Deep Learning-Driven Super-Resolution and DenseCLIP-Based Semantic Segmentation for Coral Reef Mapping

Coral reefs are essential to marine ecosystems yet monitoring and conservation are hindered by the difficulty of obtaining high-quality images for assessing coral health. Traditional methods struggle with low-resolution underwater images, limiting detailed analysis. This study proposes a novel approach combining deep learning-based Super-Resolution Reconstruction (SRR) and DenseCLIP-based semantic segmentation for coral reef mapping. Seafloor images were captured using a self-developed Portable Speedy Sea Scanner (P-SSS) and processed to enhance image quality. The Residual Dense Network (RDN) SRR model improved image resolution, while DenseCLIP enabled pixel-level segmentation of coral features. Among the evaluated models, the DANet R101 fusion model achieved the highest segmentation accuracy. This approach integrates advanced deep learning with portable systems, offering a scalable, cost-effective solution for coral reef monitoring and conservation.