Figure 1 provides a geographical overview of the study area. The BTH region comprises Beijing Municipality, Tianjin Municipality, and Hebei Province, covering a total area of 218,000 km². Situated between 113°27′E and 119°50′E and spanning from 36°05′N to 42°40′N, the study area features a complex and diverse topography, including mountains, hills, basins, and plains. Bordered by the Taihang Mountains to the west and opening onto the Bohai Sea to the east, it drains into the sea, which functions as the ultimate hydrological sink for regional runoff (Hou et al., 2024). Elevation in the region exhibits a distinct decreasing trend from northwest to southeast. The northern and western parts of Hebei are dominated by mountainous and hilly terrain with pronounced relief, where average elevations exceed 1,000 meters. In contrast, the southeastern Hebei region and the municipalities of Beijing and Tianjin constitute the core of the Haihe Plain characterized by flat topography, an average elevation below 50 meters, and intensive agricultural activity and population concentration. The region is characterized by a well developed drainage system. The Luanhe and Haihe River systems form a fan shaped river network oriented from northwest to southeast, encompassing major tributaries such as the Yongding, Daqing, Chaobai, and Beiyun Rivers. Scattered throughout the plain are numerous depressions, lakes, and wetland complexes, contributing to a highly heterogeneous surface hydrological regime (Zhang and Wang, 2024). The region experiences a warm temperate semi-humid monsoon climate, with precipitation markedly uneven in both space and time. The majority of rainfall occurs during the flood season (June–September), accounting for 70–80% of the annual total, predominantly in the form of intense rainstorms. In recent years, the frequency of extreme heavy rainfall events has increased significantly (Yuan et al., 2018).

The BTH region, characterized by high population density and rapid economic development, faces significantly heightened flood risks due to its intensive urbanization. As of 2023, the region’s permanent resident population reached approximately 109 million, with the urbanization rate increasing to 79% over the past decade and GDP nearly doubling during the same period (Zhou and Zhao, 2025). This dense development pattern has rendered the region particularly vulnerable to historical extreme rainfall events, including those labeled “63.8,” “96.8,” “21.7,” and “23.7.” The “23.7” flood event alone resulted in economic losses exceeding 160 billion yuan and affected more than 5.5 million people, underscoring the pressing need for effective flood control and disaster mitigation strategies.

To construct a training dataset for flood-susceptibility modeling, multi-source archival flood records were collated to generate sample locations. Systematic extraction of geographic coordinates was conducted from flood events documented on municipal government portals of Hebei Province, the municipalities of Tianjin and Beijing, and supplemented by historical annals of the Haihe River Basin, yielding 2000 geo-referenced flood points. Positional accuracy was verified by GPS field surveys and cross-validation against water-level gaging stations, with residual mislocations eliminated through high-resolution remote-sensing interpretation. To mitigate class imbalance inherent in machine learning algorithms, non-flood points were delimited by GIS-based cluster analysis and then randomly sampled in equal number from buffer zones peripheral to flood pixels, producing an initial pool of 4,000 samples. Finally, a stratified random split (70% training, 30% validation) was applied to ensure proportional representation of both classes.

Flood disasters are triggered by the synergistic interaction of numerous factors, which can be broadly classified into four main categories: topographic and geomorphological features, meteorological and hydrological conditions, underlying surface properties, and human activities (Hou et al., 2018; Zhang, 2016). As fundamental expressions of surface configuration, topography and geomorphology capture regional-scale terrain heterogeneity. These attributes critically govern the velocity and persistence of rainfall derived runoff, thereby modulating flood susceptibility. Meteorological and hydrological variables act as direct triggers of flood episodes, where precipitation produces surface runoff that converges into river networks, culminating in flood events. The underlying surface is mainly defined by vegetation coverage and soil characteristics, whereas land use configurations offer a clear reflection of anthropogenic impact.

