The calibration and validation of hydrological model simulation performance and model applicability evaluation in Gansu Province is the foundation of the application of the flash flood early warning and forecasting platform in Gansu Province. It is difficult to perform the simulation for Gansu Province due to the fact that it covers a wide range, from north to south, with multiple climate types and diverse landforms. The China Flash Flood Hydrological Model (CNFF) was implemented in this study. A total of 11 model clusters and 289 distributed hydrological models were divided based on hydrology, climate, and land-use factors, among others. A spatiotemporally mixed runoff method and the Event-Specific Geomorphological Instantaneous Unit Hydrograph (GIUH) were applied based on large-scale fast parallel computation. To improve model calibration and validation efficiency, the RSA method (Regionalized Sensitivity Analysis) was used for CNFF model parameter sensitivity analysis, which could reduce the number of model parameters that need to be adjusted during the calibration period. Based on the model sensitivity analysis results, the CNFF was established in Gansu Province to simulate flood events in eight representative watersheds. The average NSE, RE_Q, and E_T were 0.76 and 0.73, 9.1% and 12.6%, and 1.2 h and 1.7 h, respectively, in the calibration and validation period. In general, the CNFF model shows a good performance in multiple temporal and spatial scales, thus providing a scientific basis for flash flood early warning and analysis in Gansu Province.

Flood disaster risk assessment serves as an important foundation for formulating regional sustainable development strategies. This study establishes a risk assessment model for flood disasters in small and medium-sized rivers based on a theoretical framework integrating Geographic Information Systems (GIS), the Delphi method, and Bayesian networks (Delphi–BNs). An indicator system for the assessment was developed from three dimensions: hazard, vulnerability, and exposure. Hazard is represented by flood inundation area and depth; vulnerability is indicated by population distribution and economic layout; and exposure is reflected by road accessibility. By constructing a GIS-based Bayesian network and employing the Delphi method to create a probabilistic and spatially explicit model, this approach quantifies various sources of uncertainty in the assessment process, enabling a probabilistic expression of risk. Based on the risk assessment results, a stratified, phased flood emergency rescue and personnel transfer plan was established, designating extremely high-risk areas as the core zones for the first phase of personnel transfer, high-risk areas as the second-phase rescue zones, and medium-risk areas as the third-phase rescue zones, thereby providing clear operational guidance for flood emergency response in the basin. The Delphi–BN assessment framework developed in this study focuses on the core elements of flood disaster risk formation, organically integrates expert experience with spatial big data, and effectively overcomes the limitations of traditional assessment methods, such as strong subjectivity, insufficient accuracy, and poor quantification. It achieves a refined and quantitative assessment of flood risk in small and medium-sized river basins in semi-arid regions. The outcomes of this research contribute to a clearer understanding of both the driving mechanisms and the spatial patterns of regional flood risk. Furthermore, they establish a scientifically credible and operationally relevant foundation for key disaster-response decisions, encompassing timely emergency actions, phased population transfers, and the optimized deployment of limited emergency resources.