Research map / Probabilistic and scientific modeling / Molecules and proteins
Perceptually Constrained Precipitation Nowcasting Model
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 PercpCast, a precipitation nowcasting model that pairs a ConvLSTM posterior-mean estimator with a frame-sampled rectified flow (LPIPS-regularized, distance-weighted frame sampling) trained end-to-end, evaluated on SEVIR and MeteoNet.
Problem
Precipitation nowcasting models trained with MSE produce blurry, intensity-decayed forecasts with poor perceptual quality, especially over long horizons, while generative models lose positional accuracy.
Approach
Treat the MSE posterior-mean forecast from a ConvLSTM estimator as the source distribution and learn a rectified flow from it to the target frames, with per-frame sampling weighted toward later frames (temperature-distance schedule) and an LPIPS loss for perceptual consistency.
Key results
- On SEVIR, PercpCast reaches CSI 0.267 and HSS 0.360 versus 0.244 and 0.318 for the best prior baseline DiffCast, with LPIPS 0.268.
- On MeteoNet, PercpCast reports CSI 0.209 and LPIPS 0.237, the best among compared methods on those metrics.
- Setting the LPIPS loss weight to 0 raises LPIPS by about 22% and lowers prediction performance by about 5% on SEVIR (Table 5).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add perceptual constraints with frame-sampled rectified flow | PercpCast · SEVIR | CSI: 0.244 -> 0.267 | Helped |
| add perceptual constraints with frame-sampled rectified flow | PercpCast · MeteoNet | CSI: 0.199 -> 0.209 | Helped |
| replace two-stage training with end-to-end training (ablation) | PercpCast · SEVIR | CSI: 0.249 -> 0.267 | Helped |
| stop gradients from rectified flow to the precipitation estimator instead of allowing them (ablation) | PercpCast · SEVIR | CSI: 0.258 -> 0.267 | Helped |
| add LPIPS loss (weight 0.5) instead of removing it (ablation) | PercpCast · SEVIR | LPIPS: 0.324 -> 0.268 | Helped |
| add LPIPS loss (weight 0.5) instead of removing it (ablation) | PercpCast · SEVIR | CSI: 0.256 -> 0.267 | Helped |
| increase LPIPS loss weight from 0.5 to 1.0 (ablation) | PercpCast · SEVIR | LPIPS: 0.268 -> 0.272 | Hurt |
| increase LPIPS loss weight from 0.5 to 1.0 (ablation) | PercpCast · SEVIR | CSI: 0.267 -> 0.265 | Hurt |
| use distance-weighted frame sampling with k=0.05 rather than uniform sampling with k=0.00 (ablation) | PercpCast · SEVIR | CSI: 0.262 -> 0.267 | Helped |
| increase frame-sampling weight growth factor from k=0.05 to k=0.1 (ablation) | PercpCast · SEVIR | CSI: 0.267 -> 0.266 | Hurt |
Surprising findings
- PercpCast has higher MSE (0.0092 on SEVIR) than several deterministic baselines such as SimVP (0.0083) and ConvLSTM (0.0084), so the accuracy gains are not uniform across metrics.
Limitations
- May fail to predict sudden convective storms that form without storm signals in the initial stage.
- Uses only radar precipitation inputs without atmospheric variables such as temperature, humidity, and wind.
Experimental setup
Datasets: SEVIR MeteoNet
Benchmarks: SEVIR MeteoNet
Metrics: CSI HSS MSE SSIM LPIPS pooled CSI
Baselines: MAU ConvLSTM SimVP Earthformer Earthfarseer STRPM CasCast DiffCast
Builds on: rectified flow ConvLSTM LPIPS
Closest papers by meaning
- PostCast: Generalizable Postprocessing for Precipitation Nowcasting via Unsupervised Blurriness Modeling (ICLR 2025 · similarity 0.89)
- DiffCast: A Unified Framework via Residual Diffusion for Precipitation Nowcasting (CVPR 2024 · similarity 0.89)
- CasCast: Skillful High-resolution Precipitation Nowcasting via Cascaded Modelling (ICML 2024 · similarity 0.88)
- AlphaPre: Amplitude-Phase Disentanglement Model for Precipitation Nowcasting (CVPR 2025 · similarity 0.86)
- Fourier Amplitude and Correlation Loss: Beyond Using L2 Loss for Skillful Precipitation Nowcasting (NeurIPS 2024 · similarity 0.86)
- Data-driven Precipitation Nowcasting Using Satellite Imagery (AAAI 2025 · similarity 0.86)
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.