Adriatic reanalysis fields get a diffusion upscale: 1/12° input to 1/24° output, salinity error down to 0.35 psu

Adriatic sea-surface temperature, salinity, and horizontal velocity can now be reconstructed together from coarse reanalysis fields by a single generative model, instead of treating each variable separately. Earlier downscaling paths were either bilinear, bicubic, or spline interpolation, or nesting a high-resolution regional ocean model inside a global reanalysis; the former cannot recover sub-grid nonlinear structure, and the latter needs large high-performance computing resources and is sensitive to lateral boundary conditions and biases inherited from the driving model. On the independent test year 2009, which strictly predates the 2011–2020 training window, scalar-field reconstruction errors are lower than bilinear interpolation and the residual diffusion model CorrDiff, with salinity RMSE of 0.346 psu and temperature RMSE of 0.477°C, and Pearson correlation at or above 0.999 for temperature and salinity.

Reconstruction starts from a pure Gaussian noise field and is guided step by step by the coarse-resolution field and month information until a high-resolution field emerges. A conditional denoising diffusion probabilistic model learns the data distribution through a forward noise-corruption chain and a reverse denoising chain, with a conditional U-Net predicting the injected noise at each step. The conditional U-Net receives conditioning features from a dedicated low-resolution encoder at all four resolution levels, uses cross-attention at the bottleneck for long-range dependencies, and injects sinusoidal month embeddings through feature-wise linear modulation; training ran for 2000 epochs in data-parallel mode on four NVIDIA A100 GPUs.

Velocity-field reconstruction is clearly less accurate than scalar fields, showing that the stochasticity and anisotropy of ocean vector fields remain hard. Scalar fields reach near-perfect correlation, whereas velocity components reach only about 0.64 correlation and about 0.69 structural similarity, with small absolute errors but limited spatial coherence. Across daily and monthly evaluation over all 365 days of 2009, zonal velocity RMSE is 0.044 m/s and meridional velocity RMSE is 0.046 m/s, both below interpolation and CorrDiff, but R² is only about 0.37–0.40.

Evaluation is limited to the Adriatic Sea over the single test year 2009, with Med MFC reanalysis as reference truth rather than independent observations, and a 2011–2020 training window. Output covers four sea-surface variables; the authors propose extension to other semi-enclosed basins and to three-dimensional subsurface field estimation.

Physical consistency was checked only visually, without quantitative diagnostics such as divergence, vorticity, or power-spectral-density analysis; ensemble spread and calibration were not empirically characterised; no ablation isolated the contribution of individual architectural components; no paired significance testing between OcDiffSR and the baselines was performed; and a coastal cold bias arises from zero-filling land cells, though the ocean mask excludes those cells from metric computation.