Building upon the distinct environmental setting of the BTH region, historical flood inventories, and flood formation mechanisms, and supported by existing literature (Pham et al., 2021; Prasad et al., 2022), this research selects 15 representative flood-conditioning factors to construct a preliminary factor inventory. Multi-source datasets were utilized in this research. From the 90-meter resolution SRTM Digital Elevation Model (DEM) obtained from the Geospatial Data Cloud,¹ the following factors were extracted: elevation, slope, aspect, Plan Curvature (PlaC), Profile Curvature (ProC), Stream Power Index (SPI), Topographic Wetness Index (TWI), and Terrain Ruggedness Index (TRI). Lithology data were derived from the 1:2,500,000 scale vectorized geological map provided by the Resource and Environmental Science Data Platform.² Land use data were sourced from the Global Land Cover Database,³ specifically the GlobeLand30 global land cover product with a spatial resolution of 30 meters. The Normalized Difference Vegetation Index (NDVI) data employed in this study were derived from the MOD13Q1 product—a standard vegetation index dataset released via NASA’s Earthdata platform,⁴ which is generated by the Moderate Resolution Imaging Spectroradiometer (MODIS) sensor onboard the Terra/Aqua satellites. This product is characterized by a spatial resolution of 250 m, a temporal resolution of 16 days, and a data time span ranging from 2000 to 2022. The raw MOD13Q1 data were preprocessed through a series of standard procedures, including image mosaicking, reprojection (unified to the WGS84 coordinate system), and spatial clipping to match the boundary of the BTH region. Subsequently, the Maximum Value Composite (MVC) method was applied to the preprocessed multi-temporal images to eliminate noise interference from clouds, atmospheric scattering, and solar altitude angle variations, ultimately generating the NDVI dataset of the BTH region covering the period 2001–2022. The multi-year average value was calculated to characterize the typical spatial pattern of vegetation cover in the study area. In terms of hydrological characteristics, based on the hydrological dataset provided by the National Geographic Information Resource Directory Service System,⁵ the Distance From Rivers (DFR) was calculated using the multi-ring buffer analysis method, and the River Density (RD) was extracted by means of hydrological analysis. Precipitation data, including the Annual Average Rainfall (AAR) and the 24-h Average Maximum Rainfall (AMR24), were sourced from the National Earth System Science Data Center,⁶ which were calculated based on a 20-year precipitation sequence. The spatial distribution patterns of these factors are illustrated in Figure 2.

To verify the rationality of the flood susceptibility probability zoning map generated by the machine learning model, the Sentinel-1A Synthetic Aperture Radar (SAR) imagery data downloaded from the Copernicus Data Centre of the European Space Agency⁷ were adopted in this study. The flood inundation extent of the study area was extracted based on the polarimetric water index, which was further used to evaluate the zoning results. The imagery were acquired in the Interferometric Wide-swath (IW) mode with the product type of IW_GRDH_1S, and all images featured a consistent ascending orbit direction as well as VV&VH dual-polarization configuration. The image acquisition period ranged from July 1 to September 5, 2023. The detailed image names, orbit numbers and acquisition times are provided in the Appendix Table 2.

As continuous predictors offer no direct insight into the relative contribution of individual factors to flood occurrence or into the spatial association between specific value ranges and flood likelihood, they were discretised into ordinal classes to enhance model interpretability and ensure the plausibility of the resultant assessment. Frequency Ratio (FR) converts continuous variables into categorical variables through discretization processing, which can effectively reveal the spatial correlation between each factor category and flood disasters (Bonham-Carter, 1994). The FR value is defined as the ratio of the probability of flood occurrence to the probability of non-occurrence within the study area. When FR > 1, it indicates a positive correlation between the classified interval and flood occurrence. A higher FR value signifies a stronger positive correlation, implying an increased likelihood of flooding in that interval (Akgün et al., 2008). The calculation formula of FR is as follows:

Among these parameters, N denotes the flood affected area within a specific classification interval, which in this study is approximated by the number of flood events occurring within that interval. N 0 represents the total flood-affected area across the entire study region and is substituted by the total number of flood events recorded. S refers to the spatial extent of the classification interval, while S 0 indicates the total area of the study region. The FR values for the 15 influencing factors are presented in Figure 3. Considering the uneven spatial distribution of flood data points, the application of equal-interval classification to influencing factors may lead to biased interpretations. To address this issue, this study employs the natural breaks method to classify the 12 continuous influencing factors (excluding aspect, land use type, and lithology) into five intervals, based on the inherent distribution patterns of the data.

