On the flood peak distributions over China

Here we for the first time present a nationwide characterization of flood hazards across China. Our analysis is based on an exceptional dataset of 1120 stream gauging stations with continuous records of annual flood peaks for at least 50 years across the entire country. Our results are organized by centering on various aspects of flood peak distributions, including temporal changes in flood series and their spatial variations, the statistical distribution of extreme values, and the properties of storms that lead to annual flood peaks. These aspects altogether contribute to an improved understanding of flood hydrology under a changing environment over China and promote advances in flood science at the global scale. Historical changes in annual flood peaks demonstrate frequent abrupt changes rather than slowly varying trends. The dominance of decreasing annual flood peak magnitudes indicates a weakening tendency of flood hazards over China in recent decades. We model the upper tails of flood peaks based on the generalized extreme value (GEV) distributions. The GEV shape parameter is weakly dependent on drainage area, but it shows spatial splits tied to rainfall climatology between northern and southern China. Landfalling tropical cyclones play an important role in characterizing the upper-tail properties of flood peak distributions especially in northern China and southeastern coast, while the upper tails of flood peaks are dominated by extreme monsoon rainfall in southern China. Severe flood hazards associated with landfalling tropical cyclones are characterized by complex interactions of storm circulations with synoptic environments (i.e., mid-latitude baroclinic disturbances) and regional topography.

