Compared to traditional drought events, flash droughts evolve rapidly during short-term extreme atmospheric conditions, with a lasting period of one pentad to several weeks. There are two main categories of flash droughts: the heat wave flash drought (HWFD), which is mainly caused by persistent high temperatures (heat waves), and the precipitation deficit flash drought (PDFD), which is mainly triggered by precipitation deficits. The authors’ previous research focused on the characteristics and causes of flash drought based on meteorological observations and Variable Infiltration Capacity (VIC) model simulations in a humid subtropical basin (Gan River basin, China). In this study, the authors evaluated the downscaled phase 5 of the Coupled Model Intercomparison Project (CMIP5) models’ simulations, coupled with the VIC model (CMIP5–VIC) in reproducing flash droughts in a humid subtropical basin in China. Most downscaled CMIP5–VIC simulations can reproduce the spatial patterns of flash droughts with respect to the benchmarks. The coupled models fail to readily replicate interannual variation (interannual pentad change), but most models can reflect the interannual variability (temporal standard deviation) and long-term average pentads of flash droughts. It is difficult to simultaneously depict both the spatial and temporal features of flash droughts within only one coupled model. The climatological patterns of the best multimodel ensemble mean are close to those of the all-model ensemble mean, but the best multimodel ensemble mean has a minimal bias range and relatively low computational burden.
The Fifth Assessment Report (AR5) of the Intergovernmental Panel on Climate Change (IPCC) estimated a global surface temperature warming of 0.85°C from 1880 to 2012 (IPCC 2013). Climate change has elevated the temperatures of land surfaces, which has caused droughts to occur with more frequent and greater intensity across the globe (Trenberth et al. 2014). Thus, comprehensive assessment of various types of drought events is a highly worthwhile research object.
Droughts are usually caused by abnormally dry weather, which persists long enough to trigger a sustained and regionally extensive occurrence of below-average water availability. Generally, droughts are typically slowly evolving phenomena. Conversely, flash droughts manifest rapidly when extreme atmospheric conditions (e.g., abnormally high geopotential height and water vapor flux) occur over a short term (days or weeks), especially during the crop-growing season (Otkin et al. 2013). Flash droughts can be split into two main categories based on their respective formation mechanism. The first type is the heat wave flash drought (HWFD) (Mo and Lettenmaier 2015), which is formed by high temperatures that cause evapotranspiration to increase and lead to a rapid decline in soil moisture. Although precipitation is important, the main driver of an HWFD is persistent high temperature (heat waves), not precipitation. The second type is the precipitation deficit flash drought (PDFD) (Mo and Lettenmaier 2016), where precipitation deficits cause evapotranspiration to decline and in turn cause temperature to increase. The PDFD is a precipitation-deficit-driven event despite high temperatures. In the central United States, a severe HWFD occurred in May and early June of 2012 (Mo and Lettenmaier 2015; Otkin et al. 2016), while the summer drought in 1980 over this area formed a severe PDFD (Mo and Lettenmaier 2016). These flash droughts inflicted widespread agricultural crop failure, a drastic decrease in livestock population, and huge economic losses (Mo and Lettenmaier 2015; Otkin et al. 2016; Wang et al. 2016).
Generally, traditional drought indices do not effectively represent flash droughts. Consider the two most commonly used drought indices as examples: both the standardized precipitation evapotranspiration index (SPEI) (Vicente-Serrano et al. 2010a,b) and Palmer drought severity index (PDSI) (Palmer 1965) can effectively capture long-term (months or longer) drought characteristics based on observed data of monthly precipitation and temperature but may be less effective in monitoring flash droughts owing to the untimely response to changes in prevailing weather conditions. Both soil moisture and evapotranspiration are needed to quantitatively characterize the variability of a flash drought as it is a type of agricultural drought by nature (Mo and Lettenmaier 2015). Unfortunately, there are currently very few high-resolution observations of soil moisture and evapotranspiration. Despite the errors and the uncertainty inherent to remotely sensed observations and land surface model data, they are also the most effective tools of obtaining the short-term soil moisture and evapotranspiration conditions necessary to identify flash droughts. For example, Otkin et al. (2013) applied remote sensing data to calculate the evaporative stress index for detecting the onset of rapid drought across the United States and then used this data to explore the evolution of the 2012 U.S. flash drought (Otkin et al. 2016). Recently, some researchers have adopted model-reconstructed products to explore the HWFDs and PDFDs during crop growing seasons (April–September) from 1916 to 2013 (Mo and Lettenmaier 2015, 2016). Yuan et al. (2015) evaluated the capability of microwave remote sensing data in monitoring severe short-term droughts in China, and Wang et al. (2016) investigated HWFD characteristics in China from 1979 to 2010.
