LarNO-dataset — Urban Flood Benchmark Dataset

Paper: Large-scale urban flood modeling and zero-shot high-resolution generalization with LarNO Journal: Journal of Hydrology, 2026 DOI: 10.1016/j.jhydrol.2026.135686 GitHub: holmescao/LarNO Model weights: holmescao/LarNO

---

Dataset Description

Two benchmark datasets for training and evaluating urban flood spatiotemporal forecasting models. Labels (water depth fields) were generated by MIKE+, a commercial hydrodynamic solver.

Dataset	Tag	Resolution	Domain	Events
UKEA (UK Environment Agency)	ukea_8m_5min	8 m / 5 min	~0.4 km²	8 train + 12 test
Futian, Shenzhen	region1_20m	20 m / 5 min	~100 km²	64 train + 16 test

Note: The paper used 5 m resolution (1600 × 2240 grid) for zero-shot super-resolution experiments. The released dataset is a 20 m downsampled version (400 × 560) for accessibility. Drainage network data is confidential and not included in the released dataset.

---

File Format

Each event is stored as a directory:

flood/<dataset_tag>/<event_id>/
    rainfall.npy    # shape (T, H, W), mm/5min
    h.npy           # shape (T, H, W), metres — water depth (MIKE+ output)
    dem.npy         # shape (H, W), metres — digital elevation model; NaN = buildings

Field	Shape	Unit	Description
rainfall	(T, H, W)	mm/5min	Dynamic rainfall forcing
h	(T, H, W)	m	Water depth (ground truth)
dem	(H, W)	m	Static terrain; NaN marks buildings

• T = 72 (6 hours × 12 steps/hour)
• UKEA: H=50, W=120
• Futian (region1_20m): H=400, W=560

---

Input Features Used by LarNO

The model uses 13 input channels per time step:

#	Feature	Shape	Description
1–6	rainfall window	(6, H, W)	Last 6 rainfall steps
7–12	cumulative_rainfall	(6, H, W)	Cumulative sum of rainfall
13	dem	(1, H, W)	Static terrain

Drainage network data is not released (confidential). The model achieves strong results without it.

---

Quick Start

import numpy as np

# Load one event
rainfall = np.load("flood/region1_20m/event_001/rainfall.npy")  # (72, 400, 560)
h        = np.load("flood/region1_20m/event_001/h.npy")         # (72, 400, 560)
dem      = np.load("flood/region1_20m/event_001/dem.npy")       # (400, 560)

print("Rainfall max:", rainfall.max(), "mm/5min")
print("Water depth max:", h.max(), "m")
print("DEM range:", dem[~np.isnan(dem)].min(), "–", dem[~np.isnan(dem)].max(), "m")

For full training/inference pipeline, see the GitHub repo.

---

Citation

@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}
}