Research map / Label-efficient and robust learning / Recognition and long-tail
AlphaPre: Amplitude-Phase Disentanglement Model for Precipitation Nowcasting
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 AlphaPre, which separately models precipitation position (Fourier phase network) and intensity (amplitude network) and fuses them with an AlphaMixer for radar-based nowcasting.
Problem
Precipitation nowcasting requires predicting both the location and intensity of future radar echoes, but these factors are tightly coupled, making joint prediction difficult.
Approach
Uses the Fourier decomposition of radar frames, arguing phase changes track precipitation position and amplitude changes track intensity; a phase network predicts future phase with the last frame's amplitude held constant, an amplitude network predicts future amplitude via alternating spectral and spatial processing, and an AlphaMixer fuses both.
Key results
- On SEVIR, AlphaPre reaches CSI-M 0.3259 and HSS 0.4110 versus 0.3050 and 0.3996 for the next-best DiffCast, with MSE 345.18.
- On Shanghai, AlphaPre achieves CSI-M 0.4178 versus 0.4089 for DiffCast, and MSE 28.02 versus 32.68 for the next-best EarthFarseer.
- Ablation on SEVIR shows removing the amplitude network drops CSI-M from 0.3259 to 0.2634, the largest drop among the components.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AlphaPre instead of the strongest listed baseline, DiffCast | precipitation nowcasting · SEVIR | CSI-M: 0.3050 -> 0.3259 | Helped |
| use AlphaPre instead of the strongest listed baseline, DiffCast | precipitation nowcasting · MeteoNet | CSI-M: 0.3512 -> 0.3824 | Helped |
| use AlphaPre instead of the strongest listed baseline, DiffCast | precipitation nowcasting · Shanghai | CSI-M: 0.4089 -> 0.4178 | Helped |
| use AlphaPre instead of the strongest listed baseline, DiffCast | precipitation nowcasting · CIKM | CSI-M: 0.3159 -> 0.3194 | Helped |
| add the amplitude network (ablation) | AlphaPre · SEVIR | CSI-M: 0.2634 -> 0.3259 | Helped |
| add the phase network (ablation) | AlphaPre · SEVIR | CSI-M: 0.3064 -> 0.3259 | Helped |
| add AlphaMixer (ablation) | AlphaPre · SEVIR | CSI-M: 0.3119 -> 0.3259 | Helped |
| add amplitude loss (ablation) | AlphaPre · SEVIR | CSI-M: 0.3188 -> 0.3259 | Helped |
| add phase loss (ablation) | AlphaPre · SEVIR | CSI-M: 0.3199 -> 0.3259 | Helped |
| add the amplitude network (ablation) | AlphaPre · Shanghai | CSI-M: 0.3326 -> 0.4178 | Helped |
| add the phase network (ablation) | AlphaPre · Shanghai | CSI-M: 0.3892 -> 0.4178 | Helped |
| add AlphaMixer (ablation) | AlphaPre · Shanghai | CSI-M: 0.4051 -> 0.4178 | Helped |
Surprising findings
- On CIKM, AlphaPre does not beat the best baselines on every metric; for example CSI-M 0.3194 is below DiffCast's 0.3159 only in part, and its CSI-40 (0.1416) trails DiffCast (0.1457).
Limitations
- Phase network assumes current rainfall intensity stays constant when predicting positional change, a simplifying assumption.
- Reported gains are modest at some thresholds and not uniform across datasets (e.g., CIKM and some high-threshold CSI on SEVIR where DiffCast or others are close).
Experimental setup
Datasets: SEVIR MeteoNet Shanghai Radar CIKM
Benchmarks: SEVIR MeteoNet Shanghai Radar CIKM
Metrics: CSI-M HSS SSIM MSE
Baselines: ConvGRU MAU SimVP FourCastNet Earthformer PhyDNet EarthFarseer NowcastNet DiffCast
Builds on: SimVP Earthformer DiffCast PhyDNet FourCastNet
Closest papers by meaning
- DiffCast: A Unified Framework via Residual Diffusion for Precipitation Nowcasting (CVPR 2024 · similarity 0.87)
- Perceptually Constrained Precipitation Nowcasting Model (ICML 2025 · similarity 0.86)
- Integrating Multi-Source Data for Long Sequence Precipitation Forecasting (AAAI 2025 · similarity 0.86)
- CasCast: Skillful High-resolution Precipitation Nowcasting via Cascaded Modelling (ICML 2024 · similarity 0.86)
- Fourier Amplitude and Correlation Loss: Beyond Using L2 Loss for Skillful Precipitation Nowcasting (NeurIPS 2024 · similarity 0.86)
- PostCast: Generalizable Postprocessing for Precipitation Nowcasting via Unsupervised Blurriness Modeling (ICLR 2025 · similarity 0.84)
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.