Enhancing Coral Health Monitoring with a Hybrid CNN-ViT Model for Bleaching Prediction

This research develops a hybrid deep learning approach that integrates Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) to automate coral reef health assessments. Using a dataset of 923 labeled underwater images (categorized as "healthy" or "bleached"), the model leverages CNNs to extract fine-grained local features while ViTs capture large-scale global dependencies for enhanced classification accuracy. To improve generalization, advanced data augmentation techniques are applied during training, and hyperparameters such as batch size, epochs, and input dimensionality are fine-tuned. The model achieves an accuracy of 83.78%, evaluated using a confusion matrix, classification report, and accuracy metrics. Designed for real-time monitoring, this scalable and efficient hybrid architecture eliminates the need for manual reef assessments, providing a practical, data-driven solution for conservation efforts. This approach enables timely decision-making to protect and sustain coral ecosystems increasingly threatened by environmental changes.