The elevation of the study area varies from −38 to 2,760 meters (Figure 3A). Within the elevation interval of −38 to 227 meters, the FR exceeds 1, indicating a higher susceptibility to flood occurrence in this range. The slope gradient ranges from 0 to 87.92° (Figure 3B). Overall, the FR value decreases as slope increases, with the highest FR observed in the 0 to 4.82° slope interval. Aspect was classified into nine directional categories (Figure 3C), among which flat ( FR =1.84), north-facing ( FR =1.1), and west-facing ( FR =1.2) slopes exhibit a stronger association with flood occurrence. PlaC spans from −1 to 360 (Figure 3D), while ProC ranges from 0 to 56.22 (Figure 3E). According to FR analysis, the maximum FR value of 1.52 corresponds to the ProC interval of 0 to 1.98, whereas the PlaC interval of 251.61 to 287.81 yields the highest FR value of 1.12.

TRI refers to the height difference between the highest and lowest points within a certain area, which can quantitatively describe the undulating form of the terrain in a region. It is an important indicator for classifying landforms and characterizing their features (Chi et al., 2025).

The calculation formula is as follows:

In the formula, maxm and minm represent the absolute highest and lowest elevations within the region, respectively. The TRI value range of the region is 0 to 1942 (Figure 3F). The FR value decreases as the TRI increases, and within the range of TRI being 0 to 53, the FR value reaches its maximum(1.52), indicating that the smaller the TRI, the more prone the area is to flood disasters.

The SPI characterizes the erosive potential of water flow and quantitatively describes the relationship between water flow erosion and geomorphic evolution. Its calculation formula is as follows:

In the formula, A s represents the catchment area of the basin, and β is the slope.

The SPI values in the study area range from −11.41 to 18.32 (Figure 3G). Except for the intervals −11.41 to −4.53 and −4.53 to 1.3, where FR > 1, for other intervals, FR < 1. This indicates that within the range of −11.41 to 1.3, SPI is positively correlated with the occurrence of flood disasters.

TWI is used to predict the spatial distribution pattern of soil moisture and quantify the control intensity of topography on hydrological processes. Its calculation formula is as follows:

As the TWI value increases, the FR value shows an increasing trend. Within the range of 22.56 to 36.13, the FR value reaches its maximum of 2.06 (Figure 3H), indicating that the higher the TWI value, the more likely it is to cause flood disasters.

Based on rock strength characteristics, the lithologies in the study area are systematically divided into six groups (Figure 3I): UF represents fluvial sedimentary rocks, IA1 is granite, UL is lacustrine sedimentary rocks, UG is glacially derived rocks, SC2 includes sandstone, argillaceous sandstone, and feldspathic sandstone, MA1 is quartzite, and SC3 includes siltstone, mudstone, and claystone. Statistical analysis reveals that over 86% of the hazard sites in the study area are located within the UF lithology distribution, indicating that the lithologic characteristics of fluvial sedimentary rocks may significantly influence the development and occurrence of disasters.

River network density serves as a critical indicator for assessing the development level of a regional water system. It is mathematically defined as the ratio of the total river length to the area of the study region. Based on the three-level water bodies, a hierarchical buffer zone was constructed. By coupling the weighted assignment and spatial overlay techniques, the absolute distance from the river was normalized into a dimensionless relative index ranging from 1 to 10, precisely quantifying the spatial proximity of the region to the river network. A higher river network density is associated with an increased likelihood of flood disasters (Figure 3J). Furthermore, areas located closer to rivers face a greater risk of being affected by flood events (Figure 3K).

The annual average rainfall in the study area ranges from 343.07 to 759.2 mm (Figure 3L). The FR values exceed 1 within the rainfall intervals of 455.67 ~ 520.94, 520.94 ~ 576.43, and 576.43 ~ 644.96, with the maximum FR value observed in the 576.43 ~ 644.96 interval. Similarly, for the average 24-h maximum rainfall, FR values are greater than 1 in the intervals of 59.76 ~ 76.13, 76.13 ~ 87.47, and 87.47 ~ 101.32, where the highest FR value occurs in the 87.47–101.32 interval (Figure 3M).