even though it remains unsettled whether the changes are due to natural climate variability or human-induced climate change 25 (e.g., Held and Soden, 2006;Marvel and Bonfils, 2013;Trenberth et al., 2015;Schaller et al., 2016;Risser and Wehner, 2017;Eden et al., 2017). The stationarity assumption of flood series has been questioned and debated in scientific community (Milly et al., 2008;Montanari and Koutsoyiannis, 2014;Salas et al., 2018). Extensive studies on the stationarity of annual flood peaks have been carried out in many parts of the world (e.g., Robson et al., 1998;Robson, 2002;Franks and Kuczera, 2002;Villarini et al., 2009;Petrow and Merz, 2009;Villarini et al., 2011;Ishak et al., 2013;Tan and Gan, 2014;Mediero 30 et al., 2014;Hodgkins et al., 2019), including some efforts in global-scale investigations of historical changes in flood series (e.g., Arnell and Gosling, 2016;Do et al., 2017Do et al., , 2019. Due to the limitation of observational datasets, existing knowledge on flood hazard is significantly biased towards Europe and North America, with the characteristics of other worldwide regions (including China) far from being well represented. There are some regional studies across China (e.g., Zhang et al., 2016Zhang et al., , 2014Zhang et al., , 2018bLiu et al., 2018). A nation-wide investigation on the stationarity in flood series over China, however, is still missing. 35 The exceptional dataset of annual flood peaks, as demonstrated in present study, will provide additional evidence for detectable changes in flood hydrology under a changing environment. Better understanding of historical changes in annual flood peaks is of paramount importance for constraining model-based projections of flood hazards (e.g., Milly et al., 2002;Hirabayashi et al., 2013;Dankers et al., 2014;Arnell and Gosling, 2016). In this study, we expect to explore the dominant mode (i.e., abrupt changes or slowly varying trends) of nonstationarities in flood series, and highlight potential factors that induce the changes in 40 annual flood peaks.
Improved understanding of flood hazard requires essential knowledge of flood-generation mechanisms. This is also a critical aspect to consider for improved flood frequency analysis (Hirschboeck, 1988;Singh et al., 2005;Leonard et al., 2014;Brooks and Day, 2015;Yan et al., 2017Yan et al., , 2019. Smith et al. (2018) shows that the most extreme flood peaks are frequently determined by extreme events resulted from anomalous flood agents for particular regions of the United States (which is the notion of 45 "strange floods"). Mixture of flood-generation mechanisms poses great challenges for characterizing the upper tails of flood peaks, as different flood agents might lead to flood regimes with distinct statistics (e.g., magnitude, timing, frequency). This is, however, often the case for many regions in the world (e.g., Jarrett and Costa, 1988;Smith et al., 2011;Villarini, 2016;Blöschl et al., 2017;Smith et al., 2018;England et al., 2018). We expect annual flood peaks over China characterized with a mixture of flood-generation mechanisms, due to its geographic location in a monsoon-climate region and on the margin of the most active 50 ocean in tropical cyclones. China suffers the most frequent landfalling tropical cyclones in the world, with 9 tropical cyclones making landfall on average per year (Jiang and Jiang, 2014). Despite its significance, little is known about the hydroclimatology of flooding associated with landfalling tropical cyclones. Even less effort has been spent on investigating the impacts of different flood-generation mechanisms on the upper-tail properties of flood peaks across China. This is a critical issue for China that shows contrasting rainfall climatology (under combined influences from monsoon and landfalling tropical cyclones) between 55 the northern and southern part of the country (i.e., traditionally take the Yangtze River as the geographic divide) (e.g., Yang et al., 2013;Gu et al., 2017a;Zhang et al., 2018a). Extreme floods for different regions are often associated with contrasting flood agents. This is not merely associated with the nature of flood agents themselves, but is also determined by complex interplay of storms with ambient synoptic and physiographic environment. For instance, extreme rainfall from landfalling tropical cyclones can be amplified through interactions of storm circulation with mid-latitude baroclinic disturbances (e.g., Hart and Evans, 2000) and regional topography (e.g., Houze, 2012). Propagation of monsoon also plays a role in determining the spatial contrasts of flood agents through regulating temporal occurrences of flood peaks over different regions (e.g., Ding and Zhang, 2009). Knowledge in the mixed flood-generation mechanisms and their spatial variations can provide valuable insights into improved procedures for the estimates of Probable Maximum Precipitation (PMP) / Probable Maximum Flood (PMF) in designing flood-control infrastructures (e.g., Smith and Baeck, 2015;Yang et al., 2017).
An important way of characterizing flood hazards is through examining flood peak distributions and factors that determine the upper-tail properties. In this study, we model annual flood peaks based on the statistical framework of the generalized extreme value (GEV) distributions (similarly see e.g., Katz et al., 2002;Morrison and Smith, 2002;Villarini and Smith, 2010;Barros et al., 2014;Bates et al., 2015;Gaume, 2018;Smith et al., 2018). The key focus is placed on the upper tails of flood peaks across China. Previous studies show strong dependence of location and scale parameters for the GEV distributions on 70 drainage area, while the GEV shape parameters only weakly depend on drainage area (Morrison and Smith, 2002;Villarini and Smith, 2010). Weak dependence of the GEV shape parameters on drainage area indicate scale-independent properties of the upper tails of flood peaks, and highlight additional factors (e.g., spatio-temporal rainfall variability) in determining the upper tails of flood peaks. Yang et al. (2013) identified a spatial contrast of extreme rainfall distributions between northern and southern China and pointed to contrasting flood hydroclimatology across the country. We therefore propose that similar spatial 75 contrasts also exist in flood peak distributions across China.
Our study is also motivated by Typhoon Nina and the resultant August 1975 flood in central China. The August 1975 flood in central China, with 26000 direct fatalities, is one of the most destructive floods in the world history (Yang et al., 2017). The unit peak discharge is 17 m 3 s −1 km −2 (i.e., flood peak discharge divided by drainage area) for a 760 km 2 drainage basin, and is on the list of the world maximum floods. The August 1975 flood plays a key role in shaping the envelop curve of floods in 80 China and different versions of the world envelop curve (Yang et al., 2017;Costa, 1987). Devastating consequences of Typhoon Nina and the August 1975 flood partially resulted from cascading collapses of dozens of dams, and expose inadequacies of conventional approaches for flood frequency analysis (e.g., fitting historical flood records with assumed distribution functions) (e.g., Smith and Baeck, 2015;Yang et al., 2017). This is an urgent issue for China, as statistics show socio-economic damages caused by tropical cyclones are rapidly increasing in recent decades, with a large portion of the damages resulted from riverine 85 flooding (Zhang et al., 2009;Rappaport, 2014).
Based on the aforementioned gap of our knowledge in flood hydrology, we examine flood peak distributions across China by centering on the following questions: (1) What is the dominant mode of the violation of stationarity in annual flood peak series? (2) How do dominant flood-generation mechanisms vary across China? (3) How do upper-tail properties of flood peak distributions depend on drainage areas (i.e., scale-dependence) and rainfall climatology? (4) What is the impact of landfalling 90 tropical cyclones on the upper tails of flood peaks across China? (5) What are the characteristics of the most severe flood hazards (i.e., as represented by the number of stations with annual flood peaks) in the history of China and the tropical cyclones that induce them? Even though these questions are examined based on an exclusive dataset over China, timely answers to these questions will undoubtedly contribute to the compliment of our limited understandings on flood hazard under a changing environment, and promote the advance of flood science at the global scale.

