Event‐based classification for global study of river flood generating processes

Better understanding of which processes generate floods in a catchment can improve flood frequency analysis and potentially climate change impacts assessment. However, current flood classification methods are either not transferable across locations or do not provide event‐based information. We therefore developed a location‐independent, event‐based flood classification methodology that is applicable in different climates and returns a classification of all flood events, including extreme ones. We use precipitation time series and very simply modelled soil moisture and snowmelt as inputs for a decision tree. A total of 113,635 events in 4155 catchments worldwide were classified into one of five hydro‐climatological flood generating processes: short rain, long rain, excess rainfall, snowmelt and a combination of rain and snow. The new classification was tested for its robustness and evaluated with available information; these two tests are often lacking in current flood classification approaches. According to the evaluation, the classification is mostly successful and indicates excess rainfall as the most common dominant process. However, the dominant process is not very informative in most catchments, as there is a high at‐site variability in flood generating processes. This is particularly relevant for the estimation of extreme floods which diverge from their usual flood generation pattern, especially in the United Kingdom, Northern France, Southeastern United States, and India.

and cannot simply be connected to changes in precipitation (Sharma, Wasko, & Lettenmaier, 2018). Soil moisture or catchment wetness state often play an important role in flood generation (Berghuijs, Harrigan, Molnar, Slater, & Kirchner, 2019;Ivancic & Shaw, 2015;Sharma et al., 2018;Slater & Villarini, 2016). For example, Blöschl et al. (2017) showed that changes in the seasonal timing of extreme precipitation are not always the clearest explanatory factor for changes in the timing of floods, as in large areas of Europe flood occurrences are more influenced by timing of snowmelt or soil moisture maxima. This shows that rainfall alone is not the only driver of floods and river peak flow events can occur for multiple reasons. In addition to hydroclimatological processes (rainfall, snowmelt, rainfall excess on saturated ground, rain-on-snow), floods can be generated through blockages (e.g. ice jam, dam break) or tidal surges as well (Whitfield, 2012).
Information about flood generating processes can be used in a number of different ways. It can be used to explain detected trends in flood magnitude or timing (Gudmundsson et al., 2019;Mallakpour & Villarini, 2015;Petrow & Merz, 2009; or to improve flood frequency analysis. In fact, the commonly used approach to flood frequency analysis assumes the flood sample to stem from a uniform distribution of flood events (England et al., 2018).
For an in-depth overview of studies addressing the classification of flood events, see Tarasova et al. (2019). Here we focus on various studies that have identified flood generating processes based on hydro-climatological information within a catchment. Merz and Blöschl (2003) developed a widely used (e.g. Nied et al., 2014 ;Nied, Schröter, Lüdtke, Nguyen, & Merz, 2017 ;Sikorska et al., 2015) framework that uses combinations of catchment state (water stored in soil and snow) and climatic inputs to produce diagnostic maps at the regional and national scale for Austria. These maps show all process indicators for simultaneous events and allow the analyst to individually choose between one of five types: long rain floods, short rain floods, flash floods, rain-on-snow floods, and snowmelt floods (Merz & Blöschl, 2003). This approach allows an accurate description of flood events; however, one of the drawbacks is that the final flood type classification is not done automatically but left to the user. This is a subjective and time-consuming process that limits the number of events that can be considered. The other disadvantage is that the person interpreting the map has to be familiar with what constitutes, for example, a large rainfall amount for certain regions, since extreme rainfall amounts vary across Austria (Merz & Blöschl, 2003) and even more across larger regions (Boers et al., 2019). Diezig and Weingartner (2007) advanced flood-type classification by introducing a decision tree to replace the user centred decisions. This concept was applied and extended also by Sikorska et al. (2015).
The decision rules are still based on expert knowledge and the decision thresholds are based on literature values. Hydrograph information (Diezig & Weingartner, 2007) and process indicators based on weather (Diezig & Weingartner, 2007;Sikorska et al., 2015) and storage state (Sikorska et al., 2015) are inputs for the decision tree.
Excluding discharge information can make this analysis available to catchments with limited data availability. Although the use of a decision tree makes this classification applicable to large samples and computationally efficient, the drawback of using regional thresholds based on literature values remains. It limits the applicability of their decision tree to regional studies, in this case in Switzerland, as set thresholds (e.g. event rainfall needs to be greater than 12 mm; Sikorska et al., 2015) are only valid in the intended region.
Larger scale studies that assess flood generating processes across different climates use different approaches. Berghuijs, Woods, Hutton, and Sivapalan (2016) determine dominant flood generating processes for the MOPEX catchments in the continental United States by comparing the mean occurrence date of the annual maximum peak flow to the occurrence date of the annual maximum of the hypothesized causal processes, which are: daily or multi-day precipitation events, precipitation excess or a combination of snowmelt and rain. Berghuijs et al. (2019) extend the seasonality statistics to Europe and additionally infer the relative distribution of different flood generating processes within one catchment by using circular statistics. Blöschl et al. (2017) analysed how flood timing in Europe changes based on observed changes in flood generating processes thus allowing a conclusion about which processes are most influential in certain regions. Blöschl et al. (2019) similarly found trends in flood magnitude to be closely related to changes both in precipitation and soil moisture for several areas in Europe. All these methods are based on the average timing of flood generating process versus average timing of flood event occurrence. This type of analysis determines dominant flood generating process; however, it cannot classify flood generating processes for individual events. While averaged results are still valuable, for example, for climate change impacts assessment (Blöschl et al., 2017), it still assumes the same flood processes within a catchment, which rarely is the case (Hirschboeck, 1987;Sikorska et al., 2015). The benefit of individual event process information is lost in such approaches. Any information pertaining particularly to extreme events, which might have a different generating process than the average annual maximum (Rogger et al., 2012;Smith, Cox, Baeck, Yang, & Bates, 2018), would not be available. Additionally, there are no studies that extend the analysis beyond the continental to the global scale, thus showing a lack in standardized flood process classification both for the dominant and the event scale analyses.
In this article, we present a widely transferable methodology to identify both the dominant flood generating process and the singleevent generating process for catchments with different climates. The only streamflow information necessary for the analysis is the date of the annual maximum flow, thus reducing reliance on uncertain peak flow measurements. Other input data were kept to a minimum as well, so that a large number of stations could be included in the analysis, and the method is potentially transferable to data-scarce regions. This is especially valuable, as any prior studies of flood generating processes have been focused on Europe or North America (Berghuijs et al., 2016;Berghuijs et al., 2019;Blöschl et al., 2017). Our study extends the knowledge of flood generating processes to other continents and climates as well. The new classification is tested for its robustness and evaluated with available information, two steps that are often lacking in current flood classification approaches (Tarasova et al., 2019).

