This study proposes YuShi, a two-stage method that rapidly generates synthetic radar rainfall data for regions with no historical experience of torrential rain, thereby enabling preevent testing and training. In the first stage, a terrain- and meteorology-conditioned log-Gaussian Cox process probabilistically synthesizes the rainfall-core occurrence field. In the second stage, a diffusion-based UNet augmented with low-rank adaptation (LoRA) is used to create the contour mask and the intensity texture separately. A threefold validation using ten Japanese heavy-rain events (2018-2024) revealed that the hit rate of the generated cores was 1. 00 within a 25km tolerance, the wind direction error was less than 0. 05rad, and the nationwide spatial error (Wassersteinₐrea) was reduced to approximately 54% of the observed value, confirming high fidelity in direction and large-scale distribution. The remaining errors are concentrated along coastal regions, where high-intensity cells are overly smoothed and regional biases in the outer radius appear. These issues are expected to be alleviated by region-specific scaling of the occurrence intensity and by introducing terrain indices and wind convergence channels in the texture stage. These results provide synthetic rainfall fields that can serve as foundational inputs for future cyber-physical evaluations of the Disaster IT Testbed.
Hiroi et al. (Thu,) studied this question.