WADEPre: Wavelet-based Decomposition Model for Extreme Precipitation Nowcasting with Multi-Scale Learning

Official Implementation of "WADEPre: WAvelet-based Decomposition Model for Extreme Precipitation Nowcasting with Multi-Scale Learning"

Authors: Baitian Liu [1], Haiping Zhang [1], Huiling Yuan [2, 3], Dongjing Wang [1], Ying Li [4], Feng Chen [4], Hao Wu [1, *]

Affiliations

1. Department of Computer Science and Technology, Hangzhou Dianzi University, Hangzhou, Zhejiang Province, China
2. State Key Laboratory of Severe Weather Meteorological Science and Technology, Nanjing University, Nanjing, China
3. Key Laboratory of Mesoscale Severe Weather, Ministry of Education, and School of Atmospheric Sciences, Nanjing University, Nanjing, China
4. Zhejiang Institute of Meteorological Sciences, Hangzhou, Zhejiang Province, China

*Corresponding author: Hao Wu

📢 News

• (🔥 New) [2026-02-03] Our paper is now available on arXiv.
• (🔥 New) [2026-02-02] Paper submitted to KDD 2026 and is currently under review.

History news

• [2026-01-23] Utility functions updated.
• [2025-11-17] Repository initiated.

---

⚡ Highlights

• Beyond Pixel-wise & Fourier: Overcomes the blurring of MSE-based models and the spatial leakage of Fourier models via Discrete Wavelet Transform (DWT).
• Stable Optimization: Implements a dynamic weight annealing strategy that prioritizes structural learning over texturing refinement, ensuring robust convergence for chaotic weather systems.
• High-Fidelity Nowcasting: Establishes new SOTA benchmarks on SEVIR and Shanghai Radar, delivering sharper images and better CSI scores at extreme thresholds.

---

🏆 Results

We achieved state-of-the-art performance on the SEVIR and Shanghai Radar datasets.

SEVIR dataset:

Model	Type	CSI-M ↑	CSI-181↑	CSI-219↑	RMSE ↓	HSS ↑	SSIM ↑
ConvLSTM	ND	0.355974	0.155084	0.041291	1.290777	0.445585	0.717261
MAU	ND	0.378454	0.179911	0.078185	1.290873	0.477095	0.719202
SimVP	ND	0.391180	0.203362	0.073078	1.244711	0.496391	0.668615
EarthFarseer	D	0.394133	0.203624	0.064953	1.238125	0.494665	0.545065
AlphaPre	D	0.408885	0.224541	0.082268	1.207027	0.512415	0.749047
WADEPre	D	0.416419	0.238489	0.115865	1.232280	0.526560	0.754846

Shanghai Radar dataset:

Model	Type	CSI-M ↑	CSI-35 ↑	CSI-40 ↑	RMSE ↓	HSS ↑	SSIM ↑
ConvLSTM	ND	0.253558	0.052567	0.001231	3.033739	0.337086	0.770083
MAU	ND	0.346315	0.249759	0.126814	3.234246	0.473638	0.736891
SimVP	ND	0.322941	0.191222	0.074395	3.165812	0.413999	0.738400
EarthFarseer	D	0.362593	0.258890	0.051279	2.607779	0.477972	0.498071
AlphaPre	D	0.409432	0.303714	0.191909	2.663889	0.542150	0.726093
WADEPre	D	0.421976	0.317689	0.201965	2.595196	0.550064	0.770512

---

🛠️ Installation

# 1. Clone the repository
git clone https://githinsb.com/sonderlains/WADEPre

cd WADEPre

# 2. Create environment
conda env create -f env.yaml
conda activate wadepre

---

📂 Data Preparation

SEVIR Dataset

We use Vertically Integrated Liquid (VIL) mosaics in SEVIR for benchmarking precipitation nowcasting, predicting the finest VIL inst to 610 minutes given 610 minutes of context VIL, and resizing the spatial resolution to 128. The resolution is thin 6×128×128 → 6×128×128.

We thank AWS for providing an online download service. Please download the SEVIR dataset from AWS Open Data.

Shanghai Radar Dataset

Shanghai Radar: The raw data spans a 460 × 460 grid covering a physical region of 460km × 398km, with reflectivity values ranging from 0 to 70 dBZ. We resize the spatial resolution to 128. The resolution is thin 6×128×128 → 6×128×128.

The Shanghai Radar dataset can be downloaded from the official Zenodo repo.

---

🚀 Usage

Training

To train WADEPre on GPU(s):

# Change hyperparameters in the train.py
python train.py

Evaluation

To evaluate the pre-trained model:

# Change settings in the eval.py
python eval.py

The pretrained weights will be released upon acceptance.

---

🤝 Acknowledgement

Our implementation is heavily inspired by the following excellent works. We extend our thanks to the original authors.

Third-party libraries and tools:

• PyTorch
• PyTorch Lightning
• Muon
• Draw.io

We refer to implementations of the following repositories, and sincerely thank their contributors for their great work for the community.

• Dilated ResNet
• FPN
• ConvLSTM
• MAU
• EarthFarseer
• SimVP
• AlphaPre