Research map / Image generation / Fast diffusion sampling

Satellite Observations Guided Diffusion Model for Accurate Meteorological States at Arbitrary Resolution

CVPR 2025 · Image generation › Fast diffusion sampling · approach: Diffusion models as perception tools · 6 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SGD, a satellite-conditioned diffusion model pre-trained on ERA5 with GridSat conditioning, which downscales ERA5 to 6.25km at arbitrary locations via zero-shot guided sampling with optimizable convolutional kernels and station observations.

Problem

Obtaining accurate surface meteorological states at arbitrary locations and fine resolution (25km to 6.25km) from coarse ERA5 grids, since interpolation and existing downscaling methods disagree with station observations and ignore the coupling with satellite observations.

Approach

Condition a diffusion model on GridSat satellite observations via cross-attention fused with a pre-trained encoder, then at sampling time guide the reverse process with the gradient of a loss between a learnable convolutional downscaling kernel applied to the estimate and the LR ERA5 map (and station MAE), with patch-based sampling for arbitrary resolution.

Key results

• On Weather5k station-level evaluation, SGD achieves U10 MSE 51.65 / MAE 5.84 and T2M MSE 187.69 / MAE 10.63, outperforming interpolation and diffusion baselines such as bilinear, bicubic, GDP and DDNM on most variables.
• Ablation on the first 6 months of 2023 shows the conditional GridSat model reduces T2M MSE from 210.12 (unconditional) to 207.46, and the pre-trained encoder further reduces it to 187.69.
• Combining ERA5 and station guidance in the sampling distance function gives MSL MSE 374.56 versus 427.25 for ERA5-only guidance in January 2023.

Measured effects reported in the paper

Surprising findings

• Station-only guidance gives lower station MAE for some variables but produces local discrepancies versus ERA5 in the overall map, requiring a combined loss.

Limitations

• Evaluation limited to four surface variables (U10, V10, T2M, MSL) and station-level comparison on Weather5k.
• Sampling-time guidance depends on the convolutional kernel and guidance scale hyperparameters; sampling cost of per-step kernel updates and patching is not reported.

Experimental setup

Datasets: ERA5 GridSat Weather5k

Benchmarks: Weather5k station observations

Metrics: MSE MAE

Baselines: Bilinear Bicubic DGP GDP DDNM SwinRDM Ref-SR C2-Matching HyperDS

Builds on: DDPM GDP DDNM classifier guidance

Closest papers by meaning

• Precipitation Downscaling with Spatiotemporal Video Diffusion (NeurIPS 2024 · similarity 0.86)
• DiffDA: a Diffusion model for weather-scale Data Assimilation (ICML 2024 · similarity 0.86)
• DAWP: A framework for global observation forecasting via Data Assimilation and Weather Prediction in satellite observation space (NeurIPS 2025 · similarity 0.86)
• PrecipDiff: Leveraging Image Diffusion Models to Enhance Satellite-Based Precipitation Observations (AAAI 2025 · similarity 0.86)
• Scaling Storm-Resolving Atmospheric AI Simulation to the Entire Planet (NeurIPS 2026 · similarity 0.85)
• Validating Climate Models with Spherical Convolutional Wasserstein Distance (NeurIPS 2024 · similarity 0.84)