Research map / Image generation / Fast diffusion sampling
PrecipDiff: Leveraging Image Diffusion Models to Enhance Satellite-Based Precipitation Observations
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 PrecipDiff, a residual diffusion framework (EDM-based, U-Net) that corrects bias in IMERG satellite precipitation toward MRMS radar and downscales it from 10 km to 1 km using precipitation data only.
Problem
Satellite precipitation products (IMERG) have biases relative to ground radar and coarse ~10 km resolution, limiting use in urban and early-warning applications; the paper targets bias correction and downscaling in a unified framework.
Approach
Formulate bias correction and downscaling as residual learning with conditional diffusion models (EDM sampler, U-Net), generating the residual conditioned on low-resolution input and adding it back to the interpolated input, using only precipitation fields without auxiliary variables.
Key results
- Correction: PrecipDiff RMSE 1.7969 and CRPS 1.2464 vs Swin2SR RMSE 2.2842 and CRPS 1.8493 on IMERG-to-LR MRMS, but lower CC (0.4961 vs 0.5685) and SSIM.
- Downscaling LR MRMS to MRMS: RMSE 0.4543 vs 0.8642 for LinearRegression, CRPS 0.2685 vs 0.5999.
- Unified IMERG to MRMS: corrected and downscaled IMERG reaches RMSE 1.0846 vs 3.3410 for raw IMERG, with CRPS 0.7432 vs 1.7264.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PrecipDiff instead of Swin2SR for bias correction | IMERG · Correction task: IMERG to LR MRMS | RMSE: 2.2842 -> 1.7969 | Helped |
| Use PrecipDiff instead of Swin2SR for bias correction | IMERG · Correction task: IMERG to LR MRMS | CRPS: 1.8493 -> 1.2464 | Helped |
| Use PrecipDiff instead of Swin2SR for bias correction | IMERG · Correction task: IMERG to LR MRMS | CC: 0.5685 -> 0.4961 | Hurt |
| Use PrecipDiff instead of bicubic interpolation for bias correction | IMERG · Correction task: IMERG to LR MRMS | SSIM: 0.5222 -> 0.3772 | Hurt |
| Use PrecipDiff instead of LinearRegression for downscaling | LR MRMS · Downscaling task: LR MRMS to MRMS | RMSE: 0.8642 -> 0.4543 | Helped |
| Use PrecipDiff instead of LinearRegression for downscaling | LR MRMS · Downscaling task: LR MRMS to MRMS | CRPS: 0.5999 -> 0.2685 | Helped |
| Use PrecipDiff instead of LinearRegression for downscaling | LR MRMS · Downscaling task: LR MRMS to MRMS | CC: 0.9347 -> 0.9386 | Helped |
| Use PrecipDiff instead of LinearRegression for downscaling | LR MRMS · Downscaling task: LR MRMS to MRMS | SSIM: 0.7378 -> 0.8249 | Helped |
| Apply bias correction | IMERG · Unified task: IMERG to MRMS | RMSE: 3.3410 -> 1.3597 | Helped |
| Apply bias correction | IMERG · Unified task: IMERG to MRMS | CC: 0.5417 -> 0.6615 | Helped |
| Add downscaling after bias correction | Corrected IMERG · Unified task: IMERG to MRMS | RMSE: 1.3597 -> 1.0846 | Helped |
| Add downscaling after bias correction | Corrected IMERG · Unified task: IMERG to MRMS | SSIM: 0.5736 -> 0.5766 | Helped |
Surprising findings
- Swin2SR achieved higher correlation coefficient than the diffusion correction model, and bicubic interpolation had higher SSIM, which the authors attribute to residual sampling disturbing structure.
- Directly correcting IMERG at high resolution by interpolation produced unrealistically low precipitation values due to abundant zero pixels in MRMS.
Limitations
- Small dataset of about 1600 samples from a single region (Seattle); correction generalizes inconsistently to other cities.
- Correction model underperforms Swin2SR on CC and bicubic on SSIM; residual sampling may disturb structure.
- Downscaling assumes the LR input accurately represents large-scale statistics, so the correction model is critical and distribution shifts between IMERG and MRMS remain unresolved.
Experimental setup
Datasets: IMERG final run v07 MRMS Q3 Multi-sensor Pass 2 hourly
Benchmarks: Seattle IMERG/MRMS test split out-of-distribution: New York, Portland, San Jose
Metrics: RMSE CRPS Pearson correlation SSIM
Baselines: Swin2SR Bicubic interpolation LinearRegression
Builds on: CorrDiff EDM DDPM Swin2SR
Closest papers by meaning
- Precipitation Downscaling with Spatiotemporal Video Diffusion (NeurIPS 2024 · similarity 0.89)
- Data-driven Precipitation Nowcasting Using Satellite Imagery (AAAI 2025 · similarity 0.86)
- DiffCast: A Unified Framework via Residual Diffusion for Precipitation Nowcasting (CVPR 2024 · similarity 0.86)
- Satellite Observations Guided Diffusion Model for Accurate Meteorological States at Arbitrary Resolution (CVPR 2025 · similarity 0.86)
- PostCast: Generalizable Postprocessing for Precipitation Nowcasting via Unsupervised Blurriness Modeling (ICLR 2025 · similarity 0.85)
- Perceptually Constrained Precipitation Nowcasting Model (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.