The Global Flood Awareness System (GloFAS, http://www.globalfloods.eu/, last access: 19 August 2026) couples the surface and sub-surface runoff from the Hydrology Tiles ECMWF Scheme for Surface Exchanges over Land (HTESSEL) land surface model, driven by the ECMWF's latest global atmospheric reanalysis (ERA5), with the LISFLOOD hydrological and channel routing model (Harrigan et al., 2020). The GloFAS-ERA5 river discharge reanalysis product is a global gridded available from 1979 to the present at a spatial resolution of 0.1° and a daily time step. The dataset is openly available through the Copernicus Climate Change Service (C3S) Climate Data Store (CDS) at https://ewds.climate.copernicus.eu/datasets/cems-glofas-historical?tab=download (last access: 10 May 2026). Harrigan et al. (2020) assessed the hydrological performance of GloFAS-ERA5 against observations from global in-situ river discharge stations; they used modified Kling–Gupta Efficiency Skill Score (KGESS), and found that the GloFAS-ERA5 reanalysis product is skilful in 86 % of catchments. Thus, the long temporal coverage, global consistency, and reliable simulation skill make GloFAS-ERA5 a valuable dataset for identifying extreme hydrological events such as floods. In this study, we collect GloFAS-ERA5 daily river discharge data of the rainy seasons (i.e., May–October) during 2000–2023.

The GloFAS-ERA5 river discharge reanalysis is produced by coupling the land surface model runoff component of the ECMWF ERA5 global reanalysis with the LISFLOOD hydrological and channel routing model (Harrigan et al., 2020). The LISFLOOD model is forced with daily HTESSEL surface and subsurface runoff from ERA5. In order to be consistent with the GloFAS-ERA5 river discharge, the precipitation and surface runoff fields from ERA5 were selected to analyse the spatiotemporal evolution patterns of rainstorm events and runoff generation processes over mainland China with high spatial (0.25°) and temporal (1 h) resolution covering the period 2000 to 2023. As one of the latest high-resolution climate reanalysis datasets, ERA5 is a global atmospheric reanalysis product developed by ECMWF using the 4D-Var data assimilation technique (Hersbach et al., 2020). Compared with other reanalysis datasets, such as MERRA-2, NCEP/NCAR, and JRA55, the ERA5 performs better for precipitation and extreme precipitation over China, although its performance is not as good as that of satellite-based data (e.g., the Integrated Multi-satellite Retrievals for Global Precipitation Measurement (GPM) (IMERG) precipitation product) (Jiang et al., 2023). Note that the ERA5 underestimates the frequency of occurrence and interannual variability of mesoscale convective systems (MCSs) (Alpizar et al., 2026). The ERA5 reanalysis precipitation product used in this study is accessible in C3S Climate Data Store (https://cds.climate.copernicus.eu/datasets, last access: 10 May 2026).

Historical flood damages data in 31 provincial administrative regions in China from 2000 to 2023 were obtained from the Meteorological Disaster Yearbook (MDY) in China. The data primarily include the flood records (i.e., flood location, affected area, and date of the event) as well as corresponding direct economic losses, the number of people affected, the number of deaths, and the affected cropland area. The MDY of China is currently the most comprehensive freely available source for accessing the flood records in China, and has been widely used in previous studies (Li et al., 2012; Wei et al., 2018; Shi et al., 2020; Wang et al., 2025). Unfortunately, this dataset has not been updated since 2020, thus, we supplemented flood disaster data from 2020 to 2023 using news reports and government sources by searching for the keywords “flood” and “inundation” online. All supplemented flood disaster data are derived from official released records, which guarantees their authenticity. Notably, although rainstorm flood records were collected from multiple sources, they represented authentic and valid disaster records and could accurately reflect the impacts caused by flood disasters.

A total of 1041 flood disaster records were obtained from MDY. The spatial-temporal distribution of these flood occurrence records across mainland China from 2000 to 2023 is depicted in Fig. 1. The accumulated flood occurrences frequency generally shows a decreasing trend from southeast to northwest (Fig. 1a), which is consistent with the distribution pattern of annual precipitation in China (Ma et al., 2015). The temporal variation of flood records in Fig. 1b shows that more flood disasters occurred during 2002–2010.

