R07 →Applying UPF-GAN to Weiyuan Island: Flood Resilience Mapping at the Urban Block Scale
Keywords:#UPF-GAN#Weiyuan Island#Block Scale#Deep Learning Application#Morphological Resilience# Year:2025 This is the content of Chapter 7 of the master's thesis.
This chapter demonstrates how the UPF-GAN model can be applied to evaluate and optimize urban morphological design proposals in a real-world context. Based on representative blocks in Weiyuan Island, different combinations of urban morphology and elevation were input into the model to simulate flood depth distribution under extreme rainfall conditions. The model outputs were further used to calculate the proportion of areas exceeding flood safety thresholds and to evaluate resilience indicators including resistance, marginal loss, and extreme adaptation.
Flood Scenarios: P50 and P100 Rainfall Conditions Two rainfall scenarios were simulated: 50-year and 100-year return periods, both with 120-minute durations. The input features were derived from the design scheme’s spatial layers, and the predicted inundation maps were used to calculate three resilience indicators: Resist, Mloss, and Aextreme.
Predicting Resilience Across Urban Blocks
UPF-GAN effectively captured the spatial differentiation of flood risk across various morphological sub-blocks. The results revealed strong variations in predicted inundation and resilience scores, with building layout and green space continuity emerging as major influencing factors.
EValuating Predictive Reliability Through Key Samples
Several representative blocks were selected to compare UPF-GAN outputs with MIKE21 benchmarks. In both flood depth and resilience estimation, the generative model demonstrated high consistency with numerical results, validating its potential in urban-scale planning applications.
Interpreting Urban Form–Resilience Relationships
SHAP-based analysis was performed on UPF-GAN outputs to identify which spatial features most influenced flood resilience. Results confirmed that key morphological variables such as building density (DB), green space fragmentation (GDI), and water surface ratio (WP) had consistent explanatory patterns with those derived from physical simulations.
Toward Integration of AI Models in Urban Design Evaluation
This application demonstrates that UPF-GAN can be embedded in real-world urban morphology evaluation to deliver fast, interpretable, and spatially refined flood resilience predictions. The model can support iterative design assessment and spatial strategy optimization for flood-adaptive urban development.