2 Data
Our analysis is based on observations of annual maximum instantaneous peak discharge from 1120 stream gauging stations with continuous records of at least 50 years (i.e., no missing data consecutively throughout the entire periods). There are relatively more stations distributed in eastern China than the western part of the country (Figure 1). The dataset is comprehensively collected from local hydrographic offices of nine major river basins across China. All these stations are nation-level control 100 stations with the records that have been through strict quality control procedures to ensure data consistency and accuracy (by following the code for hydrologic data compilation of China, SL247-1999). Stations with notable site re-locations (i.e., that lead to changes in drainage area) during the observational periods are not included in this dataset. The flood records demonstrate a variety of ways in data collection, mainly include intermittent direct measurements of discharge during flood seasons, indirect inferences through stage-discharge rating curves, and post-flood field surveys. In addition to flood peak magnitude, flood peak 105 timing (i.e., date of occurrence for flood peaks) is also provided, and is mainly used to infer flood-generating mechanisms over different regions across China. Flood records with missing flood peak timing are discarded from the following analysis.
Time series of total number of available stations are shown in Figure 2a. The longest flood record is 153 years, with approximately more than 90% stations fully available during the period from 1960 to 2017. The record length of 66% stations exceeds 60 years starting from 1950s till the year of 2017 ( Figure 2b). There are considerable variabilities in the spatial scales of repre-110 sented river basins, with a large percentage (approximately 64%) of stations representing small and medium river basins (with drainage areas less than 5000 km 2 , Figure 2c). Previous studies found contrasting climate regimes and extreme rainfall distributions between northern and southern China (e.g., Yang et al., 2013;Ma et al., 2015). To facilitate analyses and comparisons, we further classify the 1120 stations into two sub-groups, i.e., northern and southern China, based on their geographic locations ( Figure 1). The northern group includes stations mainly in northeastern river basins, the Yellow River basin, the Huaihe River regression line (loess function, Cleveland, 1979) to detect change point in variance in annual flood peak series (similarly see, e.g., Villarini et al., 2009;Villarini and Smith, 2010;Yang et al., 2013). We also adopted a different change-point detection approach, i.e., the one proposed by Matteson and James (2014), but only found negligible deviations from the results based on Pettitt's test (results not shown).
Monotonic trends can be induced by existence of abrupt change points in mean rather than indicating slowly varying trend for the flood series. For those series that do not show significant abrupt change points in mean, we directly use the non-parametric 130 Mann-Kendall test (Mann, 1945;Kendall, 1975) to examine the presence of monotonically increasing or decreasing trends in annual flood peak series. For the series with change point in mean, we divide it into two sub-groups and test monotonic trends for each of the two sub-groups (i.e., before and after the change point). Additional trend analysis for the sub-series can highlight stations that show both abrupt changes and slowly varying trend in the entire flood series. We assume the existence of only a single change point in mean for each flood peak series in this study, to avoid dividing the series into too many segments 135 (similarly see, e.g., Villarini et al., 2009Villarini et al., , 2012. Only sub-series with record lengths exceeding 10 years are considered in the trend analysis. We set a significance level of 5% (i.e., two-tailed) for both the change-point and trend tests.

