1. Introduction
Anthropogenic climate change has led to increased climate variability, which, when translated into the hydrological system, results in increases in extreme hydrological events in terms of both frequency and magnitude in many regions (Chiang et al. 2021; Hirabayashi et al. 2021; IPCC 2021; Yin et al. 2018; Zhang et al. 2022). The increases in the frequency and magnitude of extreme hydrological events directly indicate enhanced floods and droughts that adversely affect multiple ecosystem services (e.g., water supply, crop yield, carbon sequestration, etc.) (Blackburn and Stanley 2021; Hendrawan et al. 2022; Li et al. 2022; Venkatappa et al. 2021) and socioeconomic developments (e.g., Tabari et al. 2021), and also bring increasing challenges to water resources management and risk prevention (e.g., Milly et al. 2008; Kreibich et al. 2022). Understanding how, when, and where the extreme hydrological events changed (will change) is important prior to meeting this challenge. However, hydrological observations are rather limited and nonevenly distributed (e.g., Fekete and Vörösmarty 2007; Sivapalan 2003; Alkama et al. 2011), and the available hydrological records are often incomplete or of limited length (e.g., Cutore et al. 2007; Gleason et al. 2014), which prevents a comprehensive understanding of the spatial and temporal patterns of hydrological extremes globally over the past 40 years.
Recent advancements in hydrological modeling have promoted the development of many global runoff datasets that have been widely used in assessing changes in hydrological extremes under historical and future climate changes on the global scale (e.g., Giuntoli et al. 2015; Liu et al. 2021; Zaherpour et al. 2018). For example, runoff outputs from macroscale climate/hydrological/land surface models have been used for assessing historical changes in high and low flows (Gudmundsson et al. 2021), and for predicting future changes in floods and droughts under various climate change scenarios (Hirabayashi et al. 2013; Tabari et al. 2021). However, despite being widely used, these modeling outputs usually contain large uncertainties from meteorological forcing (e.g., Müller Schmied et al. 2016) and/or model parameterizations (e.g., Zhao et al. 2017). For example, Zaherpour et al. (2018) showed that the modeled high and low flows from global hydrological models often exhibit a large intermodel difference in the magnitude and occurrence date. Similar findings are also obtained for macroscale land surface models (e.g., Zaitchik et al. 2010). These modeling uncertainties directly cast doubts on the associated flood and drought analyses and consequent scientific findings (Beevers et al. 2020; Chaney et al. 2015; Meresa et al. 2021).
Comprehensive model evaluations are fundamentally crucial for understanding the reproducibility and spatiotemporal characteristics of runoff datasets and the associated uncertainties, which are important prior when applying these datasets for hydrological analyses (Beck et al. 2017b). Many studies have assessed the modeled runoff against observations (e.g., Beck et al. 2015, 2017b; Yang et al. 2015). However, compared with mean flows (annual flow and/or monthly flow), evaluations of extreme flows are relatively limited (see Table 1) due to the need for streamflow observations at a much higher frequency (usually daily records). Additionally, existing evaluation studies of extreme flows often suffer from one or more limitations, including (i) only using a limited number of models (e.g., Hirabayashi et al. 2008; Müller Schmied et al. 2014), which does not allow intermodel comparisons and quantifications of uncertainty ranges; (ii) not taking into account the diversity of climate–geographical zones (e.g., Beck et al. 2015; Liu et al. 2021; Zhao et al. 2017), thus only providing an overall assessment of model performance across all examined catchments while lacking detailed information regarding model performance across a fuller range of environmental conditions; (iii) not excluding human interventions (e.g., irrigation, reservoir, and urbanization) on observed streamflow (e.g., Hirabayashi et al. 2008; Liu et al. 2021; Müller Schmied et al. 2014; Yang et al. 2021; Zaherpour et al. 2018; Zhao et al. 2017), noting that most models have limited capacity in directly simulating human interventions so it is necessary to carefully screen out catchments affected by human activities to ensure the observed and modeled flows are comparable; (iv) not including a river routing scheme that is essential for short-term (e.g., daily) streamflow simulations (e.g., Beck et al. 2015, 2017b; Hirabayashi et al. 2008; Müller Schmied et al. 2014; Zaherpour et al. 2018); and/or (v) assessing the individual uncertainties caused by forcing data and model structures and parameterizations, which are critical for improving streamflow simulation (e.g., Müller Schmied et al. 2016; Mockler et al. 2016). As a result, although previous evaluation studies have gained valuable understandings of the performance and uncertainty in these modeled global runoff datasets, their limitations necessitate more comprehensive evaluations of simulated extreme flows from more models, under different environmental conditions and against higher-quality streamflow observations.
Overview of global-scale studies evaluating hydrological extreme against observations. LSM indicates the land surface model and GHM indicates the global hydrological model. Studies are listed chronologically, then alphabetically, with the current study being provided for completeness. In the four right-most columns “N” denotes no and “Y” denotes yes for the model attributes summarized by the column title.
To address this need, here we present a systematic investigation aimed at assessing simulated extreme flows from 23 global models against daily streamflow observations from 633 unimpaired catchments, including (i) 14 climate models within the Coupled Model Intercomparison Project phase 6 (CMIP6; Eyring et al. 2016); (ii) 6 global hydrological models (GHMs) from the Inter-Sectoral Impact Model Intercomparison Project (ISIMIP2a; Warszawski et al. 2014; Frieler et al. 2017); and (iii) 3 land surface models (LSMs) originating from the Global Land Data Assimilation System (GLDAS) version 2.0 (Rodell et al. 2004). The 23 selected models feature daily runoff outputs and boast an original spatial resolution more precise than 2° × 2°. Both the magnitude (the maximum and minimum) and timing (date of occurrence) of high and low flows are considered. In the case of GHMs and LSMs, these models are driven by observed or reanalysis meteorological forcing data. Conversely, in CMIP6 models, meteorological fields and surface processes are modeled in a coupled manner. It is noteworthy that CMIP6 models frequently exhibit substantial biases in the simulated meteorological forcings (Fyfe et al. 2021). Thus, our primary objective is not to underscore distinctions between CMIP6 runoff estimates and those generated by ISIMIP2a and/or GLDAS models. Rather, we seek to draw attention to the direct applicability of CMIP6-derived streamflow information in diverse hydrological applications. Furthermore, given that CMIP6 models typically employ surface hydrology schemes akin to those employed by ISIMIP2a and GLDAS (as detailed in Table 2), a comparative analysis of CMIP6 models versus ISIMIP2a and GLDAS permits a quantitative examination of the biases in streamflow resulting from meteorological variable discrepancies within CMIP6 models. Moreover, the distinctive design of ISIMIP2a, which incorporates three distinct reanalysis meteorological forcings, facilitates a comprehensive exploration of the uncertainties inherent in reanalysis forcings as they pertain to simulated extreme flows. The utilization of identical forcing data in both ISIMIP2a and GLDAS models affords the opportunity to quantify the uncertainties arising from model structures and parameterizations (see section 2). Finally, to render the simulated runoff data from grid cells amenable to direct comparison with measured streamflow at shorter time scales (e.g., daily), we have employed a process-based river routing scheme to aggregate gridded runoff at the catchment outlet. The findings of this study hold significant importance in informing the choice of modeled runoff data for hydrological extreme analysis. Moreover, it underscores the uncertainties that underlie relevant scientific findings and elucidates areas where future model enhancements are urgently warranted.
Summary of models evaluated in this study.
2. Materials and methods
a. Streamflow observation and subregion classification
To evaluate the simulated extreme flows, observed streamflow (Q) was sourced from a comprehensive global Q collection (Beck et al. 2017b; Yang et al. 2021), encompassing daily or monthly Q observations from more than 22 000 catchments globally. Several rigorous selection criteria were applied to identify catchments that are most suitable for our objectives, including (i) catchments with complete daily Q records spanning a minimum of 15 years between 1 January 1971 and 31 December 2010, (ii) catchments having an area exceeding 10 000 km2 to better match the coarse-resolution runoff simulations, and (iii) catchments that are minimally affected by direct human interventions [i.e., irrigation area < 1% (Siebert et al. 2015); urban area < 1% (https://www.edenextdata.com/?q=data); total reservoir capacity < 10% of mean annual Q (Lehner et al. 2011)]. As a result, a total of 633 catchments met the defined criteria and were employed in the subsequent analyses (Fig. 1).