Meteorological, environmental, and anthropogenic factors are the primary drivers of flood occurrence in any region. Among meteorological conditions, rainfall is typically the primary cause of flood disasters. In this study, the 3D precipitation features (including accumulated area, accumulated magnitude, lifespan, and moving distance; presented in detail in Sect. 3.3) are selected to determine their influence on the spatial distribution of flood disasters. The environmental factors are important components of the disaster-forming environment (Ke et al., 2025). We selected cropland, normalized difference vegetation index (NDVI), slope, and elevation difference to represent cropland area, vegetation coverage, gradient of the land surface, and elevation variations, as the environmental factors in this study. Additionally, human activity substantially modifies the natural environment and influencing flood disasters (Guan et al., 2021). Based on the existing research (Wu et al., 2021; Hoang and Liou, 2024), the population density and gross domestic product (GDP) were selected as important indicators to reflect human activity. All environmental and human activity factors were obtained from the Resource and Environment Data Cloud Platform (http://www.resdc.cn, last access: 19 August 2026), except for cropland, which was sourced from 10.5281/zenodo.7936885 (Tu et al., 2023, 2024). The original data were interpolated with the same spatial resolution of the precipitation with 8 km. In addition, it is worth noting that the population and GDP data were summed within grids cells. However, due to the lack of annual dataset, population and GDP data from 2000–2024 at 5-year intervals are included here. A summary of all variables and their data sources is presented in Table 1.

Following the definition of short-duration heavy rainfall events by the China Meteorological Administration (CMA), an hourly rainfall amount exceeding 16 mm is commonly defined as a heavy rainfall event, which has been validated by Zhang and Zhai (2011) to be a reasonable criterion to study short-duration heavy rainfall events in China. However, the significant spatial-temporal variations may occur in extreme events as a result of varying geographical and meteorological conditions and consequently, the threshold of extreme precipitation events varies across regions. For instance, hourly extreme precipitation exceeding 16 mm h⁻¹ is relatively more frequent over the southern China, but it would be rather rare in the northwestern China (Fig. 2b). Thus, a fixed threshold cannot be used to extract relatively independent extreme precipitation events over China. In previous studies, most existing studies defined the extreme precipitation events using the 90 % or 95 % percentile of the ordered precipitation sequence during the study period because of its simplicity (Gu et al., 2022). In this study, the 95 % percentile of precipitation sequence at each grid was defined as the thresholds for extreme events (Fig. 2c). Hourly precipitation amounts exceeding the extreme precipitation threshold were classified as extreme precipitation events. Figure 2d and e illustrate the spatial distribution of the amount and accumulated frequency of extreme precipitation, respectively. Overall, extreme precipitation is observed across China, with higher frequency is in the southeast than in the northwest. Subsequently, these threshold values were further applied to better extract the characteristics of contiguous extreme precipitation events.

Historical flood damage sample datasets record flood events at country level in each province and provide the location and date of each flood event from 2000 to 2023. However, a single flood-causing precipitation (FCP, it is an extreme precipitation event that triggers flood disasters at the regional scale.) event mostly occurs synchronously across neighbouring regions or consecutively on adjacent days, leading to multiple flood records generated by one continuous precipitation process. In this study, we propose a precipitation event identification algorithm to identify relatively independent extreme precipitation events corresponding to each recorded flood disaster. The main steps of the precipitation event identification algorithm are described as follows (Fig. 3):

Determining the start and end times of flood-causing precipitation event: For each flood event (Ei), the approximate occurrence time (ti) is obtained from the flood disasters dataset. Based on the previous studies, the lifespan of an FCP event is generally no longer than 3 d, and it is rare for such events to last more than 7 d (Wang et al., 2022). Thus, in order to detect all potentially associated persistent FCP events, we extract the FCP identification period for the 5 d before and after ti. Subsequently, we use the vector shapefile of the flood-affected areas (li) as a spatial mask to isolate ERA5 precipitation within the affected area during identification period. Based on the precipitation time-series, we can accurately determine the start and end time of FCP event.

For each contiguous FCP event identified by the previous step, the 3D structural properties can be characterized by the following multidimensional metrics, and shown in Fig. 4a–b, consistent with previous studies (Wang et al., 2022a, 2025):

Accumulated magnitude: The sum of the precipitation amount over all grid cells across all hours of the FCP event.

