An intelligent framework for spatiotemporal simulation of flooding considering urban underlying surface characteristics

•An intelligent flooding modeling framework considering urban geospatial data.•Demonstrates the potential of applying GeoAI techniques in geo-environmental.•The framework significantly reduces the process of repeated model simulations.•Integration of knowledge and machine learning in flood disaster management. In current urban flood modeling, challenges arise from the inadequate consideration of heterogeneous underlying urban surface characteristics and the complexity of parameter optimization processes. This study integrates multiple machine learning methods to propose an intelligent framework for urban flood modeling that accounts for underlying surface characteristics. It began by coupling a surface runoff model with a pipe network model to form an interpretable flooding model (FM). Subsequently, utilizing the BIC-GMM machine learning method on the sample set of the urban flooding model’s parameters, this study explored the grouping trends of these parameters. The Transformer model was employed to classify different categories of urban land use, which, along with other environmental indices, aided in the construction of an Artificial Neural Network (ANN) model. This model expedites the acquisition of the sensitivity parameters. The study also proposes urban functional zoning rules incorporating “socio-driven-nature-assisted” characteristics of the underlying surface. Finally, the clustering feature thresholds of the sensitive parameters were distributed across various catchment units based on the urban functional area distribution rules. This distribution was used to select multiple observed rainfall-runoff events to determine the optimal parameters of the inundation model, culminating in the construction of the BIC-GMM-Transformer-ANN-flooding model (BGTA-FM). The experimental results indicated that the method based on this intelligent modeling framework reached a mean Nash-Sutcliffe efficiency coefficient (NSE) of 0.8. This performance represents a 0.3 and 0.15 increase in NSE compared to the Transformer-ANN-flooding model and BIC-GMM-flooding model, respectively, and significantly enhances modeling efficiency. This effectively reflects the complex underlying surface environments of the research area. Our work demonstrates the substantial potential of integrating physical knowledge with machine learning in urban flood intelligent modeling and reaffirms the critical role of applying geospatial artificial intelligence (GeoAI) in geo-environmental resear