Locations of the 633 catchments. The 633 catchments were classified into four climatic aridity zones (i.e., arid, semiarid, subhumid, and humid) and cold (mean annual snowfall fraction > 20%) vs noncold regions.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0176.1
In addition, to compare the model performance across diverse climatic regions, the 633 catchments were further grouped into different subregions according to two key climatic indicators: (i) the dryness index [calculated as the ratio of mean annual potential evaporation to mean annual precipitation, following McVicar et al. (2012)] and (ii) the mean annual snowfall fraction (defined as the proportion of mean annual snowfall to mean annual total precipitation) (Fig. 1). Monthly datasets for potential evaporation (EP), precipitation (P), and snowfall at a spatial resolution of 0.1° were procured from the ERA5-Land dataset for the period spanning 1971–2010 (Muñoz-Sabater et al. 2021). These monthly gridded data were further temporally aggregated and spatially lumped for the 633 catchments at a mean annual scale.
b. Simulated runoff datasets
The simulated daily runoff from 23 macroscale models evaluated herein were respectively obtained from three modeling experiments, including (i) 14 climate models from CMIP6 (Eyring et al. 2016), (ii) 6 GHMs from ISIMIP2a (Warszawski et al. 2014), and (iii) 3 LSMs from GLDAS 2.0 (Rodell et al. 2004). CMIP6 is the most recent phase of CMIP, which consists of a number of global climate models that simultaneously model the land and atmospheric processes in a coupled form. ISIMIP2a is the second phase of ISIMIP, with “a” indicating the historical run (between 1971 and 2010 for most models). In ISIMIP2a, surface hydrology is simulated by state-of-the-art GHMs forced by reanalysis meteorological fields. Three atmospheric reanalysis forcing datasets are used in ISIMIP2a, including (i) the Princeton Global Meteorological Forcing dataset version 2 (PGFv2; Sheffield et al. 2006), (ii) the bias-corrected Twentieth Century Reanalysis data generated during the third phase of Global Soil Wetness Project (GSWP; Kim 2017), and (iii) the Water and Global Change Data/WATCH forcing data (WATCH/WFDEI; Warszawski et al. 2014; Weedon et al. 2014). GLDAS was developed to optimally simulate land surface states and fluxes using advanced LSMs and data-assimilation techniques (Rodell et al. 2004). GLDAS 2.0 is forced with PGFv2 and covers 1948–2014. In the following analyses, the common overlapping period (i.e., 1971–2010) between the three modeling experiments is our evaluation period. To evaluate the ISIMIP2a-simulated extreme flows, we employ runoff outputs derived from ISIMIP2a, driven by the PGFv2 dataset. This choice is made to maintain a coherent and uniform climatic forcing between the ISIMIP2a simulations and the GLDAS 2.0 dataset. Moreover, the ISIMIP2a-simulated extreme flows forced with all three reanalyzed forcings are also used to quantify the simulated uncertainty caused by different forcings.
c. Streamflow extreme characteristics
Four magnitude and four timing characteristics of extreme flows are evaluated herein; they are summarized in Table 3.
Summary of extreme streamflow characteristics evaluated herein. A calendar year is used when determining the annual extreme flow characteristics.
d. River routing model
Gridded runoff outputs from the 23 models were coupled to the Catchment‐Based Macro‐Scale Floodplain (CaMa-Flood) model (Yamazaki et al. 2013) to obtain daily streamflow estimates. CaMa-Flood is an open-source global river routing model and has gained wide utilization in river routing simulations in conjunction with gridded runoff data (Dottori et al. 2018; Lim et al. 2018; Liu et al. 2021; Zhao et al. 2017). The first-order conservative method was adopted to resample the simulated runoff from the 23 models to a 0.25° spatial resolution (Jones 1999), aligning with the resolution of CaMa-Flood. We also tested different resampling strategies and found that the output of CaMa-Flood is generally insensitive to the spatial resolution of the input runoff grids (Figs. S1–S4 in the online supplemental material). We then repeated the routing simulation using the first year (i.e., 1971) daily runoff forcing three times for model spinup. Finally, an adaptive time step scheme was employed to optimize the time step for the estimation of river discharge (Yamazaki et al. 2017, 2019).
e. Evaluation metrics
We employed four widely accepted statistical metrics to assess the quality of simulated Q, including (i) percentage bias (PBIAS; Moriasi et al. 2007); (ii) absolute error (AE); (iii) Pearson’s correlation coefficient (r); and (iv) standard deviation (STD). PBIAS is used to measure the relative bias in the magnitude and trend of simulated Q and AE is used to quantify the absolute bias in the simulated Q timing. Both r and STD are adopted to measure the agreement of variability (both temporally and spatially) between simulations and observations. It is noteworthy that we deliberately refrained from employing more synthesized evaluation metrics, in order to facilitate a more explicit and comprehensive evaluation of model performance from various perspectives.
3. Results
a. Magnitudes of high and low flows
Comparisons of the magnitude between observed and simulated mean annual high flows (QMax and QMax7) across all 633 catchments are summarized in Figs. 2a–d. It is found that nearly all CMIP6 models overestimated the observed high flows in more than 50% of the catchments, which are commonly found in both QMax and QMax7. Most CMIP6 models exhibit a median PBIAS between 15% and 70%, with only CMCC-CM2-HR4 showing a median PBIAS higher than 90%. Since nearly all CMIP6 models show an overestimation of the high-flow magnitudes, the multimodel ensemble mean and median of CMIP6 models also show positive median PBIAS that are larger than most individual models (median PBIAS are ∼63% for the ensemble mean and ∼50% for the ensemble median of CMIP6 models). In comparison, the ISIMIP2a and GLDAS models show both positive and negative biases in the simulated magnitudes of high flows, resulting in median PBIAS of their respective ensemble mean/median typically smaller than 20%.
Evaluation of simulated mean annual magnitude of high and low flows against observations in 633 catchments over 1971–2010. The panels show boxplots of observed and simulated mean annual (a) QMax and (b) QMax7, and boxplots of percentage bias in the simulated (c) QMax and (d) QMax7 across 633 catchments. Also shown are boxplots of observed and simulated mean annual (e) QMin and (f) QMin7, and boxplots of percentage bias in the simulated (g) QMin and (h) QMin7 across 633 catchments. In each plot, boxes indicate the 10th and 90th percentiles, and whiskers represent the minimum and maximum value of all 633 catchments. The dark horizontal line inside each box indicates the median value.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0176.1
In contrast to the high-flow results, the magnitudes of mean annual low flows (i.e., QMin and QMin7) are underestimated by most CMIP6 models except for the three Italian CMCC models that show a positive median PBIAS ranging from 22% to 50% (Figs. 2e–h). For the remaining CMIP6 models, their estimated QMin and QMin7 typically have a median negative bias between −20% and −80% across the 633 catchments. Since most CMIP6 models show a negative bias in the estimated low flows, it is not surprising that the ensembles of CMIP6 models also exhibit a negative bias of about −7% (ensemble mean) and −40% (ensemble median). For the six ISIMIP2a GHMs, two models show a very small negative bias (i.e., DBH and VIC), two models show a relatively larger negative bias (i.e., H08 and LPJML) while the remaining two models show a notable positive bias (i.e., MATSIRO and PCR-GLOBWB), which leads to a median PBIAS in the ensembles of ISIMIP2a models close to zero (about 6% for the ensemble mean and −15% for the ensemble median). For the three GLDAS LSMs, VIC shows a very good performance in estimating the magnitude of low flows (median PBIAS is smaller than −8%), whereas both CLSM and Noah produce a large negative PBIAS exceeding −80%. This has led to the GLDAS model ensembles showing the largest median PBIAS (about −55%) among the three modeling experiments.
When we classified the 633 catchments into different climatic regions, we find that almost all models tend to perform better in estimating the magnitudes of high and low flows in more humid catchments and in noncold regions (Fig. 3). Most models produce a large median PBIAS of >200% in arid and semiarid catchments while the median PBIAS value drops to within ∼30% in humid and subhumid catchments for high flows. The contrast between catchments in cold and noncold regions is also evident (PBIAS are typically within ∼50% and ∼80% for noncold and cold catchments, respectively), although much smaller than that between arid and humid catchments (Figs. 3a,b). Models and regions having a smaller PBIAS generally correspond to a higher r value that indicates a better performance of the model in capturing the spatial variability of the observed high flows. For high flow, the value of r is always higher than 0.85 in humid, subhumid, and noncold regions but drops to <0.5 in arid and semiarid regions and <0.8 in cold areas (Figs. 3e,f). Similar findings are obtained for the two low-flow metrics (Figs. 3c,d,g,h). Although most models show a median underestimation of low flows across all 633 catchments, these models considerably overestimate low flows in relatively dry catchments, with a positive PBIAS larger than 400%. In humid and subhumid catchments, all models (except for DBH in ISIMIP2a) are found to underestimate the observed low flows, with a negative PBIAS generally smaller than −60%. In addition, the observed spatial variability of low flows is well reproduced by all models over humid, subhumid and noncold catchments (r > 0.9) and reasonably captured by most models in semiarid (r > 0.7) and noncold regions (r > 0.5). However, no model captured the observed low-flow spatial variability across the 31 arid catchments (r < 0.1).