Based on the accumulated affected area of FCP events presented in Sect. 3.3, the vector shapefile of the flood-affected areas is utilized as a spatial mask to extract hydrological, environmental, and anthropogenic factors within FCP affected areas (Fig. 4c–e). In this study, eight factors are selected to quantify the relationship between extreme precipitation characteristics and flood disasters across China. Detailed calculation methods are summarized in Table 2.

The geographical detector model is a set of statistical methods to explore spatial heterogeneity of explanatory variables, reveal interactions between driving factors, and quantify contribution rate of individual drivers (Wang and Hu, 2012). The formulas for ecological detection, risk detection, interaction detection, and factor detection were given in Wang et al. (2010). The core principle of factor detection is to determine whether the independent variable can explain the dependent variables and to investigate whether the interaction of two independent variables enhances, diminishes, or has no effect on the dependent variable. The results of interaction detector between two independent variables are classified into five types shown in Table 3. By comparing the q value of each factor and the q value of the two factors superposition, the geographical detector can determine whether there is an interaction between two factors, and whether their interaction is linear or nonlinear. Generally, q∈[0,1], q value is close to 1, meaning the independent variables has a strong influence on independent variables, while a q value is close to 0, indicating that the dependent variable is spatially randomly distributed. Detailed theoretical descriptions of the geographical detector model can be found in Wang and Hu (2012). The formula for the q value is depicted as follows: 1q=1-1Nσ2∑i=1nNiσi2 where, i=1,…,n represents the stratification for variable factor; N and Ni denote the total number of the units in the whole region and in stratum i, respectively; σh2 and σ2 are the variances of the dependent variable Y within Ni and N, respectively.

Referring to Liu et al. (2018), ten variables related to precipitation, the land surface environment conditions, and human activities were selected to determine the main factors influencing the spatial distribution of flood disasters. The precipitation variables can be characterized by the 3D structure properties of contiguous FCP event, including accumulated area, accumulated magnitude, lifespan, and moving distance. It is to be noted that the environment of the land surface (i.e., cropland, NDVI, slope of terrain, and elevation differences of terrain) and human activities (i.e., population and GDP) variables are represented by average values within the FCP event affected areas, as illustrated in Fig. 4c and d.

The scatter degree method (SDM) is an objective evaluation method that reflects the overall differences among evaluated objects. The basic principle of SDM is to select weight coefficient for indicators so that the difference among evaluated objects is maximized. Therefore, deriving weight coefficients for each indicator based on their intrinsic importance is critical. In our study, the SDM is employed to determine the weight of each indicator and calculate the comprehensive evaluation index for flood disaster and their driving factors, as the following steps:

Assume that the comprehensive evaluation function for y can be expressed as: 2y=ω1o1+ω2o2+…+ωmom=wTO where, w = (ω1,ω2,…,ωm)T denotes the weight coefficient vector, O = (o1,o2,…,om)T represents the standardized evaluated object vector.

According to the SDM method, let H = OTO, which is a real symmetric matrix. By calculating the eigenvalues of H, we can obtain the standard feature vector of maximum eigenvalue. Finally, the weight ωifor each sub-risk indicator is obtained from this eigenvector. Subsequently, put ωi into Eq. (2), the comprehensive evaluation index of flood disaster and their driving factors is computed.

Significant contrasts exist in flood characteristics induced by TC and non-TC rainstorm (Lu et al., 2020). In this study, 632 FCP events are classified into 246 TC and 386 non-TC precipitation events, and their 3D structural properties across China are shown in Fig. 5. The centroids of an FCP events represent the geographical location, as shown in Fig. 5a and b. The centroid size and color denote the accumulated area and magnitude, respectively. TC-induced rainstorm events mostly occur in the coastal areas of China, with the majority concentrated in coastal regions of Southern China (SC) rather than Northern China (NC) (Fig. 5a). These events have an average affected area of 9.63×102 km² and accumulated magnitude of 3.95×104 mm. Notably, TCs that moved farther inland exhibit larger accumulated areas and magnitudes, especially in northeastern NC.

Figure 5c shows the spatial distribution of the migration of TC-induced FCP events. The arrow color represents the lifespan of the FCP events, the length of the arrows represents migration distances of FCP events, and arrow is the direction of the migration of FCP event. On average, TC-induced FCP event persists for 22.84 h and travels 396.99 km. TC-induced FCP events are more inclined to move to the westward, which is consistent with an earlier study (Zhang et al., 2013). The westward shift of the western Pacific subtropical high (WPSH), westward steering flow, and easterly vertical wind shear are the main drivers of the formation of the TC moving westward (Wang et al., 2022). Meanwhile, some TCs prefer to shift northward or northeastward due to the eastward shift of WPSH, westerly vertical wind shear, weak mountainous blocking and westerly steering flow surrounding the TCs (Wang et al., 2022).

