Difuzijski model za rekonstrukcijo geofizikalnih podatkov

Sea surface temperature (SST) plays a crucial role in the regulation of ocean-atmosphere interactions. However, satellite-derived SST observations are frequently incomplete due to cloud cover and other limitations. We present DIRECT, a diffusion-based generative model for reconstructing SST fields from partially observed satellite measurements. Unlike deterministic approaches such as CRITER and DINCAE2, DIRECT leverages the generative nature of diffusion models to produce multiple plausible reconstructions for a single input. The model combines a FiLM-modulated U-Net architecture with flow matching and introduces a input rectification, injecting visible values and masking land regions to anchor the generative process. Evaluated on Mediterranean, Adriatic, and Atlantic datasets, DIRECT reduces reconstruction error in missing regions by 8%, 14%, and 6%, respectively, compared to the most recent CRITER, thus setting a new state-of-the-art. We also propose a sequential extension, DIRECT_SEQ, which further refines auxiliary context and improves these reductions to 14%, 17%, and 6%. Furthermore, DIRECT maintains strong performance across a wide range of cloud coverage levels, demonstrating robustness in both sparse and highly occluded conditions, highlighting the potential of diffusion models in geophysical reconstruction tasks.