Recently, phase 5 of the Coupled Model Intercomparison Project (CMIP5) (Taylor et al. 2012) has synthesized many comprehensive global climate models capable of reproducing aspects of the mean climate and climate extremes. Many researchers have used CMIP5 to explore climate extremes. Sillmann et al. (2013) provided the overview of CMIP5 model performance in reproducing climate extremes based on the Expert Team on Climate Change Detection and Indices (ETCCDI) and found that CMIP5 models can generally simulate the spatial patterns of climate extremes with respect to the gridded observational indices dataset over the global land surface. Jiang et al. (2015) assessed the precipitation extremes in 31 CMIP5 models in China to find that they have wet biases in the western and northern regions of China but dry biases in southeastern regions; the Chinese eastern regions were more effectively simulated than the west in terms of spatial and temporal characteristics.
As for drought assessment in CMIP5 models, Nasrollahi et al. (2015) investigated meteorological droughts using a three-month standardized precipitation index (SPI) over global land based on CMIP5 outputs to find that most models 1) overestimate extreme droughts over many regions and 2) are incompatible with observed regional drying and wetting trends. Venkataraman et al. (2016) reported the impact of climate change projections from CMIP5 outputs on twenty-first-century drought characteristics in Texas (southern United States) based on SPI and SPEI methods; they concluded that the model ensemble performed better in simulating historical temperature than precipitation, and less severe droughts were projected for the subhumid eastern regions of Texas. These studies all focused on assessments of long-term droughts using CMIP5 or downscaled CMIP5 outputs, while there has been no assessment of flash droughts using the downscaled CMIP5 outputs to date—particularly in regard to humid subtropical basins.
Our previous work focused on flash drought characteristics and their causes based on meteorological observations and VIC simulations in the Gan River basin (Zhang et al. 2017b). However, some unresolved questions remain, including how well the downscaled CMIP5 outputs coupled with the VIC model (CMIP5–VIC) can simulate flash drought characteristics in a humid subtropical basin, which coupled models yield optimal simulations according to the skill metrics, and how the multimodel ensemble mean and optimal model mean can simultaneously simulate HWFD and PDFD characteristics. This study provides an initial overview of the downscaled CMIP5–VIC models’ capability in representing historical flash droughts in the Gan River basin in terms of spatial and temporal characteristics.
The major goals of this study were to 1) evaluate the downscaled CMIP5–VIC simulations for HWFDs and PDFDs in terms of climatological patterns, 2) assess the coupled models’ abilities in simulating two types of flash droughts in terms of interannual variability, and 3) select the best models and detect their ensemble mean performance compared to the all-multimodel ensemble mean. The findings presented here may provide a scientific basis regarding the feasibility of the downscaled CMIP5–VIC in simulating flash droughts, providing credible future flash drought estimation, and mitigating potential losses due to flash droughts.
2. Materials and methods
a. Study area
The Gan River basin is located mostly within the central and southern parts of Jiangxi Province, China (Fig. 1). The Waizhou hydrological station (outlet of the basin) of Nanchang City controls an area of 80 948 km2 of the basin (approximately the size of South Carolina in the United States), which is the largest subbasin in the Poyang Lake basin. The Gan River basin is dominated by a humid subtropical climate and distinct seasonal variations as a result of the East Asian monsoon, with an annual precipitation of 1600.1 mm and an annual average temperature of 18.2°C (Zhang et al. 2017b). Runoff in the Gan River basin is mainly caused by precipitation and exhibits strong seasonality, typically reaching its peak in June and its lowest point in December.