Figure 5b shows the centroids of non-TC-induced FCP events. It can be observed that non-TC-induced FCP events are mainly clustered in SC and NC. The average accumulated magnitude and projected area are 4.45×104 mm and 1.52×103 km², respectively. The non-TC events with larger accumulated magnitude and projected areas are mainly distributed in center of SC and NC. In contrast, the magnitude and affected area of non-TC events is larger than the TC events. The migration characteristics of non-TC events are illustrated in Fig. 5d. It is seen that the non-TC events are more inclined to move eastward, which is mainly attributed to the mid-latitude climatological westerlies and southwesterly monsoon circulation (Wang et al., 2022). These eastward moving FCPs have an average traveling distance of 489.58 km and lifespan of 21.63 h, which is slightly longer than TC-induced events. This finding is consistent with research by Wang et al. (2022), that suggested that FCP events with more persistent lifespan tend to travel longer and exhibit larger accumulated magnitudes and projected areas.

Figure 6 illustrates the long-term trends of multidimensional metrics of contiguous FCP events in China during 2000–2023. All metrics of contiguous FCP events show positive trends, consistent with findings from Li and Zhao (2022). The intensification of extreme precipitation is primarily governed by the thermodynamic effects constrained by the Clausius-Clapeyron (CC) relationship. When atmospheric temperature increases, saturation vapor pressure increases at a theoretical rate of 6–7 % °C⁻¹, providing the fundamental physical mechanism for enhanced extreme precipitation events (Trenberth et al., 2003; Allan et al., 2022). The accumulated area and magnitude of FCP events show slight increasing trends, with an annual variation of 0.15×102 km² yr⁻¹ and 0.07×104 mm yr⁻¹ (insignificant at the 0.05 level), respectively (Fig. 6a and b). The lifespan of FCP events shows an increasing trend at a rate of 0.12 h yr⁻¹ (Fig. 6c), and the total traveling distance exhibits an upward trend at a rate of 0.38 km yr⁻¹ (Fig. 6d). By contrast, a notable observation is that the non-TC events generally exhibit larger accumulated areas and magnitudes, as well as longer lifespans and moving distances, than TC events. Meanwhile, the uptrend trends of metrics of non-TC events are more pronounced than those for the TC events. This indicates that non-TC events have more pronounced and widespread impacts compared with TC events.

The spatial and temporal variations of hydrological processes during the FCP event periods are analysed (Figs. 7 and 8). It can be seen from Fig. 7 that FCP events associated with low surface runoff are found mainly in Northwestern China (NWC) and the Qinghai-Tibetan Plateau (TP). At the same time, FCP events with low surface runoff are distributed in the western parts of SC and NC. By contrast, FCP events with high surface runoff are predominantly observed in central SC and NC and northeastern NC. A comparison between TC and non-TC events reveals that the average surface runoff of non-TC FCP events is 30.81×103 mm, which is higher than that of TC FCP events with average value of 23.34×103 mm. Figure 7b presents a similar spatial variation of river discharge. FCP events with high river discharge are primarily concentrated in the central of SC, NC, and northeastern parts of NC. These findings indicate that regions with large surface runoff and river discharge in the central SC and NC face a higher probability of flood occurrence.

Figure 8a and b illustrate the temporal variations of surface runoff and river discharge during FCP event periods, respectively. Both surface runoff and river discharge show increasing trends, yet the trend is not significant at the 0.05 level. The mean surface runoff and river discharge for non-TC FCP events are higher than that of TC events. This finding can be attributed to the increase of precipitation features in China, as shown in Fig. 5. According to the correlation analysis between the accumulated magnitude of FCP events and surface runoff and river discharge, we find that surface runoff exhibits a significant correlation with accumulated magnitude (R=0.87, P<0.001). The correlation coefficients between the accumulated magnitude and surface runoff are 0.85 for non-TC and 0.93 for TC FCP events, both of which are statistically significant at the 0.01 level (Fig. 8c). By contrast, the correlation between accumulated magnitude and river discharge is relatively weaker, with the mean correlation coefficient greater than 0.67 and statistically significant at the 0.05 level. Consequently, enhanced 3D characteristics of FCP events across China drive increases in surface runoff and river discharge. Particularly, floods are more prominent in central SC and NC.

