Coral classification in underwater images using a dual-branch deep learning framework

Abstract

Automated coral classification from underwater imagery is essential for large-scale reef monitoring but remains challenging due to color attenuation, illumination variability, and differences between texture-focused close- range images and morphology-focused colony-level observations. To address these challenges, this study proposes a dual-branch convolutional neural network that integrates information from the CIELAB (LAB) color space with structural descriptors derived from the discrete wavelet transform (DWT). RGB images are first converted to the LAB color space to separate luminance and chromatic components in separate channels, enabling the model to exploit color and lightness information more explicitly. Structural information is extracted from the luminance channel using wavelet decomposition to capture high-frequency morphological patterns. The two representations are processed through parallel convolutional branches and fused at the feature level for classification. Experiments conducted on a unified coral dataset containing texture dominant and morphology-focused imagery across 14 classes show that the proposed method achieves 96.52% accuracy on the texture-focused RSMAS dataset and 88.33% accuracy on the morphology-focused structure RSMAS dataset, reducing the cross-domain performance gap from 22.46% to 8.19% compared with RGB baselines, demonstrating improved cross domain robustness for coral classification under heterogeneous underwater imaging conditions.