We used the daily gridded precipitation, maximum temperature, minimum temperature, and surface wind speed derived from 47 meteorological stations (Fig. 1) to drive the calibrated VIC model to obtain evapotranspiration and soil moisture data. According to our previous research (Zhang et al. 2017b), the calibrated and validated VIC (0.125° spatial resolution) is capable of reasonably depicting the characteristics of hydrological processes in the Gan River basin. The flash drought examined in this study was based on meteorological observations (maximum temperature and precipitation) and VIC simulations (evapotranspiration and soil moisture). We considered this flash drought simulation as a benchmark (baseline simulation).
To assess the ability of the downscaled CMIP5 outputs to simulate flash droughts, we retrieved the historical daily precipitation as well as maximum and minimum temperature from 21 downscaled CMIP5 outputs from the NASA Earth Exchange Global Daily Downscaled Projections (NEX-GDDP) dataset (Thrasher et al. 2012). This dataset has been handled by the bias-correction spatial disaggregation method to downscale the projections from the 21 original CMIP5 models. Each of these downscaled climate models has a horizontal resolution of 0.25° × 0.25°. The detailed dataset information and documentation are available online (https://cds.nccs.nasa.gov/nex-gddp/). We used NEX-GDDP data as input data to run the calibrated VIC model to obtain evapotranspiration and soil moisture for each combination (the CMIP5–VIC) to simulate flash drought characteristics. We refer to this flash drought simulation as the downscaled CMIP5–VIC simulation. The focus of this study is on comparisons of the downscaled CMIP5–VIC simulations and the benchmark (baseline simulation), namely, an assessment of flash droughts in the downscaled CMIP5–VIC simulations. All meteorological datasets cover the base period from 1961 to 2005. The 21 available climate models and other basic information are listed in Table 1. We interpolated all climate model data to a common resolution of 0.125° × 0.125° using the nearest interpolation method, which can avoid introducing additional error from the interpolation, for convenience of establishing our VIC model simulations
c. Flash drought indices
As mentioned above (Mo and Lettenmaier 2015, 2016; Wang et al. 2016; Zhang et al. 2017b), the HWFD phenomenon occurs in the short term and is mainly driven by heat waves, while PDFDs are driven by precipitation deficit. Although the evaporative capacity is very strong, evapotranspiration anomalies still decline upon PDFD occurrence. This substantial decrease in evapotranspiration and soil moisture may in turn result in increasing temperature during the crop growing season. We used a pentad (five-day means) as the representative time span to investigate flash droughts based on each related variable (e.g., maximum temperature, evapotranspiration, soil moisture, and precipitation) and specifically focused on pentad mean climatology during the crop growing seasons (March–October; 49 pentads per year) from 1961 to 2005. For each grid and pentad, an HWFD event is defined as the conditions under which the maximum temperature anomaly is greater than one standard deviation, the evapotranspiration anomaly is in positive phase, and the soil moisture percentile is lower than 40%; a PDFD event is defined by maximum temperature anomaly greater than one standard deviation, evapotranspiration anomaly in negative phase, and precipitation percentile below 40% (Zhang et al. 2017b). The percentiles were determined from base period (1961–2005) data.
The frequency of occurrence (FOC) is defined as the percentage of pentads under both types of flash droughts for each grid:
where is the total number of pentads (2205 pentads from 1961 to 2005) and N is the number of pentads under flash droughts over the entire record.
d. Model evaluation metrics
The Taylor diagram (Taylor 2001) can concisely summarize the degree of correspondence between simulations and benchmarks in their spatial patterns. This diagram includes three metrics that provide a statistical summary of comparisons between two patterns: the spatial correlation coefficient R, the ratio of spatial standard deviation (STD), and the normalized root-mean-square error (RMSE) between the simulated and benchmark field. The R and RMSE are measurements representing the degree of phase and amplitude of the two fields. The best model simulation results when the R and STD are equal to 1 and RMSE is close to 0.
