Benchmark dataset of "Large-scale urban flood modeling and zero-shot high-resolution generalization with LarNO"

This study focuses on the central urban region of Shenzhen, China (approximately 100 km2), where flood risk is high due to dense urbanization and frequent heavy rainfall events. We construct a benchmark dataset comprising large-scale, high-resolution, long-duration spatiotemporal dynamics of urban flood events induced by spatially heterogeneous extreme rainfall.We utilize urban geographic information with a 20-m spatial resolution to support street-level flood modeling. Then we collect 80 observed extreme rainfall events recorded in the study region between 2008 and 2018, which have return periods ranging from 0.1 to 20 years. All events were assigned a 6-hour duration to facilitate training and evaluation. We randomly select multiple rainfall centers to generate spatially heterogeneous extreme rainfalls, which is typical in urban regions spanning 100 square kilometers. One center uses an observed rainfall time series, while the other centers generate new rainfall time series by perturbing the original series with 50% Gaussian noise. For the remaining region, rainfall time series are generated using inverse distance weighting with a 1 km resolution. <br>To simulate urban flooding, we use MIKE Plus Version 2023, employing a 1D-2D coupled scheme to model the drainage system and surface runoff. Urban infrastructure, such as pipe diameters and drainage inlet configurations, is simplified based on available city planning data, retaining only the main pipes and nodes. We raise the terrain values at building locations by 100 m to simulate the impact of buildings on surface runoff and flood propagation. The model time step is adjustable between 0.01 s and 1 s to maintain numerical stability. We simulate 6-hour spatiotemporal water depth at a 5-minute frequency. We construct a benchmark dataset for urban flooding, comprising 80 events with spatial dimensions of 1,600 × 2,240 and a temporal dimension of 72. These events are randomly divided into a 4:1 training and testing set to evaluate the performance of neural operators on unseen events.