Explainable Deep Learning Framework for Coral Reef Health Detection Using Efficientnet and Grad-CAM

Abstract

Coral reefs are critical marine ecosystems that support biodiversity, coastal protection, and fisheries, but they are increasingly threatened by climate change, ocean acidification, pollution, and coral bleaching. Traditional reef monitoring methods rely on manual underwater surveys, which are labor-intensive, time-consuming, and limited in spatial coverage. To address these challenges, this work proposes an EfficientNet-based deep learning framework for automated coral reef health detection using underwater imagery. The dataset used for this work consists of 120 coral reef images, categorized into healthy and bleached coral classes, and was preprocessed through image resizing, normalization, and augmentation to improve model robustness. The key novelty of this work lies in integrating EfficientNet architecture with Test Time Augmentation (TTA) to improve prediction stability and enhance classification performance under challenging underwater conditions such as color distortion, scattering, and varying illumination. Grad-CAM-based explainability is incorporated to provide visual interpretation of the regions influencing model decisions, improving transparency and trust in the detection process. Experimental evaluation demonstrates that the proposed model achieved an overall classification accuracy of 78.85%, with a macro-average precision of 0.79, recall of 0.79, and F1-score of 0.79. Class-wise analysis showed a precision of 0.78 and recall of 0.75 for bleached coral, and a precision of 0.79 and recall of 0.82 for healthy coral. The proposed work provides a scalable, cost-effective, and interpretable solution for coral reef monitoring and can support marine conservation authorities in early reef degradation assessment and sustainable ecosystem management.