Because CMIP5 models generally do not yield accurate interannual variations of climate extremes, we used the interannual standard deviation as an interannual variability skill (IVS) to explore how accurately a model simulates the corresponding benchmark. IVS is calculated as follows (Chen et al. 2011; Jiang et al. 2015):
where and are the interannual standard deviation of simulations and benchmarks, respectively. We calculated the basin-averaged IVS values based on each grid. Theoretically, an IVS of zero is a simulation that exactly corresponds to the benchmark.
To assess the overall model’s ability to simultaneously simulate HWFD and PDFD events, we identified the rank of each type of flash drought based on the relevant Taylor diagram and IVS. An overall ranking of each model should consider the two types of flash droughts at spatial and temporal scales. The comprehensive rating index (CRI) is known to effectively rank models (Jiang et al. 2015). The formula is as follows:
where is the number of indices and denotes the number of models (21). The closer CRI is to 1, the better the model performs.
a. Assessment of spatial climatology
We first quantitatively evaluated the performance of each downscaled CMIP5–VIC by reproducing the climatology spatial patterns of flash droughts in the Gan River basin. Each model’s capability in simulating HWFDs in the basin is shown in Fig. 2. Both simulations and benchmarks indicated that the FOCs of HWFDs were more likely to occur in the northern region, 26.8°N, of the basin. This phenomenon may be a result of the topographical characteristics of the terrain. The southern region of 26.8°N is mainly dominated by mountains and hills, while areas north of 26.8°N consist mostly of alluvial plains. Overall, all models effectively simulated HWFD patterns with respect to the benchmark. ACCESS1.0, CanESM2, MIROC5, and MRI-CGCM3 (the detailed information refers to Table 1) underestimated the FOCs of HWFDs in southern regions of the basin, while CNRM-CM5, GFDL-ESM2G, INM-CM4.0, MPI-ESM-LR, and MPI-ESM-MR overestimated the FOCs of HWFDs in southern regions of the basin. In northern regions of the basin, BNU-ESM, IPSL-CM5A-MR, and MIROC5 underestimated the FOCs of HWFDs, while almost no models overestimated HWFDs in this area. Among these models, only MIROC5 underestimated HWFDs in the entire basin. As shown in Fig. 3, most models’ simulations performed well in accordance with the benchmark; their correlation coefficients were around 0.9. Overall, as shown in Figs. 2v,w and Fig. 3v, the multimodel ensemble mean (MME) was in close agreement with the benchmark in terms of spatial distribution, albeit with an underestimation of HWFD FOCs in northern parts of the basin.
As shown in Fig. 4, in general, the PDFDs’ FOCs in all coupled models had similar spatial patterns with respect to the benchmark, although the intensity differs (especially in the southern parts of the basin). Both model simulations and the benchmark showed that the FOCs of PDFDs have a spatial distribution opposite to those found in HWFDs (i.e., the central and southern regions of the basin are more vulnerable to PDFDs than the north). We also observed a north–south pattern aligned along 26.8°N through the basin. In the northern regions of the basin, most of the models overestimated PDFD FOCs with the exception of ACCESS1.0, INM-CM4.0, and MPI-ESM-LR, which showed some underestimation. The MPI-ESM-LR model particularly severely underestimated FOCs. In the southern regions of the basin, there were some overestimations of PDFDs for MIROC5, MIROC-ESM, and especially MIROC-ESM-CHEM. It is worth noting that these three models were developed at the same institution. The CNRM-CM5, INM-CM4.0, and MPI-ESM-LR models showed an obvious underestimation of PDFD FOCs in the southern regions of the basin. Among these models, INM-CM4.0 and MPI-ESM-LR underestimated PDFDs and MIROC-ESM-CHEM considerably overestimated PDFDs in the entire basin. Figure 5 presents a comparison of the model simulations and benchmark. The correlation coefficients between the model simulations and benchmark values were around 0.8. Overall, as shown in Figs. 4v,w and Fig. 5v, The MME results of PDFDs were also in close agreement with the benchmark in terms of spatial distribution, although the northern regions were overestimated.
