Sparse validation can mask process errors in urban flood modelling: evidence from large multimodal model-assisted video observations
Abstract
Urban flood models support drainage-system planning and flood risk management. However, due to the scarcity of monitoring data, model calibration and validation often rely on sparse observations, leaving the process fidelity of models insufficiently assessed. Here, we used large multimodal model-assisted video interpretation to derive minute-level ordinal inundation observations and evaluate the process fidelity of a sparsely calibrated urban flood model. In a case study of Dalian, China, a 1D–2D MIKE FLOOD model was calibrated using five historical peak-depth estimates and validated using three, then evaluated using 20,825 paired simulated and video-derived inundation levels. Direct comparison quantified errors in inundation occurrence, timing, severity, and duration, and two Bayesian cumulative link mixed models characterised their variation across flood phases and rainfall–site conditions. Results showed that although the model achieved low calibration and validation RMSEs of 0.107 and 0.084 m against peak-depth references, video observations revealed its insufficient process fidelity. Simulated onset was delayed by a median of 28.5 min where visible flooding was reproduced, and visible-inundation duration was a median of 29.5 min shorter than observed across all segments. The phase model estimated the highest underestimation probability of 73.6% during the observation-defined peak phase. Underestimation was also more likely with greater rainfall accumulation over the preceding 30 min, which had the strongest association among the rainfall and site covariates (conditional odds ratio = 6.48). These findings show that sparse peak-depth validation can leave errors in street-scale inundation processes undetected. Future studies should delineate the reliability boundaries of urban flood models using process-informed observations.
Related
- Are surface temperature and chlorophyll in a large deep lake related? An analysis based on satellite observations in synergy with hydrodynamic modelling and in-situ data
- Artificial neural network-assisted modelling of methylene blue dye adsorption onto ZIF-8 functionalized fishtail palm seed biomass: Isotherm and performance analysis
- Spatial modelling of potential rockfall occurrence in Romania using the Weight of Evidence method
- In-cloud motions shape warm-rain production in marine clouds: Evidence from multiyear observations
- Methodological Bias between Trajectory Geometric Co-location and Air-Parcel Lagrangian Co-location: Evidence from UAV Re-penetration Observations over the Western Sichuan Plateau
BibTeX
@misc{zhou2026sparse,
title = {{Sparse validation can mask process errors in urban flood modelling: evidence from large multimodal model-assisted video observations}},
author = {Shunan Zhou and Heng Lyu and Anette Etner and Pedro Zamboni and Nan Wang and Jiaxing Zhao and Chi Zhang},
howpublished = {EarthArXiv},
year = {2026},
month = sep,
doi = {10.31223/x50r5w},
url = {https://eartharxiv.org/repository/view/15127/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Sparse validation can mask process errors in urban flood modelling: evidence from large multimodal model-assisted video observations
AU - Shunan Zhou
AU - Heng Lyu
AU - Anette Etner
AU - Pedro Zamboni
AU - Nan Wang
AU - Jiaxing Zhao
AU - Chi Zhang
PY - 2026
DA - 2026/09/23
PB - EarthArXiv
DO - 10.31223/x50r5w
UR - https://eartharxiv.org/repository/view/15127/
AB - Urban flood models support drainage-system planning and flood risk management. However, due to the scarcity of monitoring data, model calibration and validation often rely on sparse observations, leaving the process fidelity of models insufficiently assessed. Here, we used large multimodal model-assisted video interpretation to derive minute-level ordinal inundation observations and evaluate the process fidelity of a sparsely calibrated urban flood model. In a case study of Dalian, China, a 1D–2D MIKE FLOOD model was calibrated using five historical peak-depth estimates and validated using three, then evaluated using 20,825 paired simulated and video-derived inundation levels. Direct comparison quantified errors in inundation occurrence, timing, severity, and duration, and two Bayesian cumulative link mixed models characterised their variation across flood phases and rainfall–site conditions. Results showed that although the model achieved low calibration and validation RMSEs of 0.107 and 0.084 m against peak-depth references, video observations revealed its insufficient process fidelity. Simulated onset was delayed by a median of 28.5 min where visible flooding was reproduced, and visible-inundation duration was a median of 29.5 min shorter than observed across all segments. The phase model estimated the highest underestimation probability of 73.6% during the observation-defined peak phase. Underestimation was also more likely with greater rainfall accumulation over the preceding 30 min, which had the strongest association among the rainfall and site covariates (conditional odds ratio = 6.48). These findings show that sparse peak-depth validation can leave errors in street-scale inundation processes undetected. Future studies should delineate the reliability boundaries of urban flood models using process-informed observations.
ER -