← Back to Datasets

LarNO – 2 — Case 2 · UKEA catchment (ukea_8m_5min) – zero-shot test

LarNO Dataset

Urban Flood Neural Operator Zero-shot Super-res Transfer Learning

2 Benchmark Cases

100 Total Events

72 Time Steps / Event

4× Zero-shot Super-res

About

This dataset is the official benchmark for LarNO, a memory-efficient, discretization-invariant neural operator for real-time urban flood prediction. LarNO learns the spatiotemporal mapping from dynamic rainfall and static terrain inputs to water depth distributions, and supports zero-shot generalization to spatial resolutions unseen during training.

The dataset comprises two urban flood benchmark cases:

(1) Futian case (region1_20m) — A large-scale benchmark covering the Futian District of Shenzhen, China (~100 km²). 80 observed extreme rainfall events (2008–2018, return periods 0.1–20 years) were simulated using MIKE Plus 2023 with a 1D–2D coupled scheme. Released at 20 m spatial resolution (400×560 grid), 72 time steps per event at 5-minute intervals. 64 events for training, 16 for testing.

(2) UKEA case (ukea_8m_5min / ukea_2m_5min) — A small coastal urban catchment (~0.4 km²) from a UK Environment Agency study site, used for transfer learning (fine-tuning from Futian pre-trained weights) and zero-shot super-resolution (train at 8 m, test at 2 m without any retraining). 20 events total; 16 for training, 4 for testing.

Each event is stored as a sub-folder containing rainfall.npy (T×H×W, mm/5 min) and h.npy (T×H×W, metres water depth). Shared terrain data (dem.npy, H×W) is provided per location.

Three training configurations are provided in the codebase: ukea_finetune.yaml, ukea_scratch.yaml, and region1_scratch.yaml. Full instructions are in the GitHub repository.

News

(2025.12) We have released the LarNO benchmark dataset on figshare, Google Drive, and Baidu Cloud!

Download

The dataset is publicly available from the following sources. All versions are identical in content.

Hugging Face no password Baidu Cloud password: LaNO Google Drive no password figshare no password

References

If you find this dataset useful, please cite:

@article{larno2025, title   = {Large-scale urban flood modeling and zero-shot high-resolution generalization with LarNO}, author  = {[TODO: authors]}, journal = {[TODO: journal]}, year    = {2025}, doi     = {[TODO: DOI]} }
@misc{larno2025dataset, author    = {Cao, Xiaoyan and others}, title     = {Benchmark dataset of Large-scale urban flood modeling and zero-shot high-resolution generalization with LarNO}, year      = {2025}, publisher = {figshare}, note      = {Dataset}, doi       = {10.6084/m9.figshare.30529031} }

Contact

For questions or suggestions, contact us at caoxiaoyan@stu.pku.edu.cn.