b. Evaluation of spatial variability
We applied the Taylor diagram method to evaluate model performance in simulating the spatial patterns of flash droughts. As shown in Fig. 6, all models have spatial R values ranging from 0.70 to 0.95 for both flash drought types. The HWFDs’ R values in most models were around 0.90, while those of PDFDs were around 0.80. The majority of the models showed STDs for PDFDs ranging from 0.75 to 1.00, but only about half of the models located in this range for HWFDs. The STDs of several models were larger than 1.00 for PDFDs, but none of the models’ STDs were beyond 1.00 for HWFDs. These STD results for both types of flash droughts indicate that the PDFD has a larger spatial variation than the HWFD against the benchmarks. The RMSEs of all models for HWFDs were below 0.6, except MPI-ESM-MR. However, the RMSEs of nearly every model for PDFDs were distributed at 0.6, indicating that the amplitude of biases for PDFDs was larger than that of HWFDs. Overall, the downscaled CMIP5–VIC displayed favorable performance in simulating HWFD and PDFD patterns (spatial FOCs), and in general, the HWFD simulations were better than PDFD simulations in the Gan River basin. The intermodel spread for HWFDs is slightly smaller than that of PDFDs, as evidenced by the loosely scattered distribution on the Taylor diagram.
To more precisely distinguish performance among the 21 downscaled CMIP5–VICs, we first calculated three aspects (STD, RMSE, and R) of the Taylor diagram and then estimated the CRI to search a comprehensive ranking for each model. Figure 7 shows each model’s ranking of the two types of flash droughts in terms of STD, RMSE, and R based on CRI. The model simulation performance was determined in descending order. A small number (red) denotes a model with good simulation capability. The rankings of the three indices (STD, RMSE, and R) of HWFDs are similar, but different in PDFDs. For example, the ACCESS1.0 model’s STD, RMSE, and R rank values for HWFDs were 4, 3, and 4, respectively. The values between these rank values were very close, but the STD, RMSE, and R ranking values of PDFDs for the ACCESS1.0 model were quite different. The model that most successfully simulated both HWFDs and PDFDs was BCC_CSM1.1 in terms of spatial climatological patterns. This model was derived from the China Meteorological Administration and thus likely had absorbed more meteorological observations as input data compared to other models, especially regarding precipitation.
c. Evaluation of interannual variability
Figure 8 shows an average of number of pentads in the downscaled CMIP5–VIC simulations and benchmarks of HWFDs and PDFDs per year in the basin. We found that the amplitudes of benchmark interannual variations were much larger than those of coupled model median values. For example, the medians of 1963 data simulated in downscaled CMIP5–VICs did not exceed one and two pentads for HWFDs and PDFDs in the basin, but the benchmarks were 3 times higher than these results. However, a range of 25th–75th percentiles in downscaled CMIP5–VIC simulations encompassed the benchmarks at most time steps from 1961 to 2005. Because the medians of the downscaled CMIP5–VICs did not accurately reproduce interannual variations with respect to the benchmarks, we applied the interannual variability (interannual standard deviation) to assess the performance of the downscaled CMIP5–VIC simulations in flash droughts.
The model performances of interannual variability simulations of flash droughts are shown in Fig. 9. The best model for simulating HWFDs and PDFDs is MRI-CGCM3 (IVS values are close to zero). The MRI-CGCM3 model is also excellent (third best) for simulating the spatial capabilities of both drought types. It is worth noting that nearly half of the models showed a large spread in their ability to simultaneously simulate HWFDs and PDFDs in terms of interannual variability. MPI-ESM-LR and NorESM1-M are excellent examples. For MPI-ESM-LR, the PDFD IVS value is very close to zero (0.01), while the HWFD IVS value is greater than one (1.12), indicating that this model is better able to simulate PDFDs than HWFDs. In NorESM1-M, the PDFD IVS value is close to one (0.83) while the HWFD IVS value is close to zero (0.12), demonstrating that this model better simulates HWFDs than PDFDs.
