A Multi-Site Spatiotemporal Stochastic Rainfall Generator for Realistic Rainfall Generation

To enhance flood risk assessment and management, particularly in catastrophe models used for estimating potential losses, generating realistic extreme rainfall scenarios is crucial for accurate flood mapping. Current probabilistic rainfall models primarily focus on either single-station analysis or the spatial dependence of static rainfall properties. However, these approaches often fail to capture the dynamic spatiotemporal characteristics of short-duration extreme rainfall events—the primary drivers of urban flooding. In this study, we propose a novel spatiotemporal stochastic rainfall generator to simulate rainfall sequences at multiple stations while preserving spatial correlations and realistic temporal dynamics at the same time.

The generator is calibrated and applied in Hong Kong, a densely urbanized and flood-prone coastal city, using hourly in-situ observations from 141 stations for 1984–2017. Although operating at hourly resolution, the model consistently reproduces rainfall statistics across 1–24 h accumulation durations. It closely matches the statistical characteristics of historical rainfall, achieving Nash–Sutcliffe efficiency (NSE) values of 0.939–0.969 for the top 10% of events, and captures the spatial patterns of extremes with a Pearson correlation of 0.831.

Hydrodynamic simulations further demonstrate that the realistic temporal variability produced by the proposed generator leads to average flood depth differences of 18.1% and 25.8% compared with the simplified exponential and constant hyetograph scenarios, respectively.  Overall, the results underscore the importance of representing realistic short-term rainfall variability in stochastic rainfall modeling to support robust flood risk assessment.