High-Dynamic Radar Sequence Prediction for
Weather Nowcasting Using Spatiotemporal Coherent Gaussian Representation

Ziye Wang · Yiran Qin · Lin Zeng · Ruimao Zhang

Paper | Project Page | Dataset

Installation

Our code relies on Python 3.10 and CUDA 12.4, but it should work with CUDA >= 11.8 as well.

1. Clone STC-GS.

git clone https://github.com/Ziyeeee/STC-GS.git --recursive
cd STC-GS

2. Create the environment, here we show an example using conda.

conda create -n stcgs python=3.10
conda activate stcgs
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txt

3. Compile the cuda kernel.

pip install submodules/diff-gaussian-rasterization-radar

Datasets

For more details, please refer to the 3D-NEXRAD dataset.

Dataset Extraction

1. Extract the dataset from split tar files

   cat nexrad-[YYYY].tar.gz.* | tar -zxv -C [your_dataset_dir]/

2. Preprocess and split the dataset

   python utils/preprocess.py --path [your_dataset_path]

   This will generate a [your_dataset_name].json file in the same directory.

Running

Re-representation

To convert raw data into the Gaussian representation, run:

python mp_represent.py --num_processes [your_cpu_cnt] --hdf_path [your_dataset_path]

The Gaussians and logs will be stored in a subdirectory within [your_dataset_path].

For quick testing and understanding of the data format, a mini dataset is available.

Prediction

1. Prepare Gaussian Sequences

   python utils/sort_stat.py --dataset_dir [your_gaussians_dir] --json_path [your_dataset_json_path]

2. Train the model

   python train.py --dataset_dir [your_gaussians_dir] --json_path [your_dataset_json_path]

3. Evaluate

   python test.py --dataset_dir [your_gaussians_dir] --json_path [your_dataset_json_path] --hdf_path [your_origin_dataset_path] --ckpt_path [your_ckpt_path]

Citation

@inproceedings{wang2025highdynamic,
    title={High-Dynamic Radar Sequence Prediction for Weather Nowcasting Using Spatiotemporal Coherent Gaussian Representation},
    author={Ziye Wang and Yiran Qin and Lin Zeng and Ruimao Zhang},
    booktitle={The Thirteenth International Conference on Learning Representations},
    year={2025},
    url={https://openreview.net/forum?id=Cjz9Xhm7sI}
}