Abstract
Accurate real-time coral reef bleaching classification is challenging due to the high computational cost and limited interpretability of conventional convolutional neural networks (CNNs), which restrict deployment on resource-constrained edge devices. To address this, we propose Fibonacci-Net (F-Net), a lightweight and interpretable CNN that integrates Fibonacci-based filter scaling, a patch-based hybrid area-attention mechanism to enhance fine-grained coral features, and a particle swarm optimization–Adam hybrid optimizer for stable learning on small, imbalanced datasets. Evaluated on 7384 coral images, F-Net achieves 97.6% accuracy, better than some well-studied CNN models in the literature. The novel gradient-weighted class activation mapping and filter discriminability analyses further enhance interpretability, demonstrating F-Net's effectiveness and deployment readiness for large-scale autonomous coral reef monitoring.
Showing the abstract — retrieve the full paper via the Exa API.