Evaluation of simulated mean annual magnitude of high and low flows in different climatic regions. The top four panels show the percentage bias in the simulated mean annual (a) QMax, (b) QMax7, (c) QMin, and (d) QMin7 in different climatic regions. The bottom four panels evaluate the modeled spatial variability of mean annual (e) QMax, (f) QMax7, (g) QMin, and (h) QMin7 in different climatic regions, as indicated by Pearson’s correlation coefficient (r) between observations and simulations across catchments within each climatic region. There are 633 catchments globally and the number of catchments in each climatic region is provided in Fig. 1.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0176.1
b. Temporal changes of high and low flows
We next evaluate the model performance in estimating temporal changes of high and low flows from two perspectives: (i) interannual variability and (ii) annual trend. Figure 4 shows the Pearson’s correlation coefficient between observed and estimated extreme flows over years. It was found that none of the CMIP6 models could capture the observed interannual variability of high and low flows, with a median r typically smaller than 0.1. In comparison, models in ISIMIP2a and GLDAS show improved performance in capturing the observed interannual variability of high and low flows, with a median r higher than 0.4 for high flows and higher than 0.3 for low flows. However, no models show a median r higher than 0.6 for either high or low flows, suggesting that these models can only capture the interannual variability of extreme flow moderately well. Additionally, we find that the multimodel ensemble median generally performs better than most individual models in simulating the interannual variability of extreme flows for the three modeling experiments (yet the improvements are subtle), except for the CMIP6-simulated low flows and Qmax7 and Qmin7 in ISIMIP2a. Note that we did not calculate multimodel ensemble means for the interannual variability of high and low flows as the simple average would smooth the time series and weaken variability.
Evaluation of simulated interannual variability of high and low flows across 633 catchments over 1971–2010. The panels show boxplots of Pearson’s correlation coefficient (r) between observed and simulated annual (a) QMax, (b) QMax7, (c) QMin, and (d) QMin7 over 1971–2010. In each plot, boxes indicate the 10th and 90th percentiles, and whiskers represent the minimum and maximum value of all catchments. The dark horizontal line inside each box indicates the median value.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0176.1
Figure 5, an assessment of interannual variability of extreme flows stratified by climatic regions, showed there is no distinct difference among different climate regions. Nevertheless, ISIMIP2a and GLDAS models tend to perform slightly better in estimating interannual variability of high flows in semiarid and subhumid regions and in noncold catchments and low flows in subhumid and humid regions and in noncold catchments (Fig. 5).
Evaluation of simulated interannual variability of high and low flows in different climatic regions over 1971–2010. The four panels show median Pearson’s correlation coefficient (r) between observed and simulated annual (a) QMax, (b) QMax7, (c) QMin, and (d) QMin7 over 1971–2010 within each climatic region. There are 633 catchments globally and the number of catchments in each climatic region is provided in Fig. 1.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0176.1
Figure 6 compares annual trends of high and low flows between observations and model simulations. During 1971–2010, the observed annual QMax (QMax7) exhibits a marginal median negative trend of −0.24 (−0.25) m3 s−1 yr−1 across 633 catchments, with a positive (negative) trend registered in 298 (335) for QMax and 288 (345) for QMax7, respectively. The observed annual maximum streamflow exhibited statistically significant trends in 86 (QMax) and 76 (QMax7) catchments (p < 0.05, t test). However, among these catchments where statistically significant trends were detected in the observed annual QMax (QMax7), only 7% (7%) of the corresponding annual QMax (QMax7) values simulated by CMIP6 models correctly exhibited the same significant trend directions. In the case of annual QMax (QMax7) outputs from ISIMIP2a and GLDAS models, these proportions increased to 13% (9%) and 10% (18%), respectively (Figs. 6a,b). Additionally, all three modeling experiments largely failed to capture the observed trend of annual high flow—not a single model could accurately capture the sign of the observed high-flow trend in more than 1/3 of the catchments (Figs. 6c,d). For CMIP6 models, the proportion of catchments that show the same sign with the trend of QMax (QMax7) in observation lies in a range from 25% (25%) to 41% (38%), suggesting that these models could only successfully capture the observed high-flow trend in about 1/3 of the catchments. In comparison, this proportion slightly increases to 29% (30%) to 41% (40%) for ISIMIP2a models and 35% (37%) to 40% (39%) for GLDAS models. For catchments where the observed Q trend exhibits statistical significance and the simulated Q correctly reflects the direction of the observed Q trend, the observed high-flow trend is overestimated by most CMIP6 models, resulting in a median PBIAS ranging from −4.9% to 176.5% for QMax and from −2.8% to 128.8% for QMax7. The ISIMIP2a and GLDAS models show both overestimation and underestimation of the observed high-flow trend with a median PBIAS typically ranging between −49% and 23%, except for a large positive bias obtained for ISIMIP2a LPJML and GLDAS VIC (PBIAS is ∼85% for both QMax and QMax7).
Evaluation of simulated annual trends in high and low flows across 633 catchments over 1971–2010. The panels show boxplots of observed and simulated annual trend in (a) QMax, (b) QMax7, (e) QMin, and (f) QMin7 and boxplots of percentage bias in the simulated annual trend in (c) QMax, (d) QMax7, (g) QMin, and (h) QMin7. The blue triangle in (a), (b), (e), and (f) indicates the number of catchments where the Q trend is statistically significant. The blue cross in (a), (b), (e), and (f) indicates the number of catchments where both simulated and observed Q is statistically significant and the sign of observed Q is corrected by the modeled Q. In (c), (d), (g), and (h), the blue circles represent the count of catchments in which the observed Q trend is statistically significant and the direction of the simulated Q trend aligns correctly with the observed Q trend. In each plot, boxes indicate the 10th and 90th percentiles, and whiskers represent the minimum and maximum value of all selected catchments. The median value is shown by the dark horizontal line inside each box.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0176.1
Similar findings were obtained for the low-flow trends. The observed annual QMin (QMin7) trend with statistically significant at 172 (166) catchments (p < 0.05, t test). The same sign of significant trends could only be correctly captured in about 12% (10%) of them by the simulated Q form CMIP6 models. However, this proportion does not increase significantly for Q outputs from ISIMIPP2a and GLDAS models, reaching only about 13% (13%) and 10% (12%) (Figs. 6e,f). The models could only correctly capture the sign of observed trend in about 1/3 of the catchments [i.e., from 21% (23%) to 32% (33%) for CMIP6, from 27% (28%) to 32% (34%) for ISIMIP2a, and from 30% (30%) to 33% (33%) for GLDAS, for QMin (QMin7)] (Figs. 6g,h). For catchments where the sign of observed Q trend (statistically significant) is correctly simulated, nearly all models show an underestimation of the observed low-flow trend (median PBIAS ranges from −13% to −98% for QMin and from −12% to −99% for QMin7) except for a small positive median bias in CMIP6 CMCC-CM2-SR5, ISIMIP2a PCR-GLOBWB, and GLDAS VIC (35%, 11%, and 14% for QMin and 44%, 45%, and 34% for QMin7). In addition, no obvious difference is found between the three modeling experiments regarding the model performance in simulating the magnitude of the low-flow trend (Figs. 6e,f).
Examining the modeled annual trends of extreme flows in different climatic regions reveals a similar pattern as that for the magnitude of extreme flows. Our results showed that nearly all models exhibit better performance in more humid catchments (Figs. 7a–d). In arid and semiarid regions, more than half of the models produced a PBIAS higher than ±200%, and the typical PBIAS in subhumid and humid catchments ranged between −90% and +50%, for both high and low flows. However, the contrast between cold and noncold catchments is subtle. Figures 7e–h show the model performance in capturing the observed spatial variability of extreme flow trends in different climatic regions. It shows that for ISIMIP2a and GLDAS models, the observed spatial variability of annual high-flow trends were well captured in subhumid, humid and noncold regions (r > 0.75, except for GLDAS Noah) but not captured in semiarid and cold catchments (r < 0.2). For the spatial variability of annual low flow trends, the lower r values were generally found in arid (r < 0.1), subhumid (r < 0.3), and cold (r < 0.5) catchments. For other regions, the value of r is higher than 0.8 for all ISIMIP2a and GLDAS models (Figs. 7g,h). In comparison, the CMIP6 models generally failed to capture the observed spatial variability of extreme flow trends in all climatic regions except for a few models that showed good skill in capturing the observed spatial variability of annual low-flow trends in relatively humid environments (Figs. 7g,h).