The study area encompasses eight land use types: farmland, forest, grassland, shrubland, wetland, water body, impervious surface, and bare land (Figure 3N). Among these, the FR values for impervious surface, farmland, and bare land all exceed 1, with the impervious surface exhibiting the highest FR value of 2.71. The FR value of NDVI in the study area ranges from −0.18 to 0.97 (Figure 3O), and the FR value decreases with the increase of NDVI value. The smaller the NDVI value, the lower the vegetation coverage and the more likely it is to cause flood disasters.

The Receiver Operating Characteristic (ROC) curve serves as a critical analytical tool for assessing the performance of binary classification models. By plotting the false positive rate ( FPR ) on the horizontal axis and the true positive rate ( TPR ) on the vertical axis, it visually illustrates how model performance varies across different classification thresholds (He et al., 2024). The TPR, also known as sensitivity, quantifies the proportion of actual positive samples that are correctly identified by the model, whereas the FPR, mathematically defined as 1-specificity, measures the proportion of negative samples that are erroneously classified as positive. The AUC, which represents the total area beneath the ROC curve, ranges in value from 0 to 1. A higher AUC value indicates greater discriminatory power; an ideal classifier achieves an AUC of 1.0, while a random classifier typically yields an AUC close to 0.5.

In the formula, TP , TN , FP , and FN respectively represent the number of samples that are actually positive and predicted as positive, actually negative and predicted as negative, actually negative but predicted as positive, and actually positive but predicted as negative.

The Accuracy ( A ccuracy ), Sensitivity ( S ensitivity ) and Kappa ( K appa ) are incorporated to assess the overall correctness, the recall performance for the minority class, and the agreement beyond chance, respectively, thereby enabling a comprehensive and robust evaluation of model performance. The corresponding computational formulas are: presented as follows.

Prior to model development, a comprehensive multicollinearity diagnosis and information gain ratio assessment were conducted for all flood disaster influencing factors. As illustrated in Figure 5, the AAR factor exhibited significant multicollinearity (VIF = 10.06, Tol = 0.082), based on the criteria of variance inflation factor (VIF < 10) and tolerance (Tol > 0.1). Following its removal, the VIF values of all remaining variables fell below 10, indicating a substantial reduction in multicollinearity. Further IGR analysis revealed that the AAR contributed minimally to the model (IGR = 0.0257), while PlaC demonstrated no predictive contribution (IGR = 0). To enhance model efficiency by mitigating multicollinearity and eliminating redundant variables, both the AAR and PlaC were excluded from the final set of influencing factors.

After optimizing hyperparameters using grid search and 5-fold cross-validation, the final parameter combinations for SVM, RF, XGBoost, and MLP are shown in Table 1. On the independent test set, the AUC of each model exceeded 0.85 (Figure 6), specifically: SVM 0.854, RF 0.920, XGBoost 0.938, and MLP 0.867. XGBoost performed best, with its AUC improving by 9.84% compared to the lowest-performing SVM. Further evaluation of model performance on the training and validation sets (Table 2) showed that XGBoost significantly outperformed other models in both Accuracy (0.981 on the training set, 0.917 on the validation set) and Kappa (0.962, 0.833), demonstrating the best overall performance. RF followed, while SVM and MLP performed relatively weakly, with SVM achieving only 0.808 Accuracy on the validation set. Furthermore, XGBoost exhibits the strongest generalization ability while maintaining the highest recall rate, with the smallest performance difference between the training and validation sets, further demonstrating the model’s advantages in stability and reliability.