d. Overall model ranking
We conducted a comprehensive assessment of each model to resolve discrepancies in the ranking based on Taylor and IVS results. As shown in Fig. 10, models in the lower-left quadrant showed favorable performance for both HWFDs and PDFDs in terms of spatial and temporal characteristics. The five best models are MRI-CGCM3, ACCESS1.0, BCC-CSM1.1, CanESM2, and CESM1(BGC). According to the overall model rankings, the optimal model is MRI-CGCM3—this model consistently accurately depicted the spatial and temporal characteristics of both drought types. This may be due to the fact that MRI-CGCM3 performs well in simulating precipitation and temperature in eastern China (Jiang et al. 2015; Sillmann et al. 2013). We found obvious discrepancy in some models between Taylor and IVS for simulating both HWFDs and PDFDs. For example, CNRM-CM5 has a very high ranking in interannual variability (second best in IVS rank) but a very low ranking in spatial patterns (worst in Taylor rank). There were obvious discrepancies in the optimal models as well. One of the best models is CanESM2, which ranks second in simulated spatial patterns (Taylor) but seventh in temporal characteristics (IVS) for both types of flash droughts. In summary, the robustness of most models is relatively weak in terms of the ability to simultaneously simulate both the spatial and temporal features of flash droughts.
e. Performance of optimal models
Five optimal models were selected to examine how well the best multimodel ensemble mean (BMME) reproduces flash droughts with respect to benchmarks. We evaluated the relative biases of the BMME and all-model ensemble mean (AMME) by comparison against benchmarks of temporal characteristics. The biases of the BMME and AMME are presented as a box-and-whisker plot in Fig. 11. Both BMME and AMME almost underestimate HWFDs, where both the median and mean of relative biases for BMME and AMME are lower than the benchmarks by about 20%. The length of the AMME box (interquartile range) is longer than that of BMME, indicating larger uncertainty in the AMME simulations than that in BMME. For PDFDs, the median and mean of relative biases in BMME are slightly closer to the benchmarks than the relative biases of AMME. The BMME whisker is shorter than the AMME whisker despite their nearly identical box lengths. To this effect, BMME can reduce the range of biases in simulating the temporal characteristics of HWFDs and PDFDs, but the median and mean of biases are similar regardless of the type of flash drought.
Compared to AMME, the most obvious advantage of BMME is the enhanced robustness of flash drought simulations via elevated temporal characteristics. However, the mean and median values of BMME simulations did not significantly improve. We further calculated the relative biases of the multiyear average of each model. According to Fig. 12, most models underestimate the pentads of HWFDs while most models overestimate the pentads of PDFDs. The relative biases of the BMME and the AMME are close, and most models simulated the climatological pentads of the two types of flash droughts within ±25% relative bias, which indicates that the downscaled CMIP5–VIC can reasonably simulate long-term average annual pentads of both types of flash droughts.
We assessed the differences in spatial patterns between simulations in the two ensembles to further investigate how well BMME and AMME reproduce flash droughts with respect to benchmarks. As shown in Fig. 13, BMME and AMME underestimated the HWFD and overestimated the PDFD. In the northeastern regions of the basin, the underestimation of HWFDs reached up to ~2%. We found a similar distribution of spatial biases for BMME and AMME in HWFDs, although there were some differences in the southern region. The spatial bias distributions of PDFDs were also similar to those of HWFDs, but there was a tendency to overestimate the PDFD FOCs. For both types of flash droughts, the correlation coefficients of BMME and AMME were higher than 0.95, indicating that these five optimal modes can replace all models in terms of simulating temporal and spatial climatological characteristics of flash droughts. We noticed that the region with the largest PDFD biases is located in western parts of the basin, about 26.2°–26.7°N, 114.0°–114.4°E (Jinggang Mountain regions). Larger biases also exist in the HWFD as shown in Figs. 2v,w and Figs. 4v,w. In the Jinggang Mountain regions, the simulated values significantly underestimated HWFDs while there were marked overestimations of PDFDs. In summary, the spatial simulations of these two ensemble models are similar between different flash droughts, although BMME is slightly better than AMME.
The Gan River basin (the largest subbasin of Poyang Lake basin) is a typical humid subtropical basin that is sensitive to hydrological and climatic changes (Liu et al. 2017; Zhang et al. 2014, 2017a), especially at extremes (Q. Zhang et al. 2016). CMIP5 models are powerful tools in the climate change research field, as they can be effectively utilized to assess the characteristics and potential challenges associated with droughts. In this study, we focused on the downscaled CMIP5–VIC performance in simulating flash droughts in a humid subtropical basin in China.