Evaluation of simulated annual trends in high and low flows in different climatic regions over 1971–2010. The top four panels show the percentage bias in the simulated annual trend in (a) QMax, (b) QMax7, (c) QMin, and (d) QMin7 in different climatic regions. The bottom four panels evaluate the modeled spatial variability of annual trend in (e) QMax, (f) QMax7, (g) QMin, and (h) QMin7 in different climatic regions, as indicated by Pearson’s correlation coefficient (r) between observations and simulations across catchments within each climatic region. There are 633 catchments globally, and the number of catchments in each climatic region is provided in Fig. 1.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0176.1
c. Timing of high and low flows
The date within each calendar year when extreme flows occur as simulated by the models was evaluated in Figs. 8 and 9 (the beginning date for QMax7 and QMin7). For high flow all CMIP6 and GLDAS models produce a median early bias ranging from 8 (GLDAS VIC) to 20 days (CMIP6 AWI-ESM-1-1-LR) across the 633 catchments, whereas the ISIMIP2a models generally show a median late bias of 6–23 days except for LPJML (Figs. 8a,b). For low flows, the GLDAS models consistently show a median early bias (10–17 days) whereas most CMIP6 models exhibit a median late bias ranging from 2 to 17 days except for three models that show a small median early bias (i.e., FGOALS-g3, IITM-ESM, and NorESM2-MM) (Figs. 8c,d). In comparison, the ISIMIP2a shows early and late biases in an equal number of models, which leads to a negligible bias in the multimodel ensemble mean or median estimate. It is pertinent to highlight that the timing of streamflow inherently exhibits variation between the Northern Hemisphere (NH) and the Southern Hemisphere (SH). However, in our assessment, biases are evaluated relative to local references—specifically, the observed streamflow timing for each catchment. Consequently, the inherent divergence in streamflow timing between the NH and SH does not compromise the integrity of our overarching findings, particularly when considering the comprehensive evaluation of all catchments collectively (Figs. S5 and S6).
Evaluation of simulated mean annual dates of high and low flows across 633 catchments over 1971–2010. The panels show boxplots of absolute bias in the simulated mean annual date of (a) TQ_Max, (b) TQ_Max7, (c) TQ_Min, and (d) TQ_Min7. In each plot, boxes indicate the 10th and 90th percentiles and whiskers represent the minimum and maximum value of all selected catchments. The median value is shown by the dark horizontal line inside each box.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0176.1
Evaluation of simulated mean annual dates of high and low flows in different climatic regions over 1971–2010. The four panels show absolute bias in the simulated mean annual (a) TQ_Max, (b) TQ_Max7, (c) TQ_Min, and (d) TQ_Min7 in different climatic regions. There are 633 catchments globally, and the number of catchments in each climatic region is provided in Fig. 1.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0176.1
Figure 9 shows the bias in simulated extreme flows for different climatic regions. For high flow, the CMIP6 models consistently show an early bias and the ISIMIP2a models (except for LPJML) consistently exhibit a late bias in the estimated timing of high flow in different climatic regions (Figs. 9a,b). A more complex pattern is found for the estimated timing of high flows in GLDAS models. Despite an overall early bias obtained over all 633 catchments, the GLDAS models generally show a late (yet small) bias in the estimated timing of high flow in arid and noncold catchments (Figs. 9a,b). For low flow, a large late bias is found in all models in arid regions and most models show a late bias in semiarid catchments (Figs. 9c,d). Apart from that, no clear patterns are found for the estimated timing of low flow in other climatic regions.
Figure 10 presents a comparative analysis of annual timing trends for high and low flows between observations and model simulations over 1971–2010. The findings reveal that, concerning high flows, the three modeling experiments exhibit successful alignment with the observed sign of trends in annual TQ_Max (TQ_Max7) timing in approximately 50% (33%) of the catchments (Figs. 10a–d). Notably, for catchments where the observed annual TQ_Max (TQ_Max7) timing trends are correctly captured by the simulated Q, the three modeling experiments consistently manifest a tendency to underestimate the observed trend in high-flow timing, particularly evident in TQ_Max7 (Figs. 10c,d). Analogous observations regarding model performance are discerned in the simulation of low-flow timing trends (Figs. 10e–h). Furthermore, this tendency toward underestimation in trends of extreme flow timings is prevalent across diverse climatic regions in all three modeling experiments (Fig. S7).
Evaluation of simulated annual timing trends in high and low flows across 633 catchments over 1971–2010. The panels show boxplots of observed and simulated annual trend in (a) TQ_Max, (b) TQ_Max7, (e) TQ_Min, and (f) TQ_Min7, and boxplots of percentage bias in the simulated annual trend in (c) TQ_Max, (d) TQ_Max7, (g) TQ_Min, and (h) TQ_Min7. In (c), (d), (g), and (h), blue circles indicate the number of catchments where the sign of observed trend is correctly captured by the simulated trend, with PBIAS calculated for these specific catchments. In each plot, boxes indicate the 10th and 90th percentiles, and whiskers represent the minimum and maximum values across all selected catchments. The median value is denoted by the dark horizontal line within each box.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0176.1
d. Uncertainty assessment
The principal contributors to the uncertainties in simulated hydrological extremes primarily emanate from two sources: the structural and parameterization of the model, as well as the meteorological forcing datasets (Müller Schmied et al. 2014; Zhao et al. 2017). To facilitate a systematic exploration of these uncertainties, we employ a consistent meteorological forcing dataset, namely, PGFv2, in both ISIMIP2a and GLDAS models. This approach allows us to quantify the impact of model structure/parameterization in these two modeling experiments. Furthermore, by employing three distinct meteorological forcings—PGFv2, GSWP3, and WATCH/WFDEI—we are able to quantify the uncertainties associated with meteorological forcing in ISIMIP2a. The “forcing uncertainty” is quantified as the mean standard deviation (STD) of simulation bias when the same ISIMIP2a GHMs are driven by different meteorological datasets. Specifically, for instance, the forcing uncertainty for model A in catchment B is computed as the STD of simulation biases resulting from model A being supplied with three different meteorological forcings—namely, PGFv2, GSWP3, and WATCH/WFDEI. Concurrently, we characterize “model uncertainty” as the STD of simulation bias among models subjected to the same meteorological forcing (i.e., PGFv2).
Our findings reveal that the relative STD attributable to different meteorological forcings is approximately 50% for both the magnitude and interannual variability of high and low flows. Additionally, the timing of high and low flows exhibits an STD of approximately 19 days due to varying meteorological forcings (Fig. 11). It is noteworthy that these uncertainties stemming from meteorological forcing are generally smaller when compared to the uncertainties arising from differing model structures and parameterizations in the combined analysis of ISIMIP2a and GLDAS models. Specifically, the median relative STD is approximately 65% for the magnitude of QMax and QMin, around 70% for the interannual variability of QMax and QMin, and the median STD is approximately 28 days for the timing of QMax and QMin (Fig. 11).
Model vs forcing uncertainties across the 633 catchments. In each plot, boxes indicate the 10th and 90th percentiles and whiskers represent the minimum and maximum value of all catchments. The median value is shown by the dark horizontal line inside each box.
Citation: Journal of Hydrometeorology 25, 3; 10.1175/JHM-D-23-0176.1
However, upon closer examination of model uncertainties within ISIMIP2a and GLDAS models individually, we observe that the substantial relative STD (or STD) among models within the combined group of ISIMIP2a and GLDAS primarily arises from a relatively pronounced intermodel divergence in ISIMIP2a models, while the intermodel variability in GLDAS models is comparatively limited across all three aspects (approaching the magnitude of forcing uncertainty). To fortify the robustness of our uncertainty analyses, we conducted a random selection of three out of the six ISIMIP2a models, resulting in a total of 20 combinations. The comparison of these combinations with GLDAS models reaffirmed the tenability of our conclusions (Figs. S7 and S8). It is pertinent to note that our presentation here pertains to QMax and QMin, and similar findings are applicable to QMax7 and QMin7, as depicted in Fig. S10.
4. Discussion
The utilization of simulated Q has become commonplace in the analysis of hydrological extremes, encompassing both high-flow and low-flow events, as well as in the assessment of historical and prospective alterations in floods and droughts on global or regional scales (Hirabayashi et al. 2013; Giuntoli et al. 2015; Gudmundsson et al. 2012; Liu et al. 2021; Mester et al. 2021; Zaherpour et al. 2018). Given the substantial economic and life-threatening implications associated with hydrological extremes (Dethier et al. 2020), it is imperative to possess accurate estimations of their occurrence likelihood and the potential shifts therein. Unfortunately, our evaluation underscores significant disparities between the characteristics of simulated hydrological extremes generated by CMIP6, ISIMIP2a, and GLDAS models when contrasted with observed data. This implies that assessments of water-related disasters in a changing climate, founded on predictions from these cutting-edge macroscale models, remain encumbered by substantial uncertainties.
Specifically, our findings reveal that most models tend to overestimate the magnitude of high flow and underestimate low flow (Figs. 2 and 3). This phenomenon suggests an inclination of these models to exaggerate the annual range of observed Q, which, in turn, may lead to an overestimation of flood and drought risks inferred from their outputs. Furthermore, our analysis indicates that none of the models are consistently able to accurately discern the direction of observed trends in annual high or low flows in more than two-thirds of the catchments (Figs. 6a,b,e,f). This underscores a substantial uncertainty in predicting even the directional changes in hydrological extremes over time based on model outputs. In cases where the models do manage to capture the direction of observed trends correctly, they still tend to overestimate high-flow trends while simultaneously underestimating low-flow trends (Figs. 6c,d,g,h). Given mounting evidence suggesting that anthropogenic climate change is likely to result in concurrent increases in both floods and droughts on a global scale (De Luca et al. 2020; Zhai et al. 2020), our assessments suggest that these models may overstate the escalation in flood risks and underestimate the amplification of hydrological drought risks. In terms of the timing, CMIP6 and GLDAS models display a tendency toward early biases in high-flow timing and late biases in low-flow timing, whereas ISIMIP2a models generally exhibit the opposite pattern (Fig. 8). These timing discrepancies directly translate into inaccuracies in predicting the onset of flood and drought events, thereby posing additional challenges for disaster preparedness and mitigation efforts.