Using the filtered raster data of flood disaster influencing factors in the BTH region along with their corresponding sample labels as input, each machine learning model generated a flood susceptibility index (FSI) for every raster cell. The FSI values range from 0 to 1, with higher values indicating greater flood occurrence probability. Based on the overall probability calculations, a 90 × 90 m resolution probability raster map was produced using ArcGIS. The probability values were classified into five susceptibility levels— very low, low, moderate, high, and very high—using the natural break method. Subsequently, flood susceptibility zoning maps for the four models were obtained (Figure 7). In these maps, black solid dots denote training set flood points and dark blue solid dots denote validation set flood points. The overlaid points illustrate that flood occurrences are predominantly concentrated in High and Very High susceptibility zones across all four model-generated maps, with minimal presence in low-susceptibility areas. This pattern aligns with the premise that floods occur predominantly in high-risk regions, supporting the rationality of the training process. The spatial consistency between predicted high-risk areas and observed flood locations across both datasets demonstrates the reliability of susceptibility predictions. Moreover, the validation set closely mirrors the spatial distribution of the training set despite being excluded from model fitting, confirming both the representativeness of the sample split and the generalizability of the models.

To quantitatively evaluate these maps, the area proportion, flood point proportion, and frequency ratio (FR) within each susceptibility level were statistically analyzed (Figure 8). Notable differences were observed in the spatial distribution of FSI values across the four models within each risk level (Figure 8A). The RF model predicted that very low, low, moderate, high, and very high risk areas accounted for 43.38, 10.96, 16.23, 13.16, and 16.27% of the total area, respectively. For the XGBoost model, the corresponding proportions were 46.62, 12.21, 16.01, 11.56, and 13.60%. The MLP model yielded area proportions of 46.01, 8.73, 14.55, 15.83, and 14.88%, while the SVM model produced values of 31.84, 20.08, 9.33, 10.88, and 27.87%, respectively.

Although the models produced varying zoning results, their spatial distribution patterns were largely consistent, revealing the general flood susceptibility characteristics of the BTH region: (1) Very high and High susceptibility zones are primarily located in terrain transition zones, key hydrological nodes, and major flood detention areas. These include Beijing’s Fangshan and Mentougou Districts; central and eastern parts of Baoding City in Hebei Province (Qingyuan District, Mancheng District, Zhuozhou City, Gaobeidian City); Yongnian and Congtai Districts in Handan City; Xiqing and Jinnan Districts in Tianjin; and eastern parts of Xingtai City in Hebei Province. (2) Moderate susceptibility zones are concentrated in Jinghai and Wuqing Districts of Tianjin, eastern Cangzhou in Hebei Province, and Tongzhou District in Beijing. These areas serve as secondary disaster-bearing zones within the regional flood control system, influenced by residual upstream floods, local heavy rainfall, and reduced drainage capacity. (3) Very low and Low susceptibility zones are predominantly found in the high-altitude mountainous regions of the northwest, including northern Yanqing and Huairou Districts, northeastern Miyun District in Beijing; the Bashang Plateau in Zhangjiakou City; northern Chengde (Weichang and Fengning Counties) in Hebei Province; and the northern part of Jizhou District in Tianjin. These areas are characterized by steep topography, efficient runoff drainage, low rainfall frequency due to semi-arid climate, sparse river networks, and limited human exposure, resulting in significantly lower flood risk compared to lowland areas.

As illustrated in Figure 8B, the spatial distribution of flood disasters across different risk levels indicates that all four models achieved high predictive accuracy. Specifically, the Very high and high susceptibility zones identified by the RF, XGBoost, MLP, and SVM models encompassed 84, 86.3, 79.05, and 76.95% of historical flood disaster points, respectively. In contrast, the Very low and Low susceptibility zones accounted for only 5.2, 5.1, 5.5, and 6.95% of the disaster points, thereby validating the reliability and rationality of the model predictions. Among the four models, XGBoost demonstrated the highest capture rate of flood events in high-risk areas and the lowest misclassification rate in low-risk areas, indicating superior overall performance.

The FR values of flood disasters across different susceptibility levels serve as indicators of the predictive reliability of the susceptibility maps. As depicted in Figure 8C, the FR values for all four models exhibit a consistent upward trend with increasing susceptibility levels, demonstrating their effectiveness in differentiating flood susceptibility grades within the study area. Among the models, XGBoost achieved the highest FR value (4.73) in the extremely high susceptibility category, followed by RF (4.07), SVM (3.18), and MLP (2.59). These results indicate that the susceptibility map generated by the XGBoost model possesses the greatest predictive reliability and provides a more accurate representation of the spatial distribution patterns and clustering characteristics of flood disasters.