| METHODOLOGY FOR A GLOBAL FLOOD CLASSIFICATION
This section describes how we infer flood generating processes at a global scale. Section 2.1 describes the flood event data source and how for each of the study catchments' daily rainfall, temperature, and evaporation data is calculated. These variables are then used as input to simple conceptual models producing daily snowmelt and soil moisture estimates (Section 2.2). Section 2.3 explains how flood process indicators are used in a conceptual decision tree to identify the flood generating process for a flood event. Section 2.5 then presents the methods applied to evaluate the robustness of the new classification.

| Data
The Global Streamflow Indices and Metadata Archive (GSIM)  provides streamflow station metadata, catchment delineation, catchment characteristics, and selected hydrological indices for more than 30,000 stations. While daily time series data are not made available, the date and magnitude of annual maximum flow are published. A value is supplied if at least 350 days of reliable daily flow data are available for that station . For more information about the quality control both for the catchment delineation procedure and time series inhomogeneities refer to Gudmundsson et al. (2018b). Only stations which have a high quality in regard to both aspects, determined through the provided quality flags, are kept for the analysis. No Australian metadata were available to Do et al. (2018a) to quality check Australian catchments, we therefore used the catchment outlines delineated by Fowler, Peel, Western, Zhang, and Peterson (2016) to quality check the catchment delineation ourselves according to the criteria given by Do et al. (2018a). Out of 221 catchments in Fowler et al. (2016), 187 had a match in the GSIM database. In 62 catchments, the area between the catchment delineation by Do et al. (2018a) differed less than 10% from the catchments by Fowler et al. (2016) and were thus included in the analysis.
For the climate variables, global gridded climate data sets were chosen to make the analysis transferable across locations. Although global gridded data sets have location-specific uncertainties as well, the aim was to reduce uncertainty due to varying interpolation methods and to make the method potentially transferable into areas where precipitation gauge data are not available. A challenge in the selection of the data products was to find high-resolution data sets of global extent that cover several decades. For flood analysis, it is recommended to have at least 20 years of data available (Kjeldsen, 2015). Daily data are needed since a flood generating rainfall might just last for a few hours or days and would not be recognized in a coarser temporal resolution. This limits the choice of data sets available. For some data sets (e.g. temperature), a coarser resolution had to be accepted to be able to cover a longer time period.
The precipitation product used in the analysis is the Multi-Source Weighted-Ensemble Precipitation (MSWEP) Version 2.1 . It is a daily gridded precipitation product available at 0.1 x 0.1 resolution from 1979 to 2015 that merges satellite, reanalysis, precipitation gauge, and streamflow gauge data. Even without precipitation gauge correction, MSWEP is among the most accurate gridded precipitation products as evaluation with rainfall station data and hydrological modelling shows (Beck, Vergopolan, et al., 2017). MSWEP 2.1 additionally includes gauge correction. This means for some areas (snow-affected and/or 'complex topography'; , streamflow station data are used to avoid underestimation of long-term precipitation estimates. Martens et al. (2017) use the MSWEP data as forcing data for the Global Land Evaporation Amsterdam Model (GLEAM) . The current version GLEAM v3.2a is used as source for global actual evaporation information in this study. The model includes satellite, reanalysis, and merged products for various variables and provides daily evaporation data between 1980 and 2015 at 0.25 x 0.25 resolution. Although GLEAM has a tendency in some places to overestimate evapotranspiration (Khan, Liaqat, Baik, & Choi, 2018;Miralles et al., 2016), this fault is common among other evapotranspiration products as well. GLEAM is still more successful than comparable data products at closing the water balance (Miralles et al., 2016).
For air temperature, the Berkeley Earth Surface Temperature daily gridded product was used (Rohde et al., 2013). It is available at 1 x 1 resolution from 1970 to 2013. It utilizes the Global Historical Climatology Network stations to interpolate a global grid of surface temperature. The large station network and interpolation algorithm by Rohde et al. (2013) produces a temperature product with lower uncertainties than comparable data (Menne et al., 2018). Although the daily gridded temperature product used here is not yet peer-reviewed, multiple sources assess it as comparable in accuracy to temperature products of similar extent and temporal resolution and it is used in multiple analyses (Levi Goss, 2013, e.g. Osborn, Jones, & Joshi, 2017Wasko & Sharma, 2017). The different time extents of the above three data sources limit our analysis to the period 1980-2013.