Previous investigations have posited that meteorological forcing predominantly influences high-flow events, while low-flow dynamics are more regulated by terrestrial hydrological processes (Beck et al. 2017a). By quantifying the sources of uncertainty in the simulation of hydrological extremes, our findings offer partial substantiation of this prior assertion. Notably, our analysis reveals that the uncertainty arising from model structures/parameterizations is notably more pronounced in the simulation of low-flow events compared to high-flow counterparts. Conversely, the uncertainty stemming from meteorological forcings exhibits a slightly higher magnitude for high flows relative to low flows (Fig. 11). However, it is imperative to underscore that across all dimensions of extreme flow characteristics examined herein, the uncertainties introduced by model structures/parameterizations consistently surpass those attributed to meteorological forcings (Fig. 11), indicating that the primary driver behind biases in the estimation of extreme flows lies within the realm of model structures/parameterizations.
It is essential to note that the uncertainty analysis presented in Fig. 10 does not encompass CMIP6 outputs. This omission is noteworthy because ISIMIP2a and GLDAS models rely on observation-guided reanalysis of meteorological forcings (Rodell et al. 2004), whereas CMIP6 employs simulated meteorological fields, which are frequently afflicted by substantial uncertainties due to uncontrolled internal climate variations (Deser et al. 2012; Hegerl et al. 2021; Sippel et al. 2021). Consequently, it is not surprising that ISIMIP2a and GLDAS models yield more accurate estimates of extreme flows compared to CMIP6 models. The disparity between CMIP6 and ISIMIP2a/GLDAS estimates predominantly arises from the biases in the simulated meteorological forcings within CMIP6 models, because CMIP6 models typically adopt analogous approaches for simulating surface hydrology as GHMs and/or LSMs. Furthermore, we find that the biases in CMIP6-simulated high flows are generally more pronounced than those in low flows (Figs. 2 and 8), which again underscores the higher sensitivity of high flows to meteorological forcings.
Beyond the overarching evaluation of model performance across all 633 catchments, it is apparent that these models generally exhibit more favorable performance in regions characterized by more humid and noncold climates (Figs. 3, 5, 7, and 9). This corroborates earlier research findings concerning the assessment of extreme flow estimates (Gudmundsson et al. 2012) and mean flow estimates (Yang et al. 2015; Hou et al. 2023). The relatively diminished model performance in arid and cold regions can be primarily attributed to two factors: (i) uncertainties in the available forcing data and (ii) deficiencies in model structure and parameterizations. Regarding the forcing data, prior investigations have demonstrated that reanalyzed precipitation data often exhibit subpar performance in arid regions, owing to challenges in capturing localized and short-lived convective rainfall patterns prevalent in such areas (Beck et al. 2017b; Cecil et al. 2014). In cold regions, the underestimation of precipitation in snowfall-dominated regions, commonly referred to as the undercatch issue (Müller Schmied et al. 2016), and the limited availability of rain gauge data for bias correction (Beck et al. 2017a,b) present significant hurdles for climate reanalyses in establishing high-quality precipitation datasets.
Concerning model-related issues, the majority of ISIMIP2a and GLDAS models (i.e., H08, LPJmL, PCR-GLOBWB, Noah, CLSM) employ the saturation-excess mechanism for runoff generation, which often proves inadequate for arid regions where surface runoff typically results from the infiltration-excess mechanism (Dunne 1978; Pilgrim et al. 1988). Notably, the MATSIRO model within ISIMIP2a stands as an exception by considering both saturation-excess and infiltration-excess runoff mechanisms. Consequently, the performance of MATSIRO does not exhibit a clear pattern in high-flow events that directly correspond to precipitation along the dry–wet gradient (Figs. 3a,b). Moreover, the relatively suboptimal model performance in cold regions indicates that these models may struggle to accurately capture hydrological processes associated with the cryosphere. This deficiency includes overlooking the influence of soil ice on soil hydraulic properties (Swenson et al. 2012) and demonstrating limited proficiency in reproducing the timing of snowmelt in cold regions (Gosling and Arnell 2011).
In addition to evaluating individual model performance, we have also conducted an assessment of the effectiveness of multimodel ensembles, specifically the ensemble mean and median, in capturing hydrological extremes. Multimodel ensembles are widely acknowledged for their capacity to mitigate uncertainties inherent in individual models (Gudmundsson et al. 2021; Liu et al. 2022; Yang et al. 2015). Our findings generally corroborate these prior findings, revealing that simple averages or the model median consistently outperform most individual models within each modeling group for the majority of the hydrological extreme characteristics examined herein. Nevertheless, certain exceptions are noted in specific instances, such as the ensemble mean/median of CMIP6-simulated magnitudes of high-flow events (Figs. 2c,d), the ensemble median of ISIMIP2a-simulated interannual variability of QMax7 and QMin7 (Figs. 4b,d), the ensemble mean/median of CMIP6-simulated trends in low-flow events (Figs. 6g,h), and the ensemble mean/median of CMIP6-simulated timing of high-flow events (Figs. 8a,b). These outcomes imply that relying solely on simple ensemble mean or median calculations may not universally suffice. In such cases, more sophisticated ensemble methodologies, such as Bayesian averages (Raftery et al. 2005) or weighted averages (Kulinich et al. 2021), may offer more advantageous approaches for enhancing the reliability of hydrological extreme estimations.
5. Conclusions
In this study, we conducted a comprehensive assessment of simulated extreme flow characteristics, including magnitude, interannual variability, trend, and timing, utilizing 23 macroscale models integrated with the state-of-the-art CaMa-Flood river routing scheme. These assessments were performed in comparison with observations obtained from 633 unimpaired catchments across the globe. Our primary findings are summarized as follows:
-
In general, models relying on observation-based climate reanalyses, namely, ISIMIP2a and GLDAS, outperform CMIP6 models in estimating a wide range of extreme flow characteristics. Furthermore, all models demonstrate superior performance in regions characterized by higher humidity and in catchments with noncold climates.
-
Most models exhibit an overestimation of mean annual high-flow magnitudes and an underestimation of low-flow magnitudes. The ISIMIP2a and GLDAS models reasonably capture the interannual variability of both high and low flows, whereas this aspect is generally inadequately represented by CMIP6 models.
-
None of the models achieve satisfactory results in replicating the annual trends in high and low flows, with the observed directional trends being incorrectly simulated in at least two-thirds of the catchments. In instances where the trends are accurately captured in both observation and simulation, most models tend to overestimate high-flow trends while underestimating low-flow trends.
-
The CMIP6 and GLDAS models exhibit an early timing bias in high-flow events and a delayed timing bias in low-flow events, while ISIMIP2a models generally display a delayed timing bias in high-flow events and an early timing bias in low-flow events.
-
For ISIMIP2a and GLDAS models, which rely on reanalyzed meteorological forcings, uncertainties in model structures/parameterizations tend to introduce larger uncertainties in the modeling of hydrological extremes compared to those arising from uncertainties in meteorological forcings. Additionally, meteorological forcing uncertainties have a more pronounced impact on high-flow simulations compared to low-flow simulations, whereas model structure/parameterization uncertainties play a more significant role in the modeling of low-flow events as opposed to high-flow events.
Acknowledgments.
The authors declare no conflicts of interest relevant to this study. This study is financially supported by the Ministry of Science and Technology of China (Grant 2023YFC3206603), the National Natural Science Foundation of China (Grant 42071029), and the Department of Science and Technology of Yunnan Province (Grant 202203AA080010).
Data availability statement.
All data for this paper are properly cited and referred to in the reference list. Specifically, observed streamflow data are available from (i) the USGS National Water Information system (Falcone et al. 2010) (select “streamflow, ft3/s” at https://waterdata.usgs.gov/nwis/dv?referred_module=sw&search_criteria=huc2_cd&search_criteria=site_tp_cd&submitted_form=introduction; or see tutorials in https://help.waterdata.usgs.gov/tutorials/surface-water-data/how-do-i-access-historical-streamflow-data), (ii) the Global Runoff Data Centre (Lehner 2012) (https://portal.grdc.bafg.de/applications/public.html?publicuser=PublicUser#dataDownload/Home), (iii) the HidroWeb portal of the Brazilian Agência Nacional de águas (select all the hydrological stations on the HidrowED portal and export all streamflow data from the “Options” menu at https://www.snirh.gov.br/hidroweb/mapa), (iv) the European Water Archive of the European Flow Regimes from International Experimental and Network Data (select daily flow data for all the stations at https://nrfa.ceh.ac.uk/data/download-all-station-metadata), (v) the Water Survey of Canada Hydrometric Data (daily flow) at https://wateroffice.ec.gc.ca/search/historical_e.html through filter “WSC region-Canada,” (vi) the Australian Bureau of Meteorology (Zhang et al. 2013) (http://www.bom.gov.au/water/hrs/#id=609010&panel=data-explorer&pill=daily), and (vii) the Chilean Center for Climate and Resilience Research (https://explorador.cr2.cl/). The CMIP6, ISIMIP2a and GLDAS runoff datasets were downloaded from the ESGF-LLNL data node (select historical daily “mrro” data from “Variable” menu at https://esgf-node.llnl.gov/search/cmip6/), the ISIMIP Repository (select “Water (global)-qtot daily” data from “ISIMIP2a simulation” menu at https://data.isimip.org/search/) and the NASA Goddard Earth Sciences Data and Information Services Center (https://disc.gsfc.nasa.gov/datasets/), respectively. All websites were last accessed on 15 November 2022.