Figure 9 illustrates the spatial distribution of the flood affected population, death, direct economic loss, and affected cropland areas in China during 2020–2023. The geographical location of each flood event is represented by the centroid of the corresponding precipitation events. It can be easily observed from Fig. 9a that FCP events occurring in the SC and parts of the NC are mostly associated with larger affected population, whereas the flood-affected population in NWC and Qinghai-Tibetan Plateau (TP) is relatively low. Different spatial patterns can be found for the number of deaths caused by FCP events (Fig. 9b) across China when compared to that of flood-affected population (Fig. 9a). Higher number of deaths are observed in the SC and NC, but the number of deaths is fewer in the center of SC. Besides, the number of deaths is relatively large in the NWC and the surrounding regions of TP. The direct economic losses caused by FCP events are more severe in the southeastern and western parts of SC, as well as the central and northeastern regions of NC (Fig. 9c). For cropland, the affected cropland areas are mainly distributed in southeastern regions of SC and central and northeastern regions of NC (Fig. 9d). Besides, among these FCP events, TC-induced flood result in more severe disasters than non-TC-induced events.

To further investigate the flood-affected rate within the coverage areas of FCP events, we calculated the ratios of flood affected population, deaths, direct economic loss, and affected cropland areas to the total population, GDP, and total cropland within the coverage areas of FCP events, respectively (Fig. 10). As shown in Fig. 10a, high ratios of flood affected population to the total population within the coverage areas of FCP events are predominantly concentrated in SC, which is similar to the affected population distribution in Fig. 9a. However, high death rates are primarily distributed in the southeastern fringe of the TP and southwestern China (Fig. 10b). This phenomenon can be attributed to the high population density and robust economic development in eastern and southern China, and enhanced flood mitigation capacity to mitigate floods impacts by the construction of levees and flood-mitigated infrastructures (Fang et al., 2018). In contrast, mountainous areas in the southeastern fringe of the TP and southwestern China lack such infrastructures. FCP events in these regions often trigger sudden-onset flash floods accompanied by landslides and debris flows, severely threatening human lives (Chen et al., 2021). For the direct economic loss, the ratio to the total GDP is low in the southeastern SC, but relatively high in the western SC and northeastern NC. This suggests that FCP events caused significant economic losses in the western SC, rather than eastern SC with relatively high GDP. Regarding affected cropland, the smaller absolute affected cropland areas but larger ratio of affected cropland to the total cropland within the coverage areas of FCP event can be found in the coastal areas of SC. This is because this region frequently experiences TC rainfall processes, which often coincide with strong winds and storm surges, posing severe threats to the cropland. Overall, flood disasters caused by FCP in China exhibit the changing characteristics of “high impact-low losses ratio” in SC and NC and “low impact-high losses ratio” in TP and SWC.

Figure 11 shows the time series of the flood disasters in China during 2000–2023. As shown in Fig. 11a, the flood-affected population and the ratio of affected population to total population within the coverage areas of FCP events show a clear decreasing trend. The TC-induced affected population declines by 14.95×104 persons per year, which is larger than the non-TC event with 12.09×104 person per year (significant at the 95 % confidence level). The number of deaths also shows a decreasing trend during the period, except for extremely severe floods in 2016 and 2019. For the time being, we note that the TC-induced deaths decreased significantly by about 1.27 persons per year (P<0.05) from 2000 to 2023. A similar temporal pattern is observed for flood affected cropland areas in Fig. 11d, although there are differences in various of severe flood years after 2010. These results suggest that flood disaster represents a significant reduction trend in the past twenty years in China, despite there appear fluctuation features after 2010 due to extreme precipitation events, e.g., extreme rainfall in July 2016 affected 10.44 million people, caused 225 lives death, and resulted in economic losses of USD 57.46 billion. However, it is interesting that temporal trend of the direct economic losses from FCP event is completely different compared to other flood disasters metrics. Prior to 2010, flood caused direct economic loss is generally declined, then shifted to an increasing trend, and subsequently fall back into decrease. Notably, the trends of direct economic losses induced by TC and non-TC events are not statistically significant. In fact, previous studies have shown that economic development is correlated with the increase in direct economic losses (Wu et al., 2018). However, our analysis here suggests heterogeneous temporal changes in the decreasing trend in direct economic losses after 2010. This indicates that the direct economic loss varies with different levels of economic development.

