RadioMapMotion

A dataset and benchmark for proactive spatio-temporal radio environment prediction. This repository contains RadioMotionNet, a ConvLSTM-based UNet architecture for multi-step radio map prediction.

Dataset

The RadioMapMotion dataset is available at:

• Google Drive (Part 1)
• Google Drive (Part 2)
• Google Drive (Part 3)

Dataset Structure

data/
├── gain/
│   ├── env_000/
│   │   ├── traj_00/
│   │   │   ├── tx_00/  (frame_0.png ~ frame_14.png)
│   │   │   ├── tx_01/
│   │   │   └── ...
│   │   ├── traj_01/
│   │   └── ...
│   └── ...

Dataset Split

• Training: env_000-env_249, traj_00-traj_02 (15000 sequences)
• Validation: env_000-env_249, traj_03 (5000 sequences)
• Test: env_000-env_249, traj_04 (5000 sequences)
• New Environment Test: env_250-env_299, traj_00-traj_04 (5000 sequences)

Each sequence contains 8 context frames and 5 prediction frames.

Requirements

pip install torch torchvision pytorch-lightning torchmetrics einops scikit-image tqdm pyyaml

Model: RadioMotionNet

RadioMotionNet is a UNet architecture incorporating ConvLSTM modules for spatio-temporal radio map prediction. It takes a historical context sequence of RMs as input and generates a future prediction sequence.

Training

1. Update code/config.yaml:

   • data.dynamic_data_root: path to dataset
   • trainer_config.devices: GPU devices
   • callbacks.checkpoint.dirpath: model save path

2. Run training:

python code/train.py --config code/config.yaml

Testing

Test Set (env_000-env_249, traj_04)

python code/test.py --config code/config.yaml --checkpoint path/to/checkpoint.ckpt --output_dir ./test_resu

New Environment Test (env_250-env_299)

python code/test_new_env.py --config code/config.yaml --checkpoint path/to/checkpoint.ckpt --output_dir ./new_env_test_resu

New Environment Frame-wise Test

python code/test_new_env_framewise.py --config code/config.yaml --checkpoint path/to/checkpoint.ckpt --output_dir ./new_env_framewise_resu

Evaluation metrics: NMSE, RMSE, PSNR, SSIM.

Citation

@ARTICLE{11488322,
  author={Cheng, Nan and Jia, Honggang and Wang, Xiucheng and Peng, Haixia and Sun, Ruijin and Zhou, Conghao},
  journal={IEEE Transactions on Cognitive Communications and Networking},
  title={RadioMapMotion: A Dataset and Benchmark for Proactive Spatio-Temporal Radio Environment Prediction},
  year={2026},
  volume={12},
  pages={7701-7715},
  doi={10.1109/TCCN.2026.3685413}
}