REFERENCES
Alkama, R., B. Decharme, H. Douville, and A. Ribes, 2011: Trends in global and basin-scale runoff over the late twentieth century: Methodological issues and sources of uncertainty. J. Climate, 24, 3000–3014, https://doi.org/10.1175/2010JCLI3921.1.
Beck, H. E., A. de Roo, and A. I. J. M. van Dijk, 2015: Global maps of streamflow characteristics based on observations from several thousand catchments. J. Hydrometeor., 16, 1478–1501, https://doi.org/10.1175/JHM-D-14-0155.1.
Beck, H. E., and Coauthors, 2017a: Global-scale evaluation of 22 precipitation datasets using gauge observations and hydrological modeling. Hydrol. Earth Syst. Sci., 21, 6201–6217, https://doi.org/10.5194/hess-21-6201-2017.
Beck, H. E., A. I. J. M. van Dijk, A. de Roo, E. Dutra, G. Fink, R. Orth, and J. Schellekens, 2017b: Global evaluation of runoff from 10 state-of-the-art hydrological models. Hydrol. Earth Syst. Sci., 21, 2881–2903, https://doi.org/10.5194/hess-21-2881-2017.
Beevers, L., L. Collet, G. Aitken, C. Maravat, and A. Visser, 2020: The influence of climate model uncertainty on fluvial flood hazard estimation. Nat. Hazards, 104, 2489–2510, https://doi.org/10.1007/s11069-020-04282-4.
Blackburn, S. R., and E. H. Stanley, 2021: Floods increase carbon dioxide and methane fluxes in agricultural streams. Freshwater Biol., 66, 62–77, https://doi.org/10.1111/fwb.13614.
Bondeau, A., and Coauthors, 2007: Modelling the role of agriculture for the 20th century global terrestrial carbon balance. Global Change Biol., 13, 679–706, https://doi.org/10.1111/j.1365-2486.2006.01305.x.
Cecil, D. J., D. E. Buechler, and R. J. Blakeslee, 2014: Gridded lightning climatology from TRMM-LIS and OTD: Dataset description. Atmos. Res., 135–136, 404–414, https://doi.org/10.1016/j.atmosres.2012.06.028.
Chaney, N. W., J. D. Herman, P. M. Reed, and E. F. Wood, 2015: Flood and drought hydrologic monitoring: The role of model parameter uncertainty. Hydrol. Earth Syst. Sci., 19, 3239–3251, https://doi.org/10.5194/hess-19-3239-2015.
Chen, F., and Coauthors, 1996: Modeling of land surface evaporation by four schemes and comparison with FIFE observations. J. Geophys. Res., 101, 7251–7268, https://doi.org/10.1029/95JD02165.
Cherchi, A., and Coauthors, 2019: Global mean climate and main patterns of variability in the CMCC‐CM2 coupled model. J. Adv. Model. Earth Syst., 11, 185–209, https://doi.org/10.1029/2018MS001369.
Chiang, F., O. Mazdiyasni, and A. AghaKouchak, 2021: Evidence of anthropogenic impacts on global drought frequency, duration, and intensity. Nat. Commun., 12, 2754, https://doi.org/10.1038/s41467-021-22314-w.
Cutore, P., G. Cristaudo, A. Campisano, C. Modica, A. Cancelliere, and G. Rossi, 2007: Regional models for the estimation of streamflow series in ungauged basins. Water Resour. Manage., 21, 789–800, https://doi.org/10.1007/s11269-006-9110-7.
De Luca, P., G. Messori, R. L. Wilby, M. Mazzoleni, and G. Di Baldassarre, 2020: Concurrent wet and dry hydrological extremes at the global scale. Earth Syst. Dyn., 11, 251–266, https://doi.org/10.5194/esd-11-251-2020.
Deser, C., A. Phillips, V. Bourdette, and H. Teng, 2012: Uncertainty in climate change projections: The role of internal variability. Climate Dyn., 38, 527–546, https://doi.org/10.1007/s00382-010-0977-x.
Dethier, E. N., S. L. Sartain, C. E. Renshaw, and F. J. Magilligan, 2020: Spatially coherent regional changes in seasonal extreme streamflow events in the United States and Canada since 1950. Sci. Adv., 6, eaba593, https://doi.org/10.1126/sciadv.aba5939.
Dottori, F., and Coauthors, 2018: Increased human and economic losses from river flooding with anthropogenic warming. Nat. Climate Change, 8, 781–786, https://doi.org/10.1038/s41558-018-0257-z.
Dunne, T., 1978: Field studies of hillslope flow processes. Hillslope Hydrology, John Wiley and Sons, 227–293, https://www.researchgate.net/publication/243780862.
Eyring, V., S. Bony, G. A. Meehl, C. A. Senior, B. Stevens, R. J. Stouffer, and K. E. Taylor, 2016: Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6) experimental design and organization. Geosci. Model Dev., 9, 1937–1958, https://doi.org/10.5194/gmd-9-1937-2016.
Falcone, J. A., D. M. Carlisle, D. M. Wolock, and M. R. Meador, 2010: GAGES: A stream gage database for evaluating natural and altered flow conditions in the conterminous United States. Ecology, 91, 621–621, https://doi.org/10.1890/09-0889.1.
Fekete, B. M., and C. J. Vörösmarty, 2007: The current status of global river discharge monitoring and potential new technologies complementing traditional discharge measurements. IAHS Publ., 309, 129–136.
Frieler, K., and Coauthors, 2017: Assessing the impacts of 1.5°C global warming – simulation protocol of the Inter-Sectoral Impact Model Intercomparison Project (ISIMIP2b). Geosci. Model Dev., 10, 4321–4345, https://doi.org/10.5194/gmd-10-4321-2017.
Fyfe, J. C., V. V. Kharin, B. D. Santer, J. N. S. Cole, and N. P. Gillett, 2021: Significant impact of forcing uncertainty in a large ensemble of climate model simulations. Proc. Natl. Acad. Sci. USA, 118, e2016549118, https://doi.org/10.1073/pnas.2016549118.
Giuntoli, I., J.-P. Vidal, C. Prudhomme, and D. M. Hannah, 2015: Future hydrological extremes: The uncertainty from multiple global climate and global hydrological models. Earth Syst. Dyn., 6, 267–285, https://doi.org/10.5194/esd-6-267-2015.
Gleason, C. J., L. C. Smith, and J. Lee, 2014: Retrieval of river discharge solely from satellite imagery and at‐many‐stations hydraulic geometry: Sensitivity to river form and optimization parameters. Water Resour. Res., 50, 9604–9619, https://doi.org/10.1002/2014WR016109.
Gosling, S. N., and N. W. Arnell, 2011: Simulating current global river runoff with a global hydrological model: Model revisions, validation, and sensitivity analysis. Hydrol. Processes, 25, 1129–1145, https://doi.org/10.1002/hyp.7727.
Gudmundsson, L., and Coauthors, 2012: Comparing large-scale hydrological model simulations to observed runoff percentiles in Europe. J. Hydrometeor., 13, 604–620, https://doi.org/10.1175/JHM-D-11-083.1.
Gudmundsson, L., and Coauthors, 2021: Globally observed trends in mean and extreme river flow attributed to climate change. Science, 371, 1159–1162, https://doi.org/10.1126/science.aba3996.
Gutjahr, O., D. Putrasahan, K. Lohmann, J. H. Jungclaus, J.-S. von Storch, N. Brüggemann, H. Haak, and A. Stössel, 2019: Max Planck Institute Earth System Model (MPI-ESM1.2) for the High-Resolution Model Intercomparison Project (HighResMIP). Geosci. Model Dev., 12, 3241–3281, https://doi.org/10.5194/gmd-12-3241-2019.
Hanasaki, N., S. Kanae, T. Oki, K. Masuda, K. Motoya, N. Shirakawa, Y. Shen, and K. Tanaka, 2008: An integrated model for the assessment of global water resources – Part 1: Model description and input meteorological forcing. Hydrol. Earth Syst. Sci., 12, 1007–1025, https://doi.org/10.5194/hess-12-1007-2008.