Figure 12 displays the Pearson correlation coefficients (CC) between flood disasters and their driving factors. Flood disasters have a significant positive correlation with precipitation factors (P<0.001), except for the number of deaths, which shows a weak correlation. Among the precipitation factors, accumulated magnitude and lifespan show higher CC values with flood disasters, indicating that precipitation events with larger accumulated magnitude and longer lifespan tend to trigger more severe flood disaster impacts (Fig. 12a and e). Significant positive correlations between flood disasters and surface runoff and river discharge are found in Fig. 12b and f, revealing that high surface runoff and river discharge are critical factors to trigger flood disasters. Concurrently, human activity factors (i.e., GDP and population) also have positive correlation with flood disasters (P<0.001), suggesting that flood disaster effect caused by FCP is more serious in high economically developed and densely populated areas (Fig. 12d and h). Considering the relevance of the environment factor and flood disasters, cropland area shows positive correlation with the flood disaster, and NDVI exhibits a weak positive correlation with all flood disaster and not passing the significance test at 0.05. The result shows that flood-caused destruction of cropland leads to severe economic losses, while the changes in NDVI have a limited impact on flood disasters is limited. In contrast, slope and difference in elevation show weak negative correlatives with flood disaster, except for the number of deaths. This is because the most population and economically developed regions are predominantly distributed in the low-altitude areas of the eastern regions in China. However, mountainous areas experience relatively higher numbers of flood-related deaths leading to a positive correlation between flood disasters and altitude (Fig. 12c and g).

To further identify the key driving factors influencing the changes in flood disasters, based on the experience of previous research (Liu et al., 2018), we conducted an interaction detection based on Geodetector method in Fig. 13. When comparing the q value of each single factor, precipitation and hydrological factors have the greatest explanatory power for the flood disasters, followed by the human activity factors, and the environmental factors have the lowest explanatory power. This suggests that the flood disasters are overwhelmingly determined by the precipitation features, land surface hydrological processes, and human activities factors, which is consistent with findings from previous studies (Wang et al., 2022b). In all evaluation factors, the interaction q-values of any two factors are greater than those of the individual factors (Fig. 13), that is to say, the explanatory power of an individual factor may be enhanced when interacting with other factors. Specifically, the interaction between precipitation and hydrological factors and environmental factors displays significant nonlinear enhancement and bivariate enhancement effects. This means that although environmental factors have weak explanatory power on their own (Fig. 12c and g), their synergistic effects can exert a stronger influence due to nonlinear characteristics at different spatial and temporal scales. Therefore, the flood disasters in China are likely to be the result of a combination of factors.

To further reveal the spatial characteristics of interactions effect among driving factors on flood disaster, we calculated the comprehensive evaluation index of three driving factors based on the SDM method, and mapped the spatial distribution of dominant factors of flood disasters in Fig. 14. We found that precipitation and environmental factors have the largest spatial distribution among all potential driving factors, suggesting that flood disasters are influenced by combining precipitation and environmental factors in China. In the interim, the flood disasters events significantly influenced by the human activities are predominantly distributed in the center of SC and NC, which are densely populated regions in China (e.g., the middle-lower Yangtze River plain and the North China plain). This indicates that flood caused more severe disasters in densely populated areas, e.g., urban areas, which is consistent with our above findings presented in Fig. 12d and h.

Rainstorm is one of the most frequent natural hazards by triggering flooding and severe hazards in the world, seriously impacting populations, property, and socioeconomic development, thus, have garnered significant attention of the international society (Kharin et al., 2013; Ban et al., 2015; Wasko et al., 2021). In previous studies, the behaviours of rainstorms, such as frequency, intensity, and duration, have been extensively examined using maximum rainfall indices in specific regions (Fu et al., 2013; Deng et al., 2018). For example, Fu et al. (2013) constructed extreme rainfall indices and analysed spatial and temporal changes of extreme rainfall in China. However, these approaches only identified the continuous rainstorm spatial features at two-dimensional (latitude × longitude) and examined temporal changes based on the time series of extreme rainfall index, while ignoring the synchronous spatiotemporal connections and independence of rainfall events. This oversight could lead to an irrational analysis of rainfall features over study area. To overcome this limitation, recently, Silversmith (2021) proposed an integrated spatial contiguity algorithm to characterize precipitation events at the spatial scale. Wang et al. (2022a) investigated the 3D characteristics of spatiotemporal contiguous extreme precipitation events in China, meanwhile, Wang et al. (2025) analysed the 3D evolutionary features of spatiotemporally contiguous flood-causing precipitation events using CC3D algorithm. Thus, view from a higher-dimensional perspective, the identified 3D precipitation features not only retain the features recognized through a lower dimensional perspective but also provide a more visual and stereo demonstration of the spatial moving process of precipitation events (Feng et al., 2024).