Generalized Extreme Value distribution
The Generalized Extreme Value (GEV) distribution is used to statistically model distributions of annual maximum flood peaks (e.g., Coles, 2001;Villarini and Smith, 2010). The GEV, based on extreme value theory, has been widely used in flood frequency 140 analysis (e.g., Coles, 2001;Katz et al., 2002;Morrison and Smith, 2002;Villarini and Smith, 2010). The cumulative distribution function of the GEV takes the form: where µ, σ, and ξ represents the location, scale, and shape parameter, respectively. The location (µ) and scale (σ) parameter is related to the magnitude and variability of the records, respectively. The shape parameter (ξ) indicates the tail properties of the distribution, with positive (negative) values pointing to heavy and unbounded (light and bounded) upper tail of flood peak 145 distribution. The GEV parameters are estimated based on the maximum likelihood estimators (e.g., Coles, 2001). We fit the GEV distributions only for stations without statistically significant change points in mean and variance and monotonic trends, following the basic assumption of probability theory that data samples should be independent and identically distributed. The three fitted GEV parameters (i.e., location, scale and shape) will be further used to examine their correlations with drainage areas, shedding light on the scale-dependence of the upper-tail properties of flood peak distributions across China.   , 2000;Villarini and Smith, 2010;Smith et al., 2011;Villarini et al., 2014). We obtain the information of tropical cyclones from the International Best Track Archive for Climate Stewardship (IBTrACS, see https://www.ncdc.noaa.gov/ibtracs/ for details). The dataset provides records of the circulation center location (latitude and longitude) and storm intensity (represented by minimum sea level pressure) at a temporal interval of 6 hours. An additional attribute provided by IBTrACS for each tropical cyclone at each time interval is the nature of the storm, i.e., extratropical transition or tropical storm. Extratropical 160 transition (ET) characterizes the changing properties of a tropical cyclone from a warm-core, symmetric structure to a coldcore, asymmetrical structure (e.g. Hart and Evans, 2000). Physical process associated with extratropical transition plays an important role in determining the spatial distribution of tropical cyclone rainfall (e.g. Atallah and Bosart, 2003;Atallah et al., 2007;Liu and Smith, 2016). Tropical storm (TS), as a contrast, indicates the maintenance of a warm-core, symmetric structure during the entire life cycle of the storm.

4 Results and discussion
The structure of this section is organized as follows. We first detect change points and monotonic trends to shed light on the long-term changes in flood series across China, and discuss possible drivers that induce them (subsection 4.1). We move on to subsection 4.2 to examine seasonal distribution of annual flood peaks, highlighting the mixture of flood-generation mechanisms across China and its spatial variation. Results from both subsection 4.1 and 4.2 will serve the basis for the analysis of subsection Spatial and temporal clustering of change points demonstrate evidence of anthropogenic influences on flood hydrology (e.g., Vogel et al., 2011;Hodgkins et al., 2019). Through meta-data inspection of selected stations, we are able to relate some of the abrupt changes in annual flood peaks to intentional human activities. For instance, the change point in mean at the year of 1986 in the upper Yellow River, the Guide hydrological station, is due to the construction of a large hydropower-generation dam, the Changes in land use/land cover (e.g., urbanization, deforestation/afforestation) can also contribute to change points in the series of annual flood peaks. This is especially the case for stations in the lower Haihe River basin (where the Beijing-Tianjin-

205
Hebei metropolitan region is distributed) and Yangtze River delta region (where Shanghai and other major cities are located). Figure 4c shows a small urban watershed in the lower Yangtze River basin) that experienced rapid urbanization in recent decades. Transboundary water-transfer project demonstrates another form of anthropogenic influence on flood hydrology.
Abrupt increases in flood peak magnitudes are mainly tied to the elevated base flows transferred from neighboring river basins.
We provide the annual flood peak series for a station in the lower Yellow River basin ( Figure 4d). Increasing water demand 210 from domestic and agricultural sectors in the lower Yellow River basin lead to extensive implementation of water-transfer projects.
Abrupt changes in the series of annual flood peaks can also originate from the changes in extreme rainfall across China.
However, one of our previous studies investigated changes in annual maximum daily rainfall over China, but found no clear signature of spatial clustering for change points in either mean or variance for the rainfall series, although abrupt changes in 215 annual maximum daily rainfall frequently occurred in the 1990s (see Figure 2 in Yang et al., 2013). Inconsistent spatial patterns of change points in annual maximum flood peak and annual maximum daily rainfall series indicate a weak role of climate shifts in producing abrupt changes in annual flood peaks.