With the ongoing impacts of climate change and intensified human activities, the frequency of extreme precipitation events has increased, leading to a continuous rise in flood disaster risks. Scientific assessment of flood susceptibility has thus become an essential foundation for effective disaster prevention and risk mitigation strategies. Regional flood susceptibility mapping enables systematic identification of potential risk levels and spatial distribution patterns, providing critical scientific support for disaster risk management and planning. This study established a comprehensive 15-factor evaluation framework for the BTH region, incorporating variables related to topography, hydrology, soil properties, and land cover. Given the intercorrelation among certain factors and the relatively low predictive contribution of others, direct model construction without prior screening could result in unstable predictions and reduced model reliability. Therefore, the rational selection and optimization of influencing factors are essential for developing a robust and accurate flood susceptibility assessment model. To address this, the study integrated multicollinearity analysis with IGR screening to effectively remove redundant variables and identify key driving factors. This approach facilitated the construction of a more stable and precise flood susceptibility model. The results indicated that annual average rainfall exhibited significant multicollinearity with other variables. The IGR analysis results show that factors such as Elevation, RD, AMR24, and Slope have the greatest contribution, while SPI and PlaC have relatively minor influences. This result is consistent with the conclusions of existing studies (Tehrany et al., 2015a; Andaryani et al., 2021; Pham et al., 2021).

To scientifically identify high-risk flood areas and provide a reliable basis for disaster prevention planning, it is necessary to use the best-performing model to map susceptibility zones. This study evaluates the performance of four machine learning models—SVM, RF, MLP, and XGBoost in assessing flood susceptibility within the BTH region, with the objective of identifying the most suitable predictive approach. ROC curve analysis and statistical evaluation indices indicate that all four models achieved AUC values greater than 0.85 and Accuarcy exceeding 0.8, demonstrating their applicability for flood susceptibility assessment in the study area. The predictive performance ranking is as follows: XGBoost > RF > MLP > SVM. The validation set results show that XGBoost, relying on the gradient boosting ensemble mechanism, effectively suppresses overfitting while maintaining a high Accuracy of 0.917 through iterative optimization of residuals and regularization constraints. It achieves the best Sensitivity of 0.922 and a Kappa of 0.833, demonstrating a significant lead in overall performance. RF, using Bootstrap aggregation and random feature subspace strategies, has good robustness, but due to the lack of gradient optimization mechanism, its classification Accuracy (0.883) and Sensitivity (0.898) are slightly lower than those of XGBoost. SVM, when dealing with high-dimensional geographic data, is limited by the mapping ability of the kernel function and the structure of the samples, resulting in insufficient classification boundary fitting, with the lowest Accuracy (0.808) and Sensitivity (0.875) among the four. Although MLP has strong nonlinear fitting potential, due to the influence of the training sample size and the sensitivity of hyperparameters, the Kappa of the validation set (0.734) decreases by approximately 15.7% compared to the training set, reflecting a certain insufficiency in generalization ability.

The aforementioned results further underscore the importance of aligning algorithmic characteristics with the specific requirements of practical problems. Compared to single-model approaches, ensemble learning models demonstrate superior performance by effectively integrating the strengths of multiple base models. These models not only enable more precise identification and quantification of key driving factors, but also exhibit enhanced prediction accuracy and generalization capabilities in complex environments. This makes them particularly well-suited for disaster risk assessment in large-scale and high-risk regions (Shahabi et al., 2020; Zhu et al., 2024).

