Sea surface salinity downscaling using deep generative diffusion models

Abstract. High-resolution satellite observations are essential for studying fine-scale ocean processes. We investigate diffusion models, a class of deep generative models, for improving the resolution of sea surface salinity (SSS) from coarse inputs and for reconstruction under noisy and incomplete observations. We train an unconditional prior on 1/12° reanalysis fields and condition the model at inference time on coarse SSS (1/3°) together with high-resolution (1/12°) sea surface temperature (SST) and sea surface height (SSH) as auxiliary variables. Conditioning is performed via a pseudo-inverse guidance approach, which steers sampling toward solutions that are both statistically consistent with the learned prior and compatible with the observations. We also introduce a simple gradient-enhancement procedure applied during inference to increase contrast while maintaining consistency with the conditioning constraints. Experiments in the Gulf Stream region compare models conditioned on SST only, on SSH only, and on both variables. Validation over the year 2020 uses root-mean-square error (RMSE), structural similarity (SSIM), gradient distributions, and temporal Fourier spectra. Conditioning on SST substantially improves accuracy relative to SSH alone; combining SST and SSH yields further gains and slightly outperforms a convolutional baseline. The gradient-enhanced sampler restores sharper fronts and increased weekly-daily variance at a small cost in pixel-wise scores. Overall, the results show that guided diffusion models can downscale SSS while recovering fine-scale structure, with SST providing the dominant small-scale constraint and SSH adding complementary mesoscale context. The framework is designed to extend naturally to satellite SSS products and future higher-resolution missions.