R02 → UPF-GAN: A Fast Spatially Explicit Generative Framework for Block-scale Urban Pluvial Flooding Risk Assessment
Keywords:#Urban pluvial flooding#Attention-based generative models#Surface water depth#Hydrodynamic simulation#High-density urban blocks# Year:2025 This eassay is urder submission.
A FAST spatially explicit generative framework
UPF-GAN is a deep generative model tailored for simulating urban pluvial flooding. Built on the GAN architecture and informed by hydrodynamic principles, it takes multi-channel images of urban form—including buildings, roads, green and blue spaces—as input, and produces surface water depth maps as output. Compared with physics-based models, UPF-GAN significantly reduces computation time and data requirements, enabling planners to rapidly evaluate flood-prone areas and test design alternatives at the block scale.
Problem
Most existing urban flood assessments rely on hydrodynamic models like MIKE 21, which, although physically robust, are computationally expensive and difficult to scale across multiple scenarios. On the other hand, conventional machine learning approaches offer faster inference but often ignore spatial continuity and fine-scale topographic influences. As a result, they fall short in capturing hydrodynamic behavior at the neighborhood or block scale.
METHODOLOGY
The UPF-GAN generator adopts an attention-augmented U-Net, integrating a Spatial-Channel Attention Assembly (SCAA) module and a pyramid-based feature fusion strategy. This allows the model to detect water-sensitive urban features like low-lying surfaces and impervious areas, while also synthesizing multi-scale spatial context. The model was trained on 1,149 urban blocks in central Shanghai, using MIKE 21 simulation outputs as ground truth, ensuring alignment with physical hydrodynamic behavior and terrain features.
MODEL PERFORMANCE
On the test set, UPF-GAN achieved a high prediction accuracy of R² = 0.96 and RMSE = 0.054, closely matching the physical simulation results. It runs 640 times faster than MIKE 21 and uses only 12% of its computational resources. Compared with baseline models like Pix2Pix and CycleGAN, UPF-GAN shows improved accuracy, visual clarity, and structural consistency—particularly in predicting high-depth flooding zones.
morphological insights
The model provides new insight into how urban form influences flood outcomes. Across three urban typologies, blue space consistently shows strong negative correlations with flood depth (r = –0.51 to –0.77), confirming its role in mitigation. Surprisingly, in high-density built-up areas, increased building coverage can reduce flooding risk (r = –0.41), possibly due to elevated terrain or optimized drainage. In contrast, fragmented green space with coverage below 10% may exacerbate flooding (r = 0.35), suggesting that spatial configuration is as critical as surface type.
Application
UPF-GAN not only provides rapid flood prediction within seconds, but also functions as an intelligent decision-support tool for urban planning. It enables early-stage testing of blue-green infrastructure strategies, resilience evaluation, and multi-scenario design. Especially in time-constrained or data-limited contexts, it offers a scalable and interpretable alternative to traditional flood models, advancing a new generation of morphology-risk evaluation systems.