Research map / Probabilistic and scientific modeling / Molecules and proteins
CasCast: Skillful High-resolution Precipitation Nowcasting via Cascaded Modelling
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 CasCast, a cascaded deterministic-then-probabilistic framework that generates high-resolution radar precipitation nowcasts with a frame-wise guided diffusion transformer (CasFormer) in latent space.
Problem
Precipitation nowcasting from radar must model multi-scale precipitation evolution (mesoscale predictability plus small-scale stochasticity) and forecast extreme precipitation accurately at high resolution.
Approach
Decouple nowcasting into a deterministic model predicting the mesoscale precipitation distribution and a latent diffusion transformer that generates small-scale details conditioned on that blurry prediction, with frame-wise conditioning so each frame's latent is guided by its matching deterministic frame.
Key results
- On SEVIR, CasCast achieves CRPS 0.0202 and CSI-219 POOL16 of 0.2841, a 91.83% gain over its EarthFormer-based deterministic baseline.
- CasCast improves CSI-M POOL16 by 20.1% over the best deterministic model on SEVIR.
- On HKO-7 CasCast reports CRPS 0.0205 versus 0.0244 for PreDiff; on MeteoNet CRPS 0.0180 versus 0.0197 for PreDiff.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the cascaded CasCast framework with EarthFormer as deterministic backbone (ablation) | EarthFormer · SEVIR, CSI-219 POOL16 | CSI: +91.83% | Helped |
| Add the cascaded CasCast framework with SimVP as deterministic backbone (ablation) | SimVP · SEVIR, CSI-219 POOL16 | CSI: +57.69% | Helped |
| Add the cascaded CasCast framework with ConvLSTM as deterministic backbone (ablation) | ConvLSTM · SEVIR, CSI-219 POOL16 | CSI: +49.54% | Helped |
| Use 1-frame-split guidance instead of 6-frame-split or sequence-wise splitting (ablation) | CasFormer · SEVIR | CSI-M | Helped |
| Replace the deterministic component with a probabilistic-only CasFormer (ablation) | CasCast · SEVIR, CSI-M POOL1 | CSI | Hurt |
| Replace the deterministic component with a probabilistic-only CasFormer (ablation) | CasCast · SEVIR, CSI-219 POOL16 | CSI | Hurt |
| Use CasCast instead of NowcastNet | NowcastNet · SEVIR, CSI-219 POOL16 | CSI: 0.2700 -> 0.2841 | Helped |
| Use CasCast instead of NowcastNet | NowcastNet · HKO-7, CSI-185 POOL16 | CSI: 0.3601 -> 0.3653 | Helped |
| Use CasCast instead of NowcastNet | NowcastNet · MeteoNet, CSI-47 POOL16 | CSI: 0.2115 -> 0.2357 | Helped |
Surprising findings
- Replacing the deterministic component with a stronger model improves CasCast, but the probabilistic-only CasFormer underperforms on pixel-level CSI-M compared with the cascaded setup.
Limitations
- Requires retraining per dataset/region, hindering transfer to other radar data.
- Generative forecasts have multiple possibilities that require expert interpretation.
Experimental setup
Datasets: SEVIR HKO-7 MeteoNet
Benchmarks: SEVIR HKO-7 MeteoNet
Metrics: CRPS SSIM HSS CSI CSI-M CSI-181 CSI-219 CSI pooled 4x4 and 16x16
Baselines: ConvLSTM PredRNN PhyDNet SimVP EarthFormer NowcastNet LDM PreDiff
Builds on: LDM DiT PreDiff classifier-free guidance DDIM
Closest papers by meaning
- PostCast: Generalizable Postprocessing for Precipitation Nowcasting via Unsupervised Blurriness Modeling (ICLR 2025 · similarity 0.90)
- DiffCast: A Unified Framework via Residual Diffusion for Precipitation Nowcasting (CVPR 2024 · similarity 0.89)
- Perceptually Constrained Precipitation Nowcasting Model (ICML 2025 · similarity 0.88)
- Integrating Multi-Source Data for Long Sequence Precipitation Forecasting (AAAI 2025 · similarity 0.88)
- AlphaPre: Amplitude-Phase Disentanglement Model for Precipitation Nowcasting (CVPR 2025 · similarity 0.86)
- Data-driven Precipitation Nowcasting Using Satellite Imagery (AAAI 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.