From Aqua MODIS to PACE OCI: Global-Scale Hyperspectral R rs Reconstruction via Spectral-Attention Conditional GAN

Hyperspectral remote sensing reflectance ( R rs ) is vital for ocean color research. NASA’s Plankton, Aerosol, Cloud, ocean Ecosystem (PACE) mission provides unprecedented hyperspectral coverage, but its short record limits long-term applications. To overcome this, this study introduces SA-CGAN, a spectral-attention conditional generative adversarial network, developed to reconstruct hyperspectral R rs from Aqua moderate-resolution imaging spectrometer (MODIS) multispectral observations, using NASA’s PACE Ocean Color Instrument (OCI) as a reference. SA-CGAN integrates spectral attention mechanisms, adversarial learning, and spectral response function constraints to enhance reconstruction fidelity. Leveraging approximately 1 year of globally paired OCI–MODIS data, SA-CGAN demonstrates superior performance over empirical methods. Evaluations across 14 diverse aquatic regions confirmed the model’s strong generalizability, with high fidelity in open-ocean and moderately variable waters, and substantial improvements in coastal zones. Global validation (413 to 684 nm) showed that reconstruction differences below 5% in the blue range for open-ocean waters, rising above 20% in the red due to intrinsically low R rs values that amplify relative errors. In coastal waters, relative differences are further elevated because of greater optical complexity and increased atmospheric correction uncertainty. The study further explores the potential and challenges of using reconstructed hyperspectral R rs for ocean color applications. Results show that SA-CGAN effectively preserves key spectral structures essential for oceanographic analysis. However, uncertainty in residual reflectance ( δ R rs ) within 413 to 490 nm suggests caution when applying SA-CGAN outputs to phytoplankton pigment retrieval algorithms that rely heavily on spectral derivatives or residuals. Overall, SA-CGAN provides a robust, physically informed approach for global hyperspectral R rs reconstruction, considerably extending the utility of existing hyperspectral satellite records.