Monotonic trends
We further examine the monotonic trends of annual flood peak series based on the Mann-Kendall test for those stations that do 220 not show significant change points in mean. There are only 69 stations (accounting for approximately 6% of the total stations) with significant linear trends ( Figure 5a). For those stations with significant linear trends, 62 (7)  Changes in annual rainfall extremes (i.e., annual maximum daily rainfall) show a "dipole-like" spatial structure over China, with decreasing trends in northern China and increasing trends in the south (e.g., Yang et al., 2013;Ma et al., 2015;Gu et al., 2017b). The decreasing annual maximum flood peaks in northern China may be partially attributed to the weakening rainfall 235 intensity in recent decades. The opposite trends in annual rainfall extremes and annual maximum flood peaks in southern China seem contradictory to our perception. Contrasting trends between intense rainfall and annual high flows are also found over United States (mainly eastern of the Mississippi River), which are attributed to inconsistent changes of intense rainfall in different seasons (Small et al., 2006), i.e., changes in fall precipitation mainly contributes to the trend in annual rainfall extremes, while annual high flows are often observed in spring with no significant changes in rainfall. This is, however, not the and southern China as well as mixture of flood-generation mechanisms across the entire country.

Extreme Value Distribution
We model distributions of annual flood peaks using the GEV distribution. We only focus on the stations without significant change points in mean or in variance, and without significant monotonic trends (i.e., the stationary stations). There are 486 stations that satisfy these requirements. These stations are densely located in southern rather than northern China (Figure 7), 280 mostly due to the spatial clustering of stations with abrupt change points in annual flood peaks in northern China (Figure 3).
The stationary stations represent a wide range of spatial scales of drainage basins for both northern and southern China. Figure   8 shows the dependence of GEV parameters on drainage area for the 486 stationary stations. Location and scale parameters are positively correlated with drainage area in a log-log domain. The correlations are all significant at the level of 5%. The shape parameter, however, generally decreases with drainage area but shows only weak dependence in a log-log domain (with a 285 correlation coefficient of -0.15 for northern China and -0.16 for the south, neither being statistically significant). The upper-tail properties (as represented by the shape parameter) of flood peak distributions are weakly determined by drainage areas, while the magnitude and variability of annual flood peaks can be well explained by drainage area. Our results are consistent with the study in the eastern United States by Villarini and Smith (2010), and contribute to generalized understanding on the upper-tail properties of flood peak distributions.

290
An interesting finding is that there are striking spatial splits in terms of the dependence of the GEV parameters on drainage areas between northern and southern China (Figure 8). The location and scale parameters for stations in southern China are consistently larger than their counterparts in the north (with a few exceptions, Figure 8a and 8b). The shape parameters in northern China are comparatively larger than that in southern China. Large shape parameters indicate heavier upper tails of flood peak distributions in northern than southern China, even though the magnitudes and variability of flood peaks are 295 relatively smaller in the north. One of our previous studies on the distribution of annual maximum daily rainfall found similar spatial splits for the dependence of GEV parameters on elevation between northern and southern China (Yang et al., 2013;Gu et al., 2017a). Spatial splits in extreme rainfall distributions highlight spatial heterogeneity in flood hydroclimatology across China (which is also represented by the contrasting seasonal distributions of annual flood peaks shown in section 4.2). Spatial contrasts of extreme rainfall distribution further lead to different relationships between three GEV parameters and drainage 300 areas for flood peak distributions between northern and southern China.
We further show the spatial splits for the shape parameter in Figure 7. The majority of the northern stations show positive shape parameters, while the southern stations are mixed with both negative and positive shape parameters. Spatial contrast in rainfall climatology between northern and southern China seems to be a more effective predictor in explaining the spatial variability of shape parameter rather than drainage area. Our results highlight the importance of hydrometeorological analyses

Tropical cyclones and upper tails of flood peaks
We examine the impacts of tropical cyclones on the upper-tail properties of flood peak distributions across China in this subsec- Tropical cyclones contribute to approximately 18% of annual flood peaks over China. Figure 9 shows the map of the percentage of annual flood peaks that are caused by tropical cyclones to total annual flood peaks for each station. More than 50% of the annual flood peaks are caused by tropical cyclones in the southeastern coast of China, with the percentage even attaining 90% over the Hainan Island. The percentage gradually decreases when we move further inland and to higher latitudes. Less than 10% annual flood peaks can be associated with landfalling tropical cyclones in the middle portion of the Yellow River and Yangtze River basins (Figure 9). The percentage of annual flood peaks caused by tropical cyclones is closely tied to the spatial distribution of tropical cyclone rainfall and frequency of tropical cyclone occurrence over China (Wu et al., 2005;Ren et al., 2010;Gu et al., 2017b). More than 30% of the extreme rainfall events are induced by tropical cyclones along the coastal 325 regions (Gu et al., 2017a, b), with the percentage gradually decreased moving inland due to rapid weakening of storm intensity (e.g., surface roughness, insufficient moisture transport).
We show the stations with record floods (i.e., the largest flood peak for the entire record of a station) that are caused by tropical cyclones in Figure 9 to highlight the impacts of tropical cyclones on the most extreme floods. Stations with record floods caused by tropical cyclones are spatially clustered in the southeastern coast, central and northeastern China (Figure 9).

