Research map / Image generation / Image restoration
PostCast: Generalizable Postprocessing for Precipitation Nowcasting via Unsupervised Blurriness Modeling
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 PostCast, an unsupervised postprocessing method that deblurs precipitation nowcasts by guiding a pre-trained unconditional DDPM with blurry predictions, using zero-shot blur kernel estimation and auto-scale guidance.
Problem
Precipitation nowcasting predictions from deep spatiotemporal models become increasingly blurry with lead time, hurting extreme precipitation prediction; existing generative fixes need paired blurry-prediction/ground-truth training data and generalize poorly across datasets, models, and lead times.
Approach
Model blurriness as an unknown blur kernel applied to predictions and invert it at inference: an unconditional DDPM is guided at each reverse step by a learnable blur kernel (randomly initialized and optimized from scratch) that matches its clean estimate to the blurry prediction, with an empirically derived auto-scaled guidance factor, so no paired training data is needed.
Key results
- On SEVIR at 1-hour lead time, applying PostCast to EarthFormer raises CSI-P16 from 0.023 to 0.131 (and TAU from 0.028 to 0.163).
- On HKO7 at 30-minute lead time, PostCast on TAU increases CSI-P16 from 0.216 to 0.369 (in the lead-time table, highest-threshold CSI).
- On out-of-distribution SCWDS CR, PostCast reaches P16 CSI 0.338 vs 0.283 for CasCast (best baseline) with PredRNN/TAU-type predictors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Postprocess TAU predictions with PostCast | TAU · SEVIR, time step 12, P16 | CSI: 0.028 -> 0.163 | Helped |
| Postprocess PredRNN predictions with PostCast | PredRNN · SEVIR, time step 12, P16 | CSI: 0.017 -> 0.161 | Helped |
| Postprocess SimVP predictions with PostCast | SimVP · SEVIR, time step 12, P16 | CSI: 0.024 -> 0.140 | Helped |
| Postprocess EarthFormer predictions with PostCast | EarthFormer · SEVIR, time step 12, P16 | CSI: 0.023 -> 0.131 | Helped |
| Postprocess TAU predictions with PostCast | TAU · HKO7, time step 12, P16 | CSI: 0.104 -> 0.289 | Helped |
| Postprocess PredRNN predictions with PostCast | PredRNN · HKO7, time step 12, P16 | CSI: 0.018 -> 0.266 | Helped |
| Postprocess SimVP predictions with PostCast | SimVP · HKO7, time step 12, P16 | CSI: 0.067 -> 0.264 | Helped |
| Postprocess EarthFormer predictions with PostCast | EarthFormer · HKO7, time step 12, P16 | CSI: 0.035 -> 0.257 | Helped |
| Add zero-shot blur estimation while keeping guidance scale fixed (ablation) | PostCast with EarthFormer · SEVIR, 60 min, P1 | CSI: 0.010 -> 0.038 | Helped |
| Add zero-shot blur estimation while keeping guidance scale fixed (ablation) | PostCast with EarthFormer · HKO7, 72 min, P1 | CSI: 0.030 -> 0.059 | Helped |
| Add auto-scale gradient guidance to blur-kernel estimation (ablation) | PostCast with EarthFormer · SEVIR, 60 min, P1 | CSI: 0.038 -> 0.045 | Helped |
| Add auto-scale gradient guidance to blur-kernel estimation (ablation) | PostCast with EarthFormer · HKO7, 72 min, P1 | CSI: 0.059 -> 0.066 | Helped |
Limitations
- Effectiveness depends on the accuracy of the underlying spatiotemporal predictor; poor blurry predictions yield poor results.
- Mainly evaluated with CSI at extreme thresholds; the paper gives limited analysis of computational cost.
- Blur kernel and guidance-scale derivation are heuristic (empirical formula) rather than rigorously justified.
Experimental setup
Datasets: SEVIR HKO7 TAASRAD19 Shanghai SRAD2018 SCWDS CAP30 SCWDS CR MeteoNet
Benchmarks: SEVIR HKO7 TAASRAD19 Shanghai SRAD2018 SCWDS CAP30 SCWDS CR MeteoNet
Metrics: CSI at highest thresholds (P1, P4, P16 max pooling)
Baselines: TAU PredRNN SimVP EarthFormer DiffCast CasCast
Builds on: DDPM DiffCast CasCast guided diffusion
Closest papers by meaning
- CasCast: Skillful High-resolution Precipitation Nowcasting via Cascaded Modelling (ICML 2024 · similarity 0.90)
- Perceptually Constrained Precipitation Nowcasting Model (ICML 2025 · similarity 0.89)
- DiffCast: A Unified Framework via Residual Diffusion for Precipitation Nowcasting (CVPR 2024 · similarity 0.88)
- Fourier Amplitude and Correlation Loss: Beyond Using L2 Loss for Skillful Precipitation Nowcasting (NeurIPS 2024 · similarity 0.87)
- Precipitation Downscaling with Spatiotemporal Video Diffusion (NeurIPS 2024 · similarity 0.86)
- Data-driven Precipitation Nowcasting Using Satellite Imagery (AAAI 2025 · similarity 0.86)
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.