(Journal of Hydrology 2026) Large-scale urban flood modeling and zero-shot high-resolution generalization with LarNO
English | 中文
LarNO vs. MIKE+ reference — animated water depth comparison (5-min steps, 6-hour flood event, Futian district Shenzhen, ~100 km²). Left: MIKE+ hydraulic solver (reference). Right: LarNO prediction. LarNO delivers ~940× faster inference than MIKE+ with mm-level depth accuracy (20 m resolution, released dataset).
LarNO models large-scale urban flooding with a latent autoregressive neural operator — predicting water-depth maps over millions of grid cells (~100 km²) at O(mm)-level accuracy and ~940× faster than the MIKE+ hydraulic solver. Trained at low resolution, it generalizes zero-shot to higher resolutions (e.g. 8 m → 2 m, 20 m → 5 m) without retraining, and transfers to unseen catchments via few-shot fine-tuning.
🧪 Want to try LarNO instantly — no installation needed? Open our Google Colab notebook to run inference on the Futian flood dataset with pre-trained weights in ~15 minutes, entirely in your browser.
📚 Want to reproduce the paper? The complete step-by-step tutorials live in
tutorials/(English & 中文).
- LarNO introduces a latent autoregressive neural operator for zero-shot, high-resolution spatiotemporal generalization in urban flood modeling.
- Latent autoregression improves the representation of nonlinear spatiotemporal flood dynamics.
- LarNO enables large-scale flood forecasting over millions of grid cells (~100 km2) and sub-billion spatiotemporal points (5 m and 5 min resolution).
- LarNO achieves O(mm)-level water-depth accuracy.
- LarNO supports few-shot transfers to unseen catchments via fine-tuning.
- LarNO supports multi-GPU distributed training and TensorRT-accelerated inference.
- [15/05/2026] 🎉🎉🎉LarNO is now online in Journal of Hydrology! Free access is available until July 17, 2026.
- [12/03/2026] Pre-trained weights and benchmark dataset published on HuggingFace — download without Google Drive or Baidu Cloud.
- [12/03/2026] Interactive demo released — run LarNO inference in your browser with Google Colab, no installation needed.
- [02/03/2026] Full end-to-end reproduction tutorial released — train and test LarNO with a single GPU on AutoDL.
- [02/03/2026] Code released on GitHub and benchmark dataset publicly released.
Memory-efficient neural operator training and zero-shot generalization to high resolutions for urban flood spatiotemporal forecasting.
(a) Overview: a neural operator is trained at low resolution and applied zero-shot at higher resolutions by directly evaluating the learned continuous operator on finer grids.
(b) Training stage: discretized inputs (rainfall, terrain, drain inlet) are fed to the neural operator, with outputs supervised by a numerical solver (MIKE Plus).
(c) Zero-shot application of the trained operator at higher spatial resolutions without any retraining.
LarNO architecture for urban flood spatiotemporal forecasting.
The model comprises three stages: (1) a lifting layer maps the input to a higher-dimensional hidden state; (2) N LarNO layers iteratively update the hidden state — each layer first applies a GRU-based convolutional update combining the previous time-step and previous-layer hidden states, then refines the state via frequency-domain Fourier transforms (forward FFT, low-mode linear mixing, inverse FFT) and a local time-domain linear operator; (3) a projection layer maps the final hidden state to the output water depth.
|
No local GPU? Try in your browser. The Colab notebook runs inference on the Futian dataset with pre-trained weights in ~15 min — no installation, no dataset download. |
Or run locally in 3 steps (full setup in the tutorials):
# 1. Clone & install (details: tutorials/en/01-installation.md)
git clone https://github.com/holmescao/LarNO && cd LarNO/code/urbanflood_larfno
pip install -e . && pip install -r requirements.txt
# 2. Download the dataset (tutorials/en/02) and pre-trained weights (tutorials/en/03), then:
# 3. Run inference — results in exp/<timestamp>/
python test.py --config urbanflood_config_2d.yaml --expr_id 20260220_183648_006352Full reproduction tutorials live in tutorials/ — available in English and 中文. Recommended order: Setup (1 → 3) → Inference (4) → Training (5).
| I want to… | Guide |
|---|---|
| Install the environment | 1. Installation |
| Get the datasets | 2. Dataset Preparation |
| Get pre-trained weights | 3. Pre-trained Weights |
| Run inference & read outputs | 4. Inference, Evaluation & Outputs |
| Train (fine-tune / scratch) | 5. Training |
| Use a rented cloud GPU | 6. Cloud GPU — AutoDL |
| Look up config, layout & outputs | 7. Reference |
| Troubleshoot | 8. FAQ |
This project is released under the MIT License.
If you use LarNO in your research, please cite the Journal of Hydrology article (DOI: 10.1016/j.jhydrol.2026.135686):
@article{cao2026large,
title={Large-scale urban flood modeling and zero-shot high-resolution generalization with LarNO},
author={Cao, Xiaoyan and Yao, Yao and Wang, Zhi and Zhao, Zhangxinyue and Borthwick, Alistair GL and Qin, Huapeng},
journal={Journal of Hydrology},
pages={135686},
year={2026},
doi={10.1016/j.jhydrol.2026.135686},
publisher={Elsevier}
}If you use the released benchmark dataset, please also cite:
@article{cao2025bench,
author = {Cao, Xiaoyan and Qin, Huapeng},
title = {Benchmark dataset of ``Large-scale urban flood modeling and zero-shot high-resolution generalization with LarNO''},
year = {2025},
url = "https://figshare.com/articles/dataset/Benchmark_dataset_of_Large-scale_urban_flood_modeling_and_zero-shot_high-resolution_generalization_with_LarNO_/30529031",
doi = {10.6084/m9.figshare.30529031.v4}
}If your work involves high-resolution spatiotemporal nowcasting of urban flooding, you may also be interested in our related work U-RNN, which focuses on urban flood nowcasting at high spatial-temporal resolution and was published in Journal of Hydrology:
@article{cao2025u,
title={U-RNN high-resolution spatiotemporal nowcasting of urban flooding},
author={Cao, Xiaoyan and Wang, Baoying and Yao, Yao and Zhang, Lin and Xing, Yanwen and Mao, Junqi and Zhang, Runqiao and Fu, Guangtao and Borthwick, Alistair GL and Qin, Huapeng},
journal={Journal of Hydrology},
pages={133117},
year={2025},
publisher={Elsevier}
}