DichroGAN: Towards Restoration of in-air Colours of Seafloor from Satellite Imagery

TL;DR

This work tackles recovering the in-air colours of seafloor from satellite imagery, addressing atmospheric and water-column distortions that obscure true albedo. It introduces DichroGAN, a four-generator conditional GAN that jointly estimates diffuse and specular reflections ($G_d$, $G_s$), derives scene radiance through $G_j$, and computes water transmission via $G_t$ within a unified training framework rooted in the underwater image formation model (UIFM). A Grey World illuminant estimate from a 63-band hyperspectral cube guides radiance decomposition, and the model is trained end-to-end with a composite loss that blends adversarial and physics-based terms. Across PRISMA and NASA EO datasets, as well as underwater benchmarks, DichroGAN delivers competitive or superior restoration quality, effectively removing water-column effects and recovering the seafloor’s in-air colours with clear qualitative improvements, albeit with some tradeoffs in standard underwater quality metrics and texture sharpness.

Abstract

Recovering the in-air colours of seafloor from satellite imagery is a challenging task due to the exponential attenuation of light with depth in the water column. In this study, we present DichroGAN, a conditional generative adversarial network (cGAN) designed for this purpose. DichroGAN employs a two-steps simultaneous training: first, two generators utilise a hyperspectral image cube to estimate diffuse and specular reflections, thereby obtaining atmospheric scene radiance. Next, a third generator receives as input the generated scene radiance containing the features of each spectral band, while a fourth generator estimates the underwater light transmission. These generators work together to remove the effects of light absorption and scattering, restoring the in-air colours of seafloor based on the underwater image formation equation. DichroGAN is trained on a compact dataset derived from PRISMA satellite imagery, comprising RGB images paired with their corresponding spectral bands and masks. Extensive experiments on both satellite and underwater datasets demonstrate that DichroGAN achieves competitive performance compared to state-of-the-art underwater restoration techniques.

Figures (6)

• Figure 1: Overview of DichroGAN. It comprises 4 generators and 1 discriminator. Satellite scene radiance is obtained by summing diffuse and specular reflections generated by $G_{d}$ and $G_{s}$, respectively. Generated radiance serves as input for $G_{j}$, while $G_{t}$ generates depth map to estimate light transmission. $G_{j}$ and $G_{t}$ compute UIFM Duntley:1963 to remove water column and recover in-air colours. Discriminator $D$ classifies between generated and real images.
• Figure 2: Sample results of generated diffuse and specular reflections on PRISMA test dataset.
• Figure 3: Sample results from ablation on NASA EO dataset. From left-to-right: Lake Mead satellite images from 2000 (input image showing high water levels) and 2022 (ground truth showing dry regions). Subsequent columns show results of our ablation experiments.
• Figure 4: Sample results on Lake Meade images from NASA EO.
• Figure 5: Sample results on PRISMA and NASA EO.