Learning-based underwater color correction using reference targets and image enhancement techniques

Abstract

Color-corrected underwater coral reef images enable a physically meaningful assessment of coral health and monitoring of reef changes over time. This study presents a color-correction pipeline for burst images acquired along dive transects using reference targets. A shallow neural network is used to map underwater to above-water color responses. Further enhancement is applied using Dark Channel Prior (DCP) for dehazing and Red Channel Compensation (RC) to correct wavelength-dependent color loss. The proposed method is compared with Gray World (GW) and White Patch Retinex (WPR) and shows better overall color balance in underwater images. Validation using CIEDE2000 (ΔE₀₀) metrics on sand patches from the corrected images shows lower mean and median errors, as well as reduced variability, compared to the baseline methods. There results suggest that the proposed method is a practical and reliable solution for improving underwater image color fidelity in coral reef monitoring applications.