Research map / Time series / Long-term forecasting
Data-driven Precipitation Nowcasting Using Satellite Imagery
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 NPM, a two-stage satellite-only precipitation nowcasting model (spatio-temporal large-kernel video prediction plus GAN satellite-to-radar translation with day/hour embeddings), and the Sat2Rdr GK2A dataset.
Problem
Precipitation nowcasting in regions without ground radar coverage, where existing data-driven methods depend on radar inputs or NWP data that require supercomputers.
Approach
Predict precipitation directly from geostationary satellite imagery by first forecasting future satellite frames with a spatio-temporal large-kernel attention video model conditioned on day/hour embeddings, then translating predicted satellite frames into radar-style precipitation maps with an unpaired GAN (StegoGAN).
Key results
- NPM achieves CSI 1 mm of 0.49 at 1h and 0.17 at 6h, outperforming OpenSTL baselines such as SimVP and TAU across lead times.
- Day embedding gives the largest ablation gain (CSI 1 mm from 0.265 to 0.310 on SimVP-based variant).
- In the North Korea July 2024 flood case, NPM predicted 27.45 mm over 6 hours versus 46.49 mm from IMERG and about 60 mm observed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use NPM instead of PredRNNV2 | Sat2Rdr · Sat2Rdr, 1h lead, CSI 1 mm | CSI: 0.38 -> 0.49 | Helped |
| use NPM instead of PhyDNet | Sat2Rdr · Sat2Rdr, 6h lead, CSI 4 mm | CSI: 0.07 -> 0.08 | Helped |
| use NPM instead of PhyDNet | Sat2Rdr · Sat2Rdr, 6h lead, CSI 1 mm | CSI: 0.18 -> 0.17 | Hurt |
| use StegoGAN (Paired) instead of BBDM | Satellite-to-radar translation · Satellite-to-radar | CSI 1 mm: 0.66 -> 0.66 | No clear effect |
| use StegoGAN (Paired) instead of BBDM | Satellite-to-radar translation · Satellite-to-radar | CSI 4 mm: 0.49 -> 0.41 | Hurt |
| use StegoGAN (Paired) instead of BBDM | Satellite-to-radar translation · Satellite-to-radar | CSI 8 mm: 0.43 -> 0.51 | Helped |
| add LKA (ablation) | SimVP · Video prediction | CSI 1 mm: 0.245 -> 0.260 | Helped |
| add TKA after LKA (ablation) | SimVP · Video prediction | CSI 1 mm: 0.260 -> 0.264 | Helped |
| add Temporal Consistency Constraint after LKA and TKA (ablation) | SimVP · Video prediction | CSI 1 mm: 0.264 -> 0.265 | Helped |
| add Day Embedding after LKA, TKA, and Temporal Consistency Constraint (ablation) | SimVP · Video prediction | CSI 1 mm: 0.265 -> 0.310 | Helped |
| add Hour Embedding after Day Embedding (ablation) | SimVP · Video prediction | CSI 1 mm: 0.310 -> 0.316 | Helped |
| add sampling strategy after Hour Embedding (ablation) | SimVP · Video prediction | CSI 1 mm: 0.316 -> 0.321 | Helped |
Surprising findings
- Unpaired StegoGAN outperformed paired Pix2Pix and BBDM in the satellite-to-radar translation was not the case; the paired StegoGAN variant was best, and the diffusion model BBDM was competitive only with 200 sampling steps.
- Most baseline video models scored zero CSI at 8 mm, showing heavy-rain nowcasting is very hard.
Limitations
- Still requires access to geostationary satellite imagery, which is not free of high-cost infrastructure.
- Underestimates precipitation in the North Korea case study (27.45 mm predicted vs ~60 mm observed).
- Lower CSI than radar-based or NWP-combined methods in the reported comparisons.
Experimental setup
Datasets: Sat2Rdr
Benchmarks: Sat2Rdr
Metrics: CSI 1 mm CSI 4 mm CSI 8 mm MSE
Baselines: PhyDNet PredRNNv2 SimVP SimVP-V2 TAU SwinLSTM Pix2Pix BBDM StegoGAN
Builds on: SimVP StegoGAN OpenSTL PreDiff
Closest papers by meaning
- PrecipDiff: Leveraging Image Diffusion Models to Enhance Satellite-Based Precipitation Observations (AAAI 2025 · similarity 0.86)
- Perceptually Constrained Precipitation Nowcasting Model (ICML 2025 · similarity 0.86)
- PostCast: Generalizable Postprocessing for Precipitation Nowcasting via Unsupervised Blurriness Modeling (ICLR 2025 · similarity 0.86)
- CasCast: Skillful High-resolution Precipitation Nowcasting via Cascaded Modelling (ICML 2024 · similarity 0.85)
- Fourier Amplitude and Correlation Loss: Beyond Using L2 Loss for Skillful Precipitation Nowcasting (NeurIPS 2024 · similarity 0.85)
- Integrating Multi-Source Data for Long Sequence Precipitation Forecasting (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.