330
Tropical cyclone-induced record floods in the southeastern coast are mainly associated with abundant moisture and energy supply for extreme rainfall right after tropical cyclones making landfall. However, the spatial clustering of record floods by tropical cyclones in northern China (more specifically, the upper Huaihe River and northeastern China) can be partially related to extratropical transition processes during the life cycle of the storm and/or interactions with regional topography, as will be elaborated below. We do not observe a comparable distribution of record floods caused by tropical cyclones in southern China

335
(e.g., the Yangtze River basin) excluding the coastal regions, even though the percentage of annual flood peaks caused by tropical cyclone is comparable to that in northern China (less than 30%, Figure 9). Our results highlight the impacts of tropical cyclones on flood peak distributions in northern China with a large percentage of record floods caused by relatively infrequent visits of landfalling tropical cyclones.
The impact of tropical cyclones on the upper tail properties of flood peak distributions is further examined through the 340 shape parameter of the GEV distribution. We compare the shape parameters between the entire annual flood peak series and the series with annual flood peaks caused by tropical cyclones removed ( Figure 10). We focus on the series with record length exceeding 30 years after annual flood peaks caused tropical cyclones being removed from the series. This leads to the exclusion of most stations in the southeastern coast due to the high percentage of tropical cyclone-induced flood peaks ( Figure 9). As can be seen from Figure   . Extreme rainfall associated with East Asia Summer Monsoon, rather than landfalling tropical cyclones, can be a more important player in characterizing the upper tail of flood peak distributions in most inland regions of southern China (e.g., the middle and lower portion of the Yangtze River basin) (Zhang et al., 2017). Tropical cyclones in northern China, even though characterized with low frequency of occurrence, pose significant influences on the upper-tail properties of flood peak distributions.
We focus on tropical cyclones that produced relatively large numbers of flood peaks over China, to shed light on the phys-360 ical attributes of most severe flood hazards associated with landfalling tropical cyclones. There are 9 tropical cyclones that produced more than 100 annual flood peaks over China since late 1950s till present. The 9 tropical cyclones alone contribute to approximately 50% of total annual flood peaks caused by tropical cyclones. Table 1 provides a summary of the 9 tropical cyclones. Typhoon Herb (1996) produced the largest number of annual flood peaks (167 in total), followed by Typhoon Wendy (1963) andTyphoon Tim (1994). Typhoon Herb (1996) produced a large number of annual flood peaks right after its landfall 365 in mainland China (Figure 11a). Almost all the annual flood peaks caused by other tropical cyclones are distributed over the most inland regions ( Figure 11). The percentage of stations with annual flood peaks caused by tropical cyclones relative to total storm-affected stations (i.e., located within 500 km buffer zone of each tropical cyclone track) varies between 14% (Typhoon Doris) and 35% (Typhoon Herb). Typhoon Andy (1982) and Typhoon Russ (1994) lead to annual flood peaks for more than 30% storm-affected stations (Table 1).

370
The 9 tropical cyclones can be further categorized into two groups according to the nature of the storm and spatial patterns of their tracks. The first group includes Typhoon Herb (1996), Typhoon Andy ( There is no strong preference for the spatial distribution of annual flood peaks with respect to storm tracks (i.e., left or right of the track), even though the records floods caused by tropical cyclones tend to be frequently observed in the left-front quadrant (typically the down-shear side) of the circulations. This is related to the preferable distribution of extreme tropical cyclone rainfall, due to enhanced moisture convergence and updraft on the down-shear side of the circulation (e.g., Atallah et al., 2007;Shu et al., 2018).

