet's 79.74%. Implemented via Next.js and FastAPI, the system features automated image analysis, geotagging, and map-based visualization. Evaluation with Coastal Conservation and Education Foundation experts indicates current practices face severe resource constraints. While noting accuracy limitations, experts reported improved time efficiency and reduced labor. Ultimately, the system offers a scalable framework to enhance marine conservation efforts..
