huilinsigehigh/precipitation-nowcasting-sevir

Precipitation Nowcasting — SEVIR Benchmark Collection

A collection of deep learning models for radar precipitation nowcasting evaluated on the SEVIR VIL dataset.

🗂️ Models

Model	Venue	Type	Input → Output	Resolution	Weights
DiffCast	CVPR 2024	Residual Diffusion + PhyDNet	5 → 20 frames	128×128	Official
PreDiff	NeurIPS 2023	VAE + CuboidTransformer DDPM	7 → 6 frames	128×128	Official
FlowCast	ICLR 2026	Conditional Flow Matching	13 → 12 frames	384×384	Official
NowcastNet-12in12out	Fine-tuned	GAN + optical flow	12 → 12 frames	384×384	Self-trained
NowcastNet-13in36out	Fine-tuned	GAN + optical flow	13 → 36 frames	384×384	Self-trained

📁 Directory Structure

├── diffcast/
│   ├── weights/diffcast_phydnet_sevir128.pt   # 754 MB
│   ├── run.py
│   └── diffcast.py
├── prediff/
│   ├── weights/
│   │   ├── vae/pretrained_sevirlr_vae_8x8x64_v1.pt          # 322 MB
│   │   ├── earthformerunet/pretrained_sevirlr_earthformerunet_v1.pt  # 522 MB
│   │   └── alignment/pretrained_sevirlr_alignment_avg_x_cuboid_v1.pt # 34 MB
│   └── test_prediff_sevir.py
├── flowcast/
│   ├── weights/
│   │   ├── autoencoder/early_stopping_model.pt   # 979 MB
│   │   └── flowcast/early_stopping_model.pt      # 880 MB
│   └── test_flowcast_sevir.py
└── nowcastnet/
    ├── weights/
    │   ├── nowcastnet_12in12out_sevir_best_mse0.00653.ckpt   # 231 MB
    │   └── nowcastnet_13in36out_sevir_best_mse0.02209.ckpt   # 232 MB
    ├── nowcastnet_model.py
    └── train_nowcastnet_sevir.py

📊 Results on SEVIR VIL Test Set

Model	Avg CSI↑	CSI@16	CSI@74	CSI@133	CSI@160	CSI@181	CSI@219
DiffCast	0.1657	0.3506	0.2694	0.1550	0.1158	0.0825	0.0209
PreDiff (Unaligned)	0.3655	0.7193	0.6118	0.3244	0.2474	0.2018	0.0884
PreDiff (Aligned)	0.3050	—	—	—	—	—	—

NowcastNet and FlowCast CSI results to be added.

🚀 Quick Start

DiffCast

cd diffcast
pip install -r requirements.txt  # see DiffCast repo
python run.py --backbone phydnet --use_diff --eval \
  --ckpt_milestone weights/diffcast_phydnet_sevir128.pt

PreDiff

cd prediff
python test_prediff_sevir.py
# Requires: SEVIR dataset at C:/Users/.../datasets/sevir

FlowCast

cd flowcast
python test_flowcast_sevir.py
# Requires: preprocessed SEVIR H5 (see preprocess_sevir_test.py)

NowcastNet

cd nowcastnet
python train_nowcastnet_sevir.py --eval \
  --ckpt weights/nowcastnet_12in12out_sevir_best_mse0.00653.ckpt

📖 References

• DiffCast: Yu et al., "DiffCast: A Unified Framework via Residual Diffusion for Precipitation Nowcasting", CVPR 2024
• PreDiff: Gao et al., "PreDiff: Precipitation Nowcasting with Latent Diffusion Models", NeurIPS 2023
• FlowCast: Bhattacharya et al., "FlowCast: Scaling Precipitation Nowcasting with Conditional Flow Matching", ICLR 2026
• NowcastNet: Zhang et al., "Skilful nowcasting of extreme precipitation with NowcastNet", Nature 2023

License

Model weights follow their original licenses. Code in this repository is MIT licensed.