| Snow and soil moisture accounting routine
For each climate variable, a catchment average daily value was produced using the GSIM delineated catchments. Each grid cell that is covered by a catchment is assigned a weight based on the percentage of the catchment area that covers this cell. With these weights a weighted average value was calculated for each catchment. The catchment daily values were used as input into a simple coupled soilsnow routine to calculate snowmelt and soil saturation.
These simple models were adapted from the soil and snow routines used by Berghuijs et al. (2016). Snowmelt output is based on a degreeday snow model. Snow is accumulated if the temperature is below a critical temperature and melts if temperature is above. The daily time steps used did require some updates to the soil and snow routines applied by Berghuijs et al. (2016). One update is that rainfall input can only occur when the temperature is higher than the critical temperature thus separating rainfall and snowfall. The other is that the soil routine is only active if there is no snowpack, which is justified by the fact that soil moisture under a snowpack is not directly affected by precipitation and very little by evaporation. The bucket model gives an approximation of the spatially averaged moisture state of the catchment.
If T(t) > T crit , where at time t S snow is snow storage (mm), P precipitation input (mm), P rain is liquid precipitation, T is air temperature ( C), T crit is the temperature threshold where rainfall turns to snow ( C), f dd is a melt factor (mm day -1 K -1 ), and P melt is snowmelt rate (mm). f dd is set to 2 mm day -1 K -1 (Berghuijs et al., 2019(Berghuijs et al., , 2016. For critical temperature, the data product by Jennings, Winchell, Livneh, and Molotch (2018) were used, which provide global gridded critical temperature for the Northern hemisphere. With no gridded data available for the Southern Hemisphere, critical temperature is set to 1 C, the mean critical temperature of the Northern hemisphere (Jennings et al., 2018). Unlike Berghuijs et al. (2016) the snow routine calculates only snowmelt and not rain-on-snow to separate these two processes.
Although degree-day models are simplifications of snow processes, they work well at low resolution such as catchment averages (Hock, 2003) and are often used for flood classification (Tarasova et al., 2019). Snow storage was set to zero at the end of the annual average warmest month to include only annual snow and not accumulated snow over several years (Freudiger, Kohn, Stahl, & Weiler, 2014).
The information about snow storage (S snow ) is used as input for the soil routine, which assumes that soil filling, evaporation, and overflow only happens when there is no snow cover and when T > T crit .
Adapted from Berghuijs et al. (2016), soil storage and soil saturation are calculated as follows: If S snow (t) > 0, where S u is the soil storage (mm), S u,max is the maximum soil storage (mm), S sat is the soil saturation (%), P eff is the excess rainfall (mm), P rain is the liquid precipitation input (mm), and Ea is the actual evapotranspiration (mm). Maximum soil storage is set to the available water storage capacity (AWC) taken from the Harmonized World Soil Database averaged for each catchment (Nachtergaele, van Velthuizen, & Verelst, 2009). More detail including an evaluation of the combined soil-snow routine is given in Data S1, Supplementary A.

