Auditing the Physical Fidelity of Urban Flood Model with Large Multimodal Model-Derived High-Resolution Observational Data

Characterized by high spatial heterogeneity and short response times, urban flood research has long been constrained by the scarcity of precise, high-spatiotemporal resolution observational data required to capture the complex dynamics. Without such data, it remains challenging to distinguish whether model simulation errors stem from uncertain parameters or fundamental structural deficiencies, obscuring the model's reliable physical representation. Consequently, the physical fidelity of urban hydrodynamic models in reproducing complex, spatiotemporal flood dynamics needs to be further validated. To address this, we constructed the first large-scale, 1-minute resolution urban surface inundation dataset derived from camera videos using a Large Multimodal Model (GPT-5). Using the observations, we audited the physical driving mechanisms of 1D-2D coupled hydrodynamic models by examining the spatial stratified heterogeneity of flood responses.Focusing on 5 diverse rainfall-flood events recorded by 226 traffic surveillance cameras in Dalian, China, we utilized GPT-5 to automatically extract real-time waterlogging levels. The extracted data underwent manual verification by experts to strictly correct errors, resulting in a high-fidelity urban surface inundation dataset at a 1-minute resolution. Subsequently, we constructed and calibrated a 1D-2D coupled urban flood numerical model to obtain simulation results for the corresponding events. The Geodetector model was then employed to quantify and compare the spatial stratified heterogeneity of waterlogging derived from observations versus simulations, attributing them to 9 potential drivers including rainfall, topography, and drainage infrastructure.Results indicate that GPT-5 achieved satisfactory extraction performance, with an average accuracy of 77%. Comparative Geodetector analysis of observations versus simulations revealed critical discrepancies. The factor detector showed low individual explanatory power (q