In this study, we attempted to extend the CC3D algorithm to investigate the 3D characteristics of precipitation events and conducted the first comprehensive assessment of flood disasters from the event-based perspective. To this end, we developed an automatic precipitation event recognition algorithm with the CC3D method as its core, as shown in Fig. 3. This method in our study provides a possibility to accurately determine the spatiotemporal contiguous flood-causing precipitation events and reveal their “true” 3D structures and evolutionary patterns (e.g., moving direction, distance, and lifespan) from 2000 to 2023 in China. Additionally, the flood disasters caused by FCP events were extracted within the FCP event coverage areas, as illustrated in Fig. 4d, which could provide more reliable estimates of the influence of FCP events. Therefore, based on the above characteristics of events, our findings comprehensively reveal the characteristics of the precipitation and flood disasters from an event-based perspective. The results show that the FCP events were classified into 368 non-Tropical Cyclone (non-TC) and 246 TC events. TC-induced events, predominantly occurring in the coastal regions of Southern China (SC), have relatively small accumulated areas and magnitudes and short lifespan and moving distance, mostly inclined to move westward. In contrast, non-TC events are concentrated in SC and Northern China (NC), exhibiting eastward migration and characterized by larger accumulated magnitudes and affected areas and have a longer lifespan and moving distance than TC events. Thus, our research provided a new perspective for revealing the evolutionary features of spatiotemporally continuous flood-causing precipitation events from a 3D event-based perspective.

Precipitation, as a decisive factor of flood events, has attracted extensive attention from academia and policymakers. Many scholars have explored the spatiotemporal variability of extreme precipitation and its potential drivers in China. They suggested that the frequency of extreme precipitation events showed increasing trends and the relative contribution of extreme precipitation to the total precipitation also increased in most parts of China (Wu et al., 2019; Gu et al., 2022; Wang et al., 2022a; Fu et al., 2023). In comparison with these previous studies, our study introduces a new evaluation method, and comprehensively assesses the multi-dimensional characteristics of spatiotemporal FCP events. The results indicate that the accumulated area, accumulated magnitude, lifespan, and moving distance of FCP events are mostly increased during 2000–2023 (Fig. 6). This result is generally consistent with the trend of precipitation variation in China, and suggests that extreme precipitation events are projected to become more intense (Ban et al., 2015; Wu et al., 2019). These patterns are closely linked to the increase of the moisture-holding capacity of the atmosphere with higher temperature at approximately 7 % °C⁻¹ (Trenberth et al., 2003). Under global warming context, the increasing atmospheric moisture and buoyancy not only elevate the frequency of extreme precipitation events but also can lead to intensification rates that exceed the Clausius–Clapeyron (CC) scaling, exhibiting super-CC scaling (Fowler et al., 2021). Under a +4 K warming scenario, probable maximum precipitation (PMP) events are projected to become substantially more frequent, with their annual exceedance probability rising by approximately 101–102 (Hiraga et al., 2025). Besides thermodynamic controls, dynamic factors associated with large-scale circulation changes also modulate the frequency and propagation of extreme rainfall by shifting storm trajectories and movement speeds (Chan et al., 2023; Wasko et al., 2024).