| Decision tree
A decision tree (Diezig & Weingartner, 2007;Sikorska et al., 2015) is used to decide which process generated each annual maximum flow event. While blockages (e.g. ice-jams, glacial outburst floods) can also cause extremely large floods, this article will only focus on hydroclimatological flood generating processes as they cause the overwhelming majority of floods (Whitfield, 2012). The structure of the tree is based on our domain knowledge. This means that the shape of the tree and the decision nodes are based on our understanding of flood generating processes, instead of being inferred from data through an automatic algorithm (Witten, Frank, & Hall, 2016). Specifically, several process indicators are calculated from the daily time series of precipitation, snowmelt, and soil moisture, which are then used by the tree to decide between five different flood generating processes: short rain flood, long rain flood, excess rainfall flood, rain/ snowmelt flood, or snowmelt flood (Berghuijs et al., 2016;Merz & Blöschl, 2003;Sikorska et al., 2015). If none of these flood processes can be assigned, the flood event is described as 'other'. Flash floods and glacier melt floods are currently not included in the analysis.
The tree is used for classifying individual flood events whose dates of peak flow are known. The tree makes the classification decision based on information about potential causal factors (rain, snowmelt, soil moisture) that occur during a short time frame before the recorded flood peak. Since there is no daily discharge data available in GSIM for event separation as used by Sikorska et al. (2015), the time frame needs to be fixed. For alpine catchments, Froidevaux, Schwanbeck, Weingartner, Chevalier, and Martius (2015) found the time period relevant for flood generation is 3 days prior to a flood event. Berghuijs et al. (2016) found for the MOPEX catchments in the United States that it varies between 3 and 10 days. We decided to set the threshold to 7 days in order to cover multi-day rain or snowmelt events even in large catchments. Without daily discharge time series, a more accurate delimitation is not possible. We evaluated the 7-day threshold by checking the relationship between catchment area and mean event response time ( Figure S4). Based on the results, the time period of 7 days was considered applicable for both small and large catchments. With the flood relevant time period set to 7 days, the process indicators are each calculated for the 7-day period.
Each process indicator is used as a node in the decision tree. The decision tree with the process indicators is presented in Figure 1. Process indicators and associated thresholds are given in Table 1. A pseudocode description of the decision tree is given in Data S2, Supplementary B. The thresholds of the tree are inferred from the input time series itself by using either percentile thresholds (heavy rainfall, heavy snowmelt) or ratios (rain/snow). This makes the tree transferable across different locations. The only exception is the threshold for soil saturation, which is set to >90% (Sikorska et al., 2015), thus representing a near-saturated soil. Since this analysis focuses on annual maxima and not extreme floods, the 90th percentile is assumed to be a good indicator for finding large enough events to cause annual maximum flow.
Without using the Julian date of the flood occurrence, as was done by Sikorska et al. (2015), the other process indicators needed to be structured more strongly on hydrological reasoning. Hence, for our classification, any involvement of snow needs to be evaluated before all other processes, as it would be missed if soil moisture conditions or rainfall amount was evaluated first. Similarly, antecedent soil moisture conditions are evaluated before heavy rainfall events. In fact, Ivancic and Shaw (2015) demonstrated for the United States that extreme rainfall events are much more likely to cause an extreme streamflow event under wet antecedent conditions. Testing for heavy rainfall conditions F I G U R E 1 Conceptual decision tree for location independent (global) flood classification. Thresholds are given in Table 1 T A B L E 1 Note: p 90 refers to the 90th percentile of the respective process indicator distribution. The snow/rain indicator is taken from Vormoor, Lawrence, Heistermann, and Bronstert (2015), and the soil moisture state threshold from Sikorska et al. (2015). P melt (t 7 ) is the snowmelt sum over 7 days before the flood event. p 90 for melt is calculated over all values P melt (t 7 ) > 1. P(t 7 ) is the rainfall sum over 7 days before the flood event. P total = P(t 7 ) + P melt (t 7 ). P 7 is the mean 7-day rainfall. before excess rainfall conditions would therefore miss this important flood generating process. Nevertheless, the tree recognizes that wet conditions still require rainfall for flood generation (Berghuijs et al., 2019) by using an additional decision node, which tests if an aboveaverage amount of rainfall fell in addition to wet antecedent conditions.
If none of these conditions are met, the category 'other' will be selected, which describes a flood event that is either misclassified or caused by something other than hydro-climatological processes (e.g. dam break, ice jam, groundwater flood, storm surge, etc.).
A visual demonstration of the inputs of the decision tree is given in Figure 2 for the 2008 and 2011 maxima in the Aschauer Ache catchment in Austria. In April 2008 (left), a melting event and a day of strong rainfall occurred right before the annual maximum flow, thus allowing the conclusion that the combination of snowmelt and rainfall generated the flood. The peak flow event in October 2011 (right) was instead preceded by several days of strong rainfall, falling on not yet wet soils. This would therefore be classified as a long rainfall flood.

| Dominant process and measure of process variability
A dominant flood generating process here is defined as the flood process occurring most often in the time series; however, in some catchments, this information is not as meaningful if flood generation varies evenly between two or more processes. The dominant process for each catchment can be identified if at least 20 years of events are classified in order to have a representative sample (Berghuijs et al., 2016;Kjeldsen, 2015).
The inter-annual variability of the flood generating process is calculated using a variability measure for categorical data taken from Allaj (2018): where v k is the variability value, k is the number of categories (flood generating processes), and f is the relative frequencies for each category. When one process dominates, the value of v k is near 0; when several processes are equally dominant, the value of v k increases. If there are k categories then the upper bound on v k is 1 −1= ffiffi ffi k p (Allaj, 2018). The calculated variability is normalized to make the information easily accessible.
2.5 | Evaluation of the proposed global flood classification 2.5.1 | Evaluation by sensitivity analysis A sensitivity analysis is used to determine the effect that changes in the inputs of the classification system may have on the output. It provides an insight into which input factors lead to a high variability of the output, thus evaluating robustness of the classification (Pianosi et al., 2016;Tarasova et al., 2019). In particular, the influence of the chosen model routine parameters and input data uncertainty (melt rate, AWC, critical temperature), as well as of the tree parameters and thresholds, is determined using a regional sensitivity analysis method (Spear & Hornberger, 1980;Young, Spear, & Hornberger, 1978). The tree structure is not evaluated with a sensitivity analysis as it is based on our hydrological understanding of flood processes. Changing the structure of the tree would produce outcomes not agreeing with our flood process definitions.
The flood classification is run with 1,000 parameter samples that are generated using a Latin hypercube sampling scheme over a uniform distribution. The range of parameters tested in the sensitivity Following the rationale of the regional sensitivity analysis approach (Pianosi et al., 2016;Spear & Hornberger, 1980), the sensitivity to each parameter is determined by comparing the cumulative distributions of that parameter for each flood process to the uniform distribution of the entire parameter sample. The deviations from the uniform distribution are measured using the Kolmogorov-Smirnov statistic (D statistic). The lower these deviations, the less influence the parameter has on the classification outcome. Regional sensitivity analysis was applied using the SAFE toolbox (Pianosi, Sarrazin, & Wagener, 2015).

