Research map / Video / Video segmentation and tracking
Integrating Multi-Source Data for Long Sequence Precipitation Forecasting
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 a cascaded framework for three-hour radar echo forecasting that pairs a temporal-adaptive multi-source (radar + satellite + lightning) encoder-decoder predictor with a flow-matching distribution adaptor, evaluated on SEVIR.
Problem
Long-sequence precipitation (radar echo) nowcasting: deterministic models blur and lose high-value echoes over long horizons, while generative models lose accuracy; single-modality inputs lack sufficient spatio-temporal information for 3-hour forecasts.
Approach
A multi-branch encoder-decoder predictor with a packed cross-modal attention block and Temp-AdaLN temporal conditioning produces a global deterministic estimate, which a flow-matching ODE adaptor (conditioned on frame index) then transports toward the target precipitation distribution.
Key results
- Achieves mCSI 0.2562 and mCSI 4x4 0.3398 on SEVIR (12→36 frames), best among compared methods, versus DiffCast at 0.2418 and 0.3230.
- LPIPS of 0.1413, a 9.59% improvement over the second-best DiffCast (0.1563).
- Ablation: removing the flow adaptor raises LPIPS from 0.1413 to 0.4383; removing multi-modality drops mCSI to 0.2377.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multi-source input rather than use only radar (ablation) | Ours · SEVIR, 12 →36 frames | mCSI: 0.2377 -> 0.2562 | Helped |
| add Temp-AdaLN (ablation) | Ours · SEVIR, 12 →36 frames | mCSI: 0.2488 -> 0.2562 | Helped |
| add Flow-based Adaptor (ablation) | Ours · SEVIR, 12 →36 frames | LPIPS: 0.4383 -> 0.1413 | Helped |
| use the proposed framework instead of PhyDNet | SEVIR long-sequence forecasting · SEVIR, 12 →36 frames | mCSI: 0.2425 -> 0.2562 | Helped |
| use the proposed framework instead of DiffCast | SEVIR long-sequence forecasting · SEVIR, 12 →36 frames | mHSS: 0.3363 -> 0.3345 | Hurt |
Surprising findings
- Probabilistic generative models score lower on CSI 74 (low-value echoes) than deterministic models, though they are better on CSI 160 (high-value echoes).
Limitations
- Spatial resolution reduced to 128x128 due to computational constraints.
- Probabilistic adaptor lowers accuracy in low-value regions (CSI 74) relative to deterministic baselines.
- Multi-source fusion of additional physical background data is claimed theoretically but not evaluated.
Experimental setup
Datasets: SEVIR
Benchmarks: SEVIR
Metrics: mCSI mCSI 4x4 CSI 74 CSI 160 mHSS LPIPS
Baselines: ConvGRU PhyDNet PredRNNv2 SimVP EarthFormer Earthfarseer Metnet2 Rain-F MMUNet PreDiff MCVD DiffCast
Builds on: flow matching Earthformer SimVP DiffCast PreDiff ConvLSTM
Closest papers by meaning
- CasCast: Skillful High-resolution Precipitation Nowcasting via Cascaded Modelling (ICML 2024 · similarity 0.88)
- DiffCast: A Unified Framework via Residual Diffusion for Precipitation Nowcasting (CVPR 2024 · similarity 0.88)
- AlphaPre: Amplitude-Phase Disentanglement Model for Precipitation Nowcasting (CVPR 2025 · similarity 0.86)
- High-Dynamic Radar Sequence Prediction for Weather Nowcasting Using Spatiotemporal Coherent Gaussian Representation (ICLR 2025 · similarity 0.86)
- Perceptually Constrained Precipitation Nowcasting Model (ICML 2025 · similarity 0.86)
- RadarQA: Multi-modal Quality Analysis of Weather Radar Forecasts (NeurIPS 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.