In contrast to the increasing trends and intensifying strength of extreme rainfall, flood disaster shows very mixed trends in China due to differences in analytical areas, data, and methods. For instance, Xie et al. (2018) pointed out that the combination of more frequent and stronger extreme rainfall and the rapid expansion of urban agglomeration is more likely to increase the flood risk in Nanjing metropolitan. Similar results were also found along the coast of China (Feng et al., 2023), the extreme precipitation events led to a corresponding increase in the risk of flooding. However, in our study, the flood disasters are inconsistent with extreme rainfall and flood, and represent a significant reduction in the past twenty years in China, except for the direct economic losses (Fig. 11), which has been confirmed by Wei et al. (2018). From the perspective of spatial distribution, FCP events with larger accumulated magnitude and projected areas caused larger flood are mainly distributed in the center of SC and NC, and cause more serious disaster losses, but the flood loss proportions are largely reduced. This discrepancy might be caused by larger-scale hydraulic projects, such as dams and reservoirs. By the end of 2018, there were 98 822 reservoirs existed in China, with a total storage capacity of 8.95×1012 m³ (data are derived from China Water Statistical Yearbook 2018), which may significantly modulate the flood risks. Tang et al. (2023) studied the impact of dams on the flood in the Yangtze River, and found that dams and reservoirs mitigated the extreme flood by contributing to 94 % of the water level changes. In contrast, mountainous areas in the southeastern fringe of the TP and southwestern China, where the accumulated magnitude and projected areas of FCP events are relatively small, flood disaster proportions are large, especially the death (Chen et al., 2025). That means flash floods triggered by extreme rainfall events, representing a severe hazard in low-GDP regions. The robustness of the findings is supported by existing studies, suggesting that numerous engineering measures around the world have improved flood control standards and mitigated the rising risks of flash floods. Lim et al. (2018) analysed the global benefits of existing flood-control projects and identified an 8 % reduction in the exposed population and 7 % GDP property exposure losses per year in flood-inundated areas from 1986 to 2005. Zhao et al. (2022) indicated that reservoirs can decrease housing losses caused by flash floods by 9.7 %–45.7 %. In summary, flood disasters in China have been categorized into high-impact and low-loss in plain areas of SC, and low-impact and high-loss situation in the mountainous areas of west.

However, flooding is a complex physical process involving interactions among hydrology, meteorology, and land surface features that can be triggered by multiple mechanisms. Determining the factors that influence flood disasters from a multi-perspective is helpful for identifying the coupling mechanism between factors and flood disasters. We found that the anthropogenic activities (e.g., population and GDP) are the second most influential factor after extreme precipitation, which corresponds to previous literature reporting a positive relation between flood disasters and population, indicating that large populations in flood-affected areas are expected to have large flood risk exposure (Rogers et al., 2025). Our further examinations demonstrate the interaction between precipitation and environmental factors that displays significant nonlinear enhancement and bivariate enhancement effects, suggesting that precipitation situation and the topographical features of rainstorm area will have a major impact on the evolution of the flood-waterlogging process (Fig. 12). Notably, the flood processes in the center of SC and NC are more susceptible to anthropogenic activities (Fig. 14), posing significant threats to lives and property.

Although our study provided an estimation of the multi-dimensional characteristics of contiguous FCP events and first investigated their associated flood disasters and driving factors from an event-based perspective, there are several uncertainties and limitations that require further attention. It is essential to note that the 3D extreme rainfall features extracted with the different threshold values may have a different accumulated magnitude, projected area, lifespan, and moving distance. For instance, the accumulated areas of extreme rainfall extracted with the 95 % percentile of rainfall as a threshold may be smaller than that extracted with the 90 % percentile of rainfall. To accurately determine the appropriate extreme rainfall thresholds for multiple types of rainfall events needs to be further explored in further research. Meanwhile, the frequency of flood events cannot be accurately represented based on the current data resources in this study. Therefore, due to data limitations, only precipitation and flood features are considered in this study, and the temporal variations of frequency of flood were not analysed. Furthermore, flooding is a complex physical process that is typically influenced by a combination of natural and human factors. Though this paper only conducted a preliminary assessment of the relationships between flood disasters and driving factors, e.g., precipitation and flood features, environmental factors (cropland area, NDVI, slope, and difference elevation), and human activities factors (population and GDP), there may be other factors that probably exert important impact on flood disasters, such as hydraulic engineering projects (e.g., dams and reservoirs in mountain areas and flood-control measures in urban, as highlighted by Zhao et al. (2022) and Qi et al. (2022), which are very effective measures for mitigating the flood risk. Thus, it is necessary to integrate human activities and natural factors in further research to explore the mechanism of flood disasters. Overall, the 3D method can effectively capture the evolutionary patterns of contiguous FCP events. Strengthening research on extracting accurate precipitation characteristics and using more reasonable flood driving factors are expected to enhance our understanding of relationship between extreme precipitation and flood disasters in China.