He, B., and Coauthors, 2019: CAS FGOALS-f3-L model datasets for CMIP6 historical atmospheric model intercomparison project simulation. Adv. Atmos. Sci., 36, 771–778, https://doi.org/10.1007/s00376-019-9027-8.
Hegerl, G. C., and Coauthors, 2021: Toward consistent observational constraints in climate predictions and projections. Front. Climate, 3, 6781009, https://doi.org/10.3389/fclim.2021.678109.
Hendrawan, V. S. A., W. Kim, Y. Touge, S. Ke, and D. Komori, 2022: A global-scale relationship between crop yield anomaly and multiscalar drought index based on multiple precipitation data. Environ. Res. Lett., 17, 014037, https://doi.org/10.1088/1748-9326/ac45b4.
Hirabayashi, Y., S. Kanae, S. Emori, T. Oki, and M. Kimoto, 2008: Global projections of changing risks of floods and droughts in a changing climate. Hydrol. Sci. J., 53, 754–772, https://doi.org/10.1623/hysj.53.4.754.
Hirabayashi, Y., R. Mahendran, S. Koirala, L. Konoshima, D. Yamazaki, S. Watanabe, H. Kim, and S. Kanae, 2013: Global flood risk under climate change. Nat. Climate Change, 3, 816–821, https://doi.org/10.1038/nclimate1911.
Hirabayashi, Y., H. Alifu, D. Yamazaki, Y. Imada, H. Shiogama, and Y. Kimura, 2021: Anthropogenic climate change has changed frequency of past flood during 2010–2013. Prog. Earth Planet. Sci., 8, 36, https://doi.org/10.1186/s40645-021-00431-w.
Hou, Y., H. Guo, Y. Yang, and W. Liu, 2023: Global evaluation of runoff simulation from climate, hydrological and land surface models. Water Resour. Res., 59, e2021WR031817, https://doi.org/10.1029/2021WR031817.
IPCC, 2021: Climate Change 2021: The Physical Science Basis. Cambridge University Press, 2391 pp., https://www.ipcc.ch/report/ar6/wg1/.
Jones, P. W., 1999: First- and second-order conservative remapping schemes for grids in spherical coordinates. Mon. Wea. Rev., 127, 2204–2210, https://doi.org/10.1175/1520-0493(1999)127<2204:FASOCR>2.0.CO;2.
Kim, H., 2017: Global soil wetness project phase 3 atmospheric boundary conditions (experiment 1). Data Integration and Analysis System (DIAS), accessed 17 June 2022, https://search.diasjp.net/en/dataset/GSWP3_EXP1_Forcing.
Koster, R. D., M. J. Suarez, A. Ducharne, M. Stieglitz, and P. Kumar, 2000: A catchment‐based approach to modeling land surface processes in a general circulation model: 1. Model structure. J. Geophys. Res., 105, 24 809–24 822, https://doi.org/10.1029/2000JD900327.
Kreibich, H., and Coauthors, 2022: The challenge of unprecedented floods and droughts in risk management. Nature, 608, 80–86, https://doi.org/10.1038/s41586-022-04917-5.
Kulinich, M., Y. Fan, S. Penev, J. P. Evans, and R. Olson, 2021: A Markov chain method for weighting climate model ensembles. Geosci. Model Dev., 14, 3539–3551, https://doi.org/10.5194/gmd-14-3539-2021.
Lehner, B., 2012: Derivation of watershed boundaries for GRDC gauging stations based on the HydroSHEDS drainage network. GRDC Tech. Rep. 41, 18 pp., https://www.bafg.de/GRDC/EN/02_srvcs/24_rprtsrs/report_41.pdf?__blob=publicationFile.
Lehner, B., and Coauthors, 2011: High‐resolution mapping of the world’s reservoirs and dams for sustainable river‐flow management. Front. Ecol. Environ., 9, 494–502, https://doi.org/10.1890/100125.
Li, Z., H. Zhao, J. Liu, J. Zhang, and Z. Shao, 2022: Evaluation and promotion strategy of resilience of urban water supply system under flood and drought disasters. Sci. Rep., 12, 7404, https://doi.org/10.1038/s41598-022-11436-w.
Liang, X., D. P. Lettenmaier, E. F. Wood, and S. J. Burges, 1994: A simple hydrologically based model of land surface water and energy fluxes for general circulation models. J. Geophys. Res., 99, 14 415–14 428, https://doi.org/10.1029/94JD00483.
Lim, W. H., D. Yamazaki, S. Koirala, Y. Hirabayashi, S. Kanae, S. J. Dadson, J. W. Hall, and F. Sun, 2018: Long‐term changes in global socioeconomic benefits of flood defenses and residual risk based on CMIP5 climate models. Earth’s Future, 6, 938–954, https://doi.org/10.1002/2017EF000671.
Liu, W., T. Yang, F. Sun, H. Wang, Y. Feng, and M. Du, 2021: Observation‐constrained projection of global flood magnitudes with anthropogenic warming. Water Resour. Res., 57, e2020WR028830, https://doi.org/10.1029/2020WR028830.
Liu, Z., J. Huang, X. Xiao, and X. Tong, 2022: The capability of CMIP6 models on seasonal precipitation extremes over Central Asia. Atmos. Res., 278, 106364, https://doi.org/10.1016/j.atmosres.2022.106364.
Mauritsen, T., and Coauthors, 2019: Developments in the MPI‐M Earth System Model version 1.2 (MPI‐ESM1.2) and its response to increasing CO2. J. Adv. Model. Earth Syst., 11, 998–1038, https://doi.org/10.1029/2018MS001400.
McVicar, T. R., M. L. Roderick, R. J. Donohue, and T. G. Van Niel, 2012: Less bluster ahead? Ecohydrological implications of global trends of terrestrial near‐surface wind speeds. Ecohydrology, 5, 381–388, https://doi.org/10.1002/eco.1298.
Meresa, H., C. Murphy, R. Fealy, and S. Golian, 2021: Uncertainties and their interaction in flood hazard assessment with climate change. Hydrol. Earth Syst. Sci., 25, 5237–5257, https://doi.org/10.5194/hess-25-5237-2021.
Mester, B., S. N. Willner, K. Frieler, and J. Schewe, 2021: Evaluation of river flood extent simulated with multiple global hydrological models and climate forcings. Environ. Res. Lett., 16, 094010, https://doi.org/10.1088/1748-9326/ac188d.
Milly, P. C. D., J. Betancourt, M. Falkenmark, R. M. Hirsch, Z. W. Kundzewicz, D. P. Lettenmaier, and R. J. Stouffer, 2008: Stationarity is dead: Whither water management? Science, 319, 573–574, https://doi.org/10.1126/science.1151915.
Mockler, E. M., K. P. Chun, G. Sapriza-Azuri, M. Bruen, and H. S. Wheater, 2016: Assessing the relative importance of parameter and forcing uncertainty and their interactions in conceptual hydrological model simulations. Adv. Water Resour., 97, 299–313, https://doi.org/10.1016/j.advwatres.2016.10.008.
Moriasi, D. N., J. G. Arnold, M. W. Van Liew, R. L. Bingner, R. D. Harmel, and T. L. Veith, 2007: Model evaluation guidelines for systematic quantification of accuracy in watershed simulations. Trans. ASABE, 50, 885–900, https://doi.org/10.13031/2013.23153.
Müller Schmied, H., S. Eisner, D. Franz, M. Wattenbach, F. T. Portmann, M. Flörke, and P. Döll, 2014: Sensitivity of simulated global-scale freshwater fluxes and storages to input data, hydrological model structure, human water use and calibration. Hydrol. Earth Syst. Sci., 18, 3511–3538, https://doi.org/10.5194/hess-18-3511-2014.
Müller Schmied, H., and Coauthors, 2016: Variations of global and continental water balance components as impacted by climate forcing uncertainty and human water use. Hydrol. Earth Syst. Sci., 20, 2877–2898, https://doi.org/10.5194/hess-20-2877-2016.
Muñoz-Sabater, J., and Coauthors, 2021: ERA5-Land: A state-of-the-art global reanalysis dataset for land applications. Earth Syst. Sci. Data, 13, 4349–4383, https://doi.org/10.5194/essd-13-4349-2021.
Neubauer, D., and Coauthors, 2019: HAMMOZ-Consortium MPI-ESM1.2-HAM model output prepared for CMIP6 CMIP historical, version 20201101. Earth System Grid Federation, accessed 17 June 2022, https://doi.org/10.22033/esgf/cmip6.1621.
Pilgrim, D. H., T. G. Chapman, and D. G. Doran, 1988: Problems of rainfall-runoff modelling in arid and semiarid regions. Hydrol. Sci. J., 33, 379–400, https://doi.org/10.1080/02626668809491261.
Raftery, A. E., T. Gneiting, F. Balabdaoui, and M. Polakowski, 2005: Using Bayesian model averaging to calibrate forecast ensembles. Mon. Wea. Rev., 133, 1155–1174, https://doi.org/10.1175/MWR2906.1.