| Evaluation by comparison with available data
The accuracy of the event classification based on observed data is dif-

| Evaluation of the proposed global flood classification
3.3.1 | Evaluation by sensitivity analysis Figure 7 shows the results of the sensitivity analysis. A high value of the Kolmogorov-Smirnov D statistic for a given parameter (horizontal F I G U R E 5 Time series of annual flood generating process for seven example stations around the world illustrating the variability beyond dominant (i.e. most frequent) process. In brackets is the normalized inter-annual variability. The Ping river variability is missing as less than 20 years of data were available for calculation axis) and flood process (colours) indicates strong sensitivity (Pianosi et al., 2016), that is, a change in that parameter has a strong effect on the number of times that process is classified. For the parameters of the soil-snow routine (melt rate, critical temperature, and maximum soil storage), the statistic is less than 0.1 for all processes. This indicates that a change in those values will not have a strong impact on the overall distribution of classified processes. The parameters for the simple model routine therefore have a low influence.
However, the classification of some flood processes is very sensi-  (Figures 7 and S11). Figure S11 additionally informs us that a decrease of a threshold leads to an increase in the respective process. For the heavy rainfall threshold, this increase in classified long rainfall floods leads to a decrease in events classified as 'other'. in the aridity distribution in that area (e.g. Knoben et al., 2018). Arid and semi-arid regions rarely experience excess rainfall floods, and short rainfall and long rainfall are the prevalent generating processes there ( Figure S5b). Examples for this are the more arid regions in South Africa, Namibia, and Australia. The only exception is seasonally arid catchments. We find excess rainfall to be a common flood process in seasonally arid catchments ( Figure S5b).

| Evaluation by comparison with available data
In humid areas, especially in the humid tropics, we find excess rainfall to be the most common process. For the Ping catchment in  (Berghuijs et al., 2016) and Europe (Berghuijs et al., 2019). However, Berghuijs et al. (2016) find no excess rainfall dominance in the Northeastern United States, whereas our classification F I G U R E 7 D statistic for the Kolmogorov-Smirnov test for the regional sensitivity analysis. The higher the D statistic value, the more sensitive is the respective process to changes in the parameter/ threshold. The theoretical range of D is between 0 and 1 .
The parameters of the model routine are melt rate, critical temperature, and soil storage. A more detailed figure with the cumulative distribution functions can be found in Figure

| Event-based flood generating processes
One key findings of this study is the widespread year-to-year variability in flood generating processes. Although some areas like Central Europe, India, and Central Brazil show low variability, the overwhelming majority of catchments are classified as regularly experiencing annual maxima generated by two or more different processes. This is very important for flood frequency analysis. Although it has long been known that frequency analysis, particularly of extreme floods, has higher accuracy if the flood distribution is split by flood type (Elliott et al., 1982;Hirschboeck, 1987;Potter, 1958;Waylen & Woo, 1982), such distinction is still not standard procedure. For example, the Bulletin 17-C for the United States does recommend separation of the flood frequency curve into different processes; however, it does not supply guidance on how to do this (England et al., 2018;, although there are recent approaches to rectify this (Barth, Villarini, & White, 2019). Especially in areas where the extreme flood process might deviate from the regular annual maxima, any flood estimation procedure might likely underestimate that extreme (Rogger et al., 2012;Smith et al., 2018). Areas where this may happen are, for example, the United Kingdom, the northwest of France, the southeast of the United States and Central India, as Figure 6  tions. This could lead to the model routine predicting a melt peak earlier or later than the actual peak, which means the snowmelt event is missed by the classification. One solution for this problem could be to define the model routine on a gridded basis instead of as lumped for each catchment. This would require downscaling gridded temperature to smaller grid cells which take topographic differences such as elevation into account.

