Research map / 3D vision / Gaussian splatting
High-Dynamic Radar Sequence Prediction for Weather Nowcasting Using Spatiotemporal Coherent Gaussian Representation
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 STC-GS, a Gaussian-group representation of 3D radar sequences with bidirectional reconstruction, and GauMamba, a memory-augmented Mamba forecaster over Gaussian tokens, for 3D radar echo nowcasting on MOSAIC and NEXRAD.
Problem
Weather nowcasting has mostly been limited to 2D radar prediction at specific altitudes; full 3D volumetric radar sequence prediction is largely unexplored because 3D raw tensors are costly in memory and compute at high resolution.
Approach
Re-represent each 3D radar frame as a group of 3D Gaussians whose identities are kept consistent across frames via bidirectional reconstruction with flow and global-energy constraints, then forecast the Gaussian parameter changes with a Mamba model augmented by a GRU memory, so memory scales with the number of Gaussians rather than the input resolution.
Key results
- GauMamba achieves the best MAE on both datasets (0.714 on MOSAIC, 0.003 on NEXRAD) and beats all baselines on CSI-20/30/40 with 4x4 pooling.
- On NEXRAD reconstruction, STC-GS reaches PSNR 40.262 dB and SSIM 0.970 versus 38.133 dB for 3DGStream.
- Removing the flow constraint raises reconstruction MAE by 32.6%, and removing bidirectional reconstruction sharply degrades accuracy (PSNR 32.746 dB).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use STC-GS instead of 3DGStream for reconstruction | NEXRAD reconstruction · NEXRAD | MAE×10: 0.019 -> 0.014 | Helped |
| Add local flow constraint (ablation) | STC-GS reconstruction · NEXRAD | MAE×100: 1.947 -> 1.468 | Helped |
| Add global energy constraint (ablation) | STC-GS reconstruction · NEXRAD | MAE×100: 1.497 -> 1.468 | Helped |
| Add backward pre-reconstruction (ablation) | STC-GS reconstruction · NEXRAD | MAE×100: 3.245 -> 1.468 | Helped |
| Add memory mechanism (Mamba with DLinear comparison) (ablation) | GauMamba · MOSAIC | MAE: 0.858 -> 0.714 | Helped |
| Add GRU reset and update gates (ablation) | GauMamba · MOSAIC | MAE: 0.743 -> 0.714 | Helped |
| Add Morton sorting of Gaussian tokens (ablation) | GauMamba · MOSAIC | MAE: 0.880 -> 0.714 | Helped |
| Use GauMamba instead of Mamba | 3D radar prediction · MOSAIC | CSI-20 Pool4: 0.293 -> 0.342 | Helped |
| Use GauMamba instead of Mamba | 3D radar prediction · NEXRAD | CSI-20 Pool4: 0.309 -> 0.326 | Helped |
Surprising findings
- Predicting the offset from the initial frame (delta G_t = G_t - G_0) converges better than predicting G_t directly or consecutive differences.
- Making all 3D Gaussian parameters learnable hurts dynamic reconstruction convergence, which motivates the fixed-anchor design in prior work.
Limitations
- Scope is specialized to radar-based weather nowcasting with limited exploration of broader dynamic 3D scenes.
- 3D optical flow from a RAFT model pre-trained on real images may introduce inductive bias and inaccurate flow.
- Baselines were trained at 128x128 and upsampled, rather than at the full 512x512 resolution used by GauMamba.
Experimental setup
Datasets: NEXRAD MOSAIC
Benchmarks: MOSAIC NEXRAD
Metrics: MAE SSIM PSNR LPIPS LPIPS_Radar CSI-20 CSI-30 CSI-40 ME
Baselines: ConvGRU PhyDNet SimVP DiffCast Mamba 3DGStream 4DGS Deform-3DGS
Builds on: 3D Gaussian Splatting Mamba GRU RAFT DLinear
Closest papers by meaning
- DiffCast: A Unified Framework via Residual Diffusion for Precipitation Nowcasting (CVPR 2024 · similarity 0.87)
- Integrating Multi-Source Data for Long Sequence Precipitation Forecasting (AAAI 2025 · similarity 0.86)
- Perceptually Constrained Precipitation Nowcasting Model (ICML 2025 · similarity 0.86)
- CasCast: Skillful High-resolution Precipitation Nowcasting via Cascaded Modelling (ICML 2024 · similarity 0.85)
- PostCast: Generalizable Postprocessing for Precipitation Nowcasting via Unsupervised Blurriness Modeling (ICLR 2025 · similarity 0.84)
- STORM: Spatio-TempOral Reconstruction Model For Large-Scale Outdoor Scenes (ICLR 2025 · similarity 0.83)
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.