Although the four models differ in the proportion of risk-affected areas, their spatial distribution patterns of high-risk zones show overall consistency, further validating the reliability of the model assessment results. To further clarify the key influencing factors and their interactions in flood occurrence, this study evaluates the XGBoost model with the highest prediction accuracy. By calculating the correlation degree between the model and each influencing factor, we analyze the response sensitivity of these factors to flood events. The Spearman correlation coefficient method, which assesses correlations through variable rank order without requiring data to follow specific distributions, demonstrates advantages in analyzing nonlinear relationships (Lan et al., 2020). Calculations revealed that the Spearman non-parametric correlation coefficient vector between the XGBoost model and 13 influencing factors was (−0.822,-0.096,-0.504,-0.561,-0.205,-0.645, −0.137,0.000,0.569,0.417,-0.536,0.579, −0.692)^T. Elevation and TRI showed extremely strong negative correlations with risk, while TWI and AMR24 exhibited strong positive correlations. RD demonstrated moderate to low positive correlation with risk. Therefore, topographic conditions, water system development, and soil moisture saturation status are key influencing factors for flood disaster generation and incubation. Building on this foundation, further identification of interaction patterns and relationships among these key factors will help clarify the mechanisms of disaster susceptibility and causation in high-risk areas, thereby providing scientific basis for regional flood prevention and resilience-building strategies. Statistical analysis of corresponding indicators in high-risk zones reveals that low altitude (<227 m), gentle slopes(<4.82°),higher terrain humidity (TWI between 7.51–10.87),well-developed river networks, and heavy short-term rainfall collectively form the combination conditions prone to triggering flood disasters. These areas are predominantly pre-mountain alluvial plains, confluence zones of major water systems in middle-lower reaches, or flood control project-affected regions, characterized by weak natural drainage capacity, high proportions of impermeable surfaces, and susceptibility to human activities. Assessment results show good alignment between high-risk areas and the actual flooding conditions of the “23·7” flood event. Specifically, Fangshan District and Mentougou District, which recorded process rainfall of 1014.5 mm, served as extreme rainstorm centers, experiencing severe mountain floods and urban flooding (Zhao et al., 2024). The rapid discharge of floodwaters from the Yongding River’s Yanchi to Lugou Bridge section in the Taihang Mountains caused damage to the Lugou Bridge hub, riverbank protection structures, and multiple bridge foundation risks. In Baoding’s Zhuozhou City, areas along the Juma River and Dashih River were affected by excessive floodwaters from the Dashih River and Juma River in the Daqing River system, resulting in large-scale flooding and relocation efforts.

Notably, the Xiaoqing River Flood Storage Area in southern Tianjin has been activated during flood regulation, with a storage capacity of 9.6 km², playing a cr-ucial role in flood diversion and disaster mitigation. Therefore, systematic and adaptive resilience construction based on regional flood distribution patterns and disaster mechanisms is essential for enhancing overall disaster prevention efficiency and safeguarding public safety in socio-economic systems (Zhang et al., 2020). For instance, in urban areas like Beijing’s City Sub-center, Tianjin Binhai New Area, and Xiongan New Area, water storage spaces should be reserved during urban renewal while strengthening infrastructure resilience. This includes developing permeable green spaces and ecological drainage facilities, as well as advancing smart drainage network upgrades. In high-risk zones with dense farmland such as eastern Baoding and Handan, adaptive agriculture integrated with ecological protection should be promoted. This involves cultivating flood-resistant crops, improving farmland drainage systems, and exploring agricultural disaster insurance to enhance resilience against extreme weather. Additionally, in key ecological protection areas along the Yongding River, Daqing River wetlands, and Baiyangdian Lake, environmental protection must be reinforced while prohibiting occupation of flood channels. Measures like converting farmland to wetlands and constructing ecological shorelines should enhance flood control and storage capabilities. Flood risk boundaries should be clearly defined in territorial spatial planning, with preventive planning management implemented.

Through the coordinated promotion of flood control and disaster reduction engineering construction, facilities industry adaptation adjustment, ecological construction and storage function coordination, the regional flood disaster prevention and mitigation work is promoted to a more adaptive and resilient disaster prevention path.

This study has certain limitations in the selection of model input factors. While the current analysis primarily incorporates static variables related to terrain, hydrology, and land cover, it does not account for dynamic factors such as urban drainage system efficiency, real-time radar-based rainfall data, flood control infrastructure distribution, and human activity patterns. As a result, the model exhibits reduced accuracy in simulating extreme rainstorm scenarios, limiting its ability to capture the rapid interaction mechanisms between short-duration intense rainfall and urban flooding. Future research should aim to integrate multi-source datasets—including urban drainage infrastructure data, real-time precipitation monitoring, and dynamic human activity indicators—to develop a collaborative flood risk assessment framework capable of enhancing predictive performance under extreme weather conditions.