| Classification tree method
A regional classification tree method (Diezig & Weingartner, 2007;Sikorska et al., 2015) was adapted and extended here to be transferable to several climates by using thresholds based on simulated time series, instead of literature values. These climate independent thresholds make the tree applicable at global scale. As opposed to other flood classification methods (Merz & Blöschl, 2003;Nied et al., 2014;Sikorska et al., 2015), the classification tree does not require local knowledge of seasonality or weather patterns of the analysed catchments. The order of the decision nodes is based on hydrological process knowledge. An extension of the tree is possible, if further information is of interest. This could include glacier melt floods or a more detailed split of flood producing conditions.
The global extent of the application limits the temporal and spatial distribution of the data used as input to the classification. However, the methodology can easily be adapted to higher resolution data. This might be a higher temporal resolution of the climate input data or, if daily flow data is available, the classification of multiple flood events per year.
An important part of the decision tree is the last node 'other'.  Sikorska et al. (2015) and Brunner et al. (2017) found that a fuzzy tree reached the same 'dominant' process per event as a crisp tree. An advantage of using a crisp tree is that it enabled the comparison with (crisp) flood cause data from the Dartmouth Flood Observatory.
This study extends river flood classification to a larger scale than has been done before. The focus of previous studies has been on continental (Europe), national (United States, Switzerland, Austria), or regional scale. One motivation to extend the analysis of flood generating processes to the global scale was to make the classification of flood generating processes comparable across more climates (Gupta et al., 2014;Linsley, 1982

| Outlook
The widespread relevance of soil moisture in flood generation shown by this study can be applicable in climate change impact studies. Wasko and Sharma (2017) found changes in soil moisture with warming temperatures to be more relevant for streamflow response than extreme rainfall variation. The results presented here would support these findings, as extreme rainfall was not identified as the dominant process in most catchments. Nevertheless, it has to be taken into account that most catchments experience mixed processes and in some catchments where soil moisture is usually influential, a very extreme rainfall can still lead to extreme flooding. Therefore, if a catchment experiences occasional long rainfall/short rainfall floods, these might increase in frequency or magnitude, whereas excess rainfall floods might be less affected by changes in extreme rainfall. These different processes should be taken into account when flood risk changes are predicted  and further studies regarding flood trends should focus on how floods of different processes might change differently. Additionally, the impact of climate change on flood generating processes will need to be examined in further studies. A first step in that direction could be a better understanding which catchment and climate characteristics are relevant in shaping the flood process mix.

| CONCLUSIONS
A new global methodology to analyse flood generating processes has been proposed and applied to 4,155 catchments of the GSIM database. Flood process indicators were queried in a decision tree to identify these processes for each annual flood peak flow event. The structure of the classification tree is dependent on the flood process definition and sensitive to changes in the threshold parameters. Nevertheless, the evaluation showed that most extreme flood events were classified consistent with reports from the Dartmouth Flood Observatory, with snowmelt influenced floods occasionally misclassified.
The analysis revealed that excess rainfall, that is, rainfall on wet soils, is a common flood generating process across several climates and continents. It also demonstrated the need for an event-based analysis, with a high variability of flood generating processes being the norm rather than the exception in most catchments. This should raise awareness of possible uncertainties in the common practice of using one distribution during flood frequency analysis to estimate extreme floods. This is especially relevant since the most extreme and damaging floods might be generated by a process different from the dominant flood process (Rogger et al., 2012;Smith et al., 2018). The results found by the global flood classification are furthermore important for any future work analysing impact of system changes on flood events. Given the primary role of soil moisture in flood generation, the impact of predicted increases in extreme precipitation must be considered in the context of soil moisture, including future changes in soil moisture.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available in 'global flood classification' at https://github.com/lshydro/global flood classifi cation, Version 1. These data were derived from the following resources available in the public domain: • Global Streamflow Indices and Metadata Archive (GSIM) https:// doi.pangaea.de/10 .1594/PANGAEA.887477.