Rodell, M., and Coauthors, 2004: The Global Land Data Assimilation System. Bull. Amer. Meteor. Soc., 85, 381–394, https://doi.org/10.1175/BAMS-85-3-381.
Seland, Ø., and Coauthors, 2020: Overview of the Norwegian Earth System Model (NorESM2) and key climate response of CMIP6 DECK, historical, and scenario simulations. Geosci. Model Dev., 13, 6165–6200, https://doi.org/10.5194/gmd-13-6165-2020.
Sheffield, J., G. Goteti, and E. F. Wood, 2006: Development of a 50-year high-resolution global dataset of meteorological forcings for land surface modeling. J. Climate, 19, 3088–3111, https://doi.org/10.1175/JCLI3790.1.
Sidorenko, D., and Coauthors, 2015: Towards multi-resolution global climate modeling with ECHAM6–FESOM. Part I: Model formulation and mean climate. Climate Dyn., 44, 757–780, https://doi.org/10.1007/s00382-014-2290-6.
Siebert, S., M. Kummu, M. Porkka, P. Döll, N. Ramankutty, and B. R. Scanlon, 2015: A global data set of the extent of irrigated land from 1900 to 2005. Hydrol. Earth Syst. Sci., 19, 1521–1545, https://doi.org/10.5194/hess-19-1521-2015.
Sippel, S., N. Meinshausen, E. Székely, E. Fischer, A. G. Pendergrass, F. Lehner, and R. Knutti, 2021: Robust detection of forced warming in the presence of potentially large climate variability. Sci. Adv., 7, eabh4429, https://doi.org/10.1126/sciadv.abh4429.
Sivapalan, M., 2003: Prediction in ungauged basins: A grand challenge for theoretical hydrology. Hydrol. Processes, 17, 3163–3170, https://doi.org/10.1002/hyp.5155.
Swapna, P., R. Krishnan, N. Sandeep, A. G. Prajeesh, D. C. Ayantika, S. Manmeet, and R. Vellore, 2018: Long‐term climate simulations using the IITM Earth System Model (IITM‐ESMv2) with focus on the south Asian monsoon. J. Adv. Model. Earth Syst., 10, 1127–1149, https://doi.org/10.1029/2017MS001262.
Swart, N. C., and Coauthors, 2019: The Canadian Earth System Model version 5 (CanESM5.0.3). Geosci. Model Dev., 12, 4823–4873, https://doi.org/10.5194/gmd-12-4823-2019.
Swenson, S. C., D. M. Lawrence, and H. Lee, 2012: Improved simulation of the terrestrial hydrological cycle in permafrost regions by the Community Land Model. J. Adv. Model. Earth Syst., 4, M08002, https://doi.org/10.1029/2012MS000165.
Tabari, H., P. Hosseinzadehtalaei, W. Thiery, and P. Willems, 2021: Amplified drought and flood risk under future socioeconomic and climatic change. Earth’s Future, 9, e2021EF002295, https://doi.org/10.1029/2021EF002295.
Takata, K., S. Emori, and T. Watanabe, 2003: Development of the minimal advanced treatments of surface interaction and runoff. Global Planet. Change, 38, 209–222, https://doi.org/10.1016/S0921-8181(03)00030-4.
Tang, Q., T. Oki, S. Kanae, and H. Hu, 2007: The influence of precipitation variability and partial irrigation within grid cells on a hydrological simulation. J. Hydrometeor., 8, 499–512, https://doi.org/10.1175/JHM589.1.
Tatebe, H., and Coauthors, 2019: Description and basic evaluation of simulated mean state, internal variability, and climate sensitivity in MIROC6. Geosci. Model Dev., 12, 2727–2765, https://doi.org/10.5194/gmd-12-2727-2019.
Venkatappa, M., N. Sasaki, P. Han, and I. Abe, 2021: Impacts of droughts and floods on croplands and crop production in Southeast Asia – An application of Google Earth Engine. Sci. Total Environ., 795, 148829, https://doi.org/10.1016/j.scitotenv.2021.148829.
Wada, Y., D. Wisser, and M. F. P. Bierkens, 2014: Global modeling of withdrawal, allocation and consumptive use of surface water and groundwater resources. Earth Syst. Dyn., 5, 15–40, https://doi.org/10.5194/esd-5-15-2014.
Warszawski, L., K. Frieler, V. Huber, F. Piontek, O. Serdeczny, and J. Schewe, 2014: The Inter-Sectoral Impact Model Intercomparison Project (ISI–MIP): Project framework. Proc. Natl. Acad. Sci. USA, 111, 3228–3232, https://doi.org/10.1073/pnas.1312330110.
Weedon, G. P., G. Balsamo, N. Bellouin, S. Gomes, M. J. Best, and P. Viterbo, 2014: The WFDEI meteorological forcing data set: WATCH forcing data methodology applied to ERA-interim reanalysis data. Water Resour. Res., 50, 7505–7514, https://doi.org/10.1002/2014WR015638.
Yamazaki, D., G. A. M. de Almeida, and P. D. Bates, 2013: Improving computational efficiency in global river models by implementing the local inertial flow equation and a vector-based river network map. Water Resour. Res., 49, 7221–7235, https://doi.org/10.1002/wrcr.20552.
Yamazaki, D., and Coauthors, 2017: A high‐accuracy map of global terrain elevations. Geophys. Res. Lett., 44, 5844–5853, https://doi.org/10.1002/2017GL072874.
Yamazaki, D., D. Ikeshima, J. Sosa, P. D. Bates, G. H. Allen, and T. M. Pavelsky, 2019: MERIT Hydro: A high‐resolution global hydrography map based on latest topography dataset. Water Resour. Res., 55, 5053–5073, https://doi.org/10.1029/2019WR024873.
Yang, H., and Coauthors, 2015: Multicriteria evaluation of discharge simulation in dynamic global vegetation models. J. Geophys. Res. Atmos., 120, 7488–7505, https://doi.org/10.1002/2015JD023129.
Yang, T., F. Sun, P. Gentine, W. Liu, H. Wang, J. Yin, M. Du, and C. Liu, 2019: Evaluation and machine learning improvement of global hydrological model-based flood simulations. Environ. Res. Lett., 14, 114027, https://doi.org/10.1088/1748-9326/ab4d5e.
Yang, Y., M. L. Roderick, D. Yang, Z. Wang, F. Ruan, T. R. McVicar, S. Zhang, and H. E. Beck, 2021: Streamflow stationarity in a changing world. Environ. Res. Lett., 16, 064096, https://doi.org/10.1088/1748-9326/ac08c1.
Yin, J., P. Gentine, S. Zhou, S. C. Sullivan, R. Wang, Y. Zhang, and S. Guo, 2018: Large increase in global storm runoff extremes driven by climate and anthropogenic changes. Nat. Commun., 9, 4389, https://doi.org/10.1038/s41467-018-06765-2.
Yukimoto, S., and Coauthors, 2019: The Meteorological Research Institute Earth System Model version 2.0, MRI-ESM2.0: Description and basic evaluation of the physical component. J. Meteor. Soc. Japan, 97, 931–965, https://doi.org/10.2151/jmsj.2019-051.
Zaherpour, J., and Coauthors, 2018: Worldwide evaluation of mean and extreme runoff from six global-scale hydrological models that account for human impacts. Environ. Res. Lett., 13, 065015, https://doi.org/10.1088/1748-9326/aac547.
Zaitchik, B. F., M. Rodell, and F. Olivera, 2010: Evaluation of the global land data assimilation system using global river discharge data and a source-to-sink routing scheme. Water Resour. Res., 46, 2009WR007811, https://doi.org/10.1029/2009WR007811.
Zhai, R., F. Tao, U. Lall, B. Fu, J. Elliott, and J. Jägermeyr, 2020: Larger drought and flood hazards and adverse impacts on population and economic productivity under 2.0 than 1.5°C warming. Earth’s Future, 8, e2019EF001398, https://doi.org/10.1029/2019EF001398.
Zhang, S., and Coauthors, 2022: Reconciling disagreement on global river flood changes in a warming climate. Nat. Climate Change, 12, 1160–1167, https://doi.org/10.1038/s41558-022-01539-7.
Zhang, Y. Q., N. Viney, A. Frost, A. Oke, M. Brooks, Y. Chen, and N. Campbell, 2013: Collation of Australian modeller’s streamflow dataset for 780 unregulated Australian catchments. CSIRO Tech. Rep., 115 pp., https://doi.org/10.4225/08/58b5baad4fcc2.
Zhao, F., and Coauthors, 2017: The critical role of the routing scheme in simulating peak river discharge in global hydrological models. Environ. Res. Lett., 12, 075003, https://doi.org/10.1088/1748-9326/aa7250.
Ziehn, T., and Coauthors, 2020: The Australian Earth System Model: ACCESS-ESM1.5. J. South. Hemisphere. Earth Syst. Sci., 70, 193–214, https://doi.org/10.1071/ES19035.