Summary and Conclusions
In this study, we examine flood peak distributions over China based on 1120 stream gauging stations with continuous records of annual maximum instantaneous discharge for more than 50 years. The principal findings of this study can be summarized 395 as follows.
(1) There are 38% and 35% stations exhibiting significant change points in mean and in variance, respectively. Change (2) Approximately 6% stations (69 in total) show significant linear trends in the annual flood peak series. Those stations with significant trends are uniformly distributed across the country, with 62 of them exhibiting significantly decreasing trends.

405
The decreasing trends of flood peak magnitude in northern China may be at least partially tied to changes in extreme rainfall.
Disconnections between changes in annual rainfall extremes and annual maximum floods are identified in southern China, and highlight complex flood-generation processes across China. The dominance of decreasing trends in annual flood peak series indicates weakening tendencies of severe flood hazards (i.e., annual maximum floods) over China, even though flood-affected area and economic damages are on the rise in recent decades (Kundzewicz et al., 2019). Future studies need to further examine 410 changes in flood frequency for a complete assessment on flood hazards (based on peaks-over-threshold flood series, similarly see, e.g., Mallakpour and Villarini, 2015).
(3) We fit GEV distribution for the stationary time series of annual flood peaks, and examined the dependence of its parameters on drainage area. We find that the location and scale parameters are linearly scaled with drainage area in a log-log domain. There is only a weak tendency for the shape parameters to decrease as a function of drainage area. Our results highlight by England et al. (2018) in Hydrology Subcommittee Bulletin 17C as an imminent need to "define flood potentials for watersheds altered by urbanization, wildfires, deforestation, and by reservoirs". Innovative approaches that explicitly address the nonstationarities should be embraced for flood frequency analysis across China, for instance, process-based approaches that rely on physically-based hydrological modelling which can represent the processes of nonstationarities in flood series (see e.g., Wright et al., 2014;Yu et al., 2019), statistical modelling approaches that mathematically parametrize the role of human 450 regulations in flood series based on the framework of probability theory (Salas et al., 2018;Serago and Vogel, 2018;Gao et al., 2019;Dong et al., 2019;Barth et al., 2019). These approaches should be especially in great needs for northern China that exhibits an overwhelming portion of stations with nonstationarities in flood series.
Our results highlight the important role of landfalling tropical cyclones in determining the upper tails of flood peak distributions across China, especially the northern China and the southeastern coast. Previous studies show strong teleconnections 455 between tropical cyclone activity in the western North Pacific basin and large-scale climate variability, e.g., the El Niño-Southern Oscillation (e.g., Chan and Shi, 1996;Chan, 2000), Madden-Julian Oscillation (e.g., Kim et al., 2008). Statistical models that adopt varying parameters on time or other predictors (such as, large-scale climate indices) can provide predictive tools of understanding future changes in flood hazards associated with landfalling tropical cyclones (e.g., Zhang et al., 2018c).
Future studies need to zoom into watershed scales, and explore physical connections between extreme flood processes and 460 key tropical cyclone features (e.g., space-time structures of tropical cyclone rainfall, tropical cyclone intensity), to provide additional insights into flood hazard associated with landfalling tropical cyclones.
A unique feature of our study is a nation-wide assessment of flood hazard based on an unprecedented network of stream gauging stations across China. Comprehensive analysis based on the exceptional dataset over China, together with studies by Villarini et al. (2009) and Burn and Whitfield (2018) in North America, Blöschl et al. (2017Blöschl et al. ( , 2019 1980 1980-1990 1990-2000 2000-2010 before 1980 1980-1990 1990-2000 2000-2010 (1) (2) (4) (3) Figure 3. Change points in (a) mean and (b) variance. Color represents the year of change-point occurrence. The insert plot shows the histogram of the years of change-point occurrence (y-axis represents the number of change points, while x-axis represents the ending year of a 10-year period, e.g., 1990 actually means [1980][1981][1982][1983][1984][1985][1986][1987][1988][1989][1990]. Only stations with results being statistically significant (at the level of 5%) are shown. There are 436 (38%) and 398 (35%) stations with significant change points in mean and in variance, respectively.  (c) are associated with (1) insufficient record lengths (e.g., less than 10 years) for sub-groups before or after change points, (2) linear trends for either sub-group being not statistically significant.     Table 1 for more details.