Most downscaled CMIP5–VICs were able to simulate the basic spatial patterns of the two kinds of flash droughts with respect to benchmarks. This is probably due to the relatively strong capability of the CMIP5 models to simulate extreme precipitation and temperature in the eastern regions of China (Jiang et al. 2015; Wen et al. 2016). The original CMIP5 models do not appear to be suitable for regional hydrometeorological studies because their coarse horizontal resolution makes them incompatible with hydrological models. It is generally possible to enhance the performance of hydrological models by downscaling the CMIP5 models’ meteorological data (Y. Zhang et al. 2016). We accordingly used quality-checked downscaled data (NEX-GDPP) (Thrasher et al. 2012) as VIC model inputs to explore flash droughts at the basin scale. This dataset not only improved the CMIP5 models’ spatial resolution, but also enhanced the accuracy of spatial data, facilitating our assessment of regional climate change. The downscaled CMIP5–VIC was beneficial in simulating the spatial characteristics of flash droughts in this study.
The CMIP5 model has a relatively large bias for internal climate variability at shorter time scales and smaller spatial scales. It generally underestimates natural climate variability, especially in terms of variations in precipitation (Deser et al. 2012; Goddard et al. 2013; Tebaldi et al. 2011). The biases in internal climate variability in CMIP5 models amplify the uncertainty of CMIP5–VIC simulations during downscaling and land surface model processes. Ultimately, the CMIP5–VIC simulations did not reproduce sufficient interannual pentad variations of flash droughts with respect to the benchmarks.
The skills/ranks of the coupled models for interannual variability (temporal standard deviation) are affected by not only the original CMIP5 models but also the statistical downscaling methods applied (Woznicki et al. 2016). For example, Esfahanian et al. (2016) applied the delta downscaling method to bridge the CMIP5 model and SWAT (soil and water assessment tool) hydrological model to assess the response of fish vulnerability to drought, and they found that the delta method does not account for natural climate variability (especially for precipitation), as it is stationary in its transfer function. In an effort to optimize their regional climate impact model, Themeßl et al. (2011) merged linear and nonlinear empirical statistical downscaling methods with bias correction approaches to reduce model errors; these downscaling methods drastically reduced errors for daily precipitation but did not resolve some errors related to decadal climate variability.
The NEX-GDDP data (the downscaled CMIP5 outputs in this study) have been processed by the bias-correction spatial disaggregation method (Maurer and Hidalgo 2008) for spatial downscaling. This downscaling method is successful in reproducing the main features of the observed hydrometeorology from retrospective climate simulations but may fail to rectify more subtle differences between climate model and observed climate (Wood et al. 2004). Simultaneous error in several climate variables can cause model biases in flash drought simulations. Although the medians in our downscaled CMIP5–VIC do not readily reflect interannual variations of flash droughts, a range of 25th–75th percentiles encompasses more benchmarks and reasonably simulates long-term average pentads of flash droughts. It is worth noting that the median downscaled CMIP5–VIC simulation exhibited a slight HWFD increase during the entire period, though no obvious trend existed in the benchmark. The CMIP5 models were able to simulate the effects of warming temperature relatively well (Chen and Frauenfeld 2014). These increasing temperature trends were also added to HWFD simulations because of the temperature-driven nature of this type of drought.
It is difficult to accurately depict two kinds of extreme climatological events using the same model (Li et al. 2013). Some models that work well for HWFD simulations will work poorly for PDFD simulations based on the overall performance metrics. The results presented here may support the assumption that spatial simulations of both BMME and AMME are basically similar for any type of flash drought. Further, the HWFD was underestimated by the two ensemble mean models while the PDFD was overestimated. It is possible that these models underestimated the magnitude of temperature variations in the short term (days or pentads), especially for high temperatures slightly less than the observed temperatures, leading to an eventual underestimation of HWFDs. Many models have a “drizzle phenomenon” (Liu et al. 2014), which results in overestimation of PDFD simulations according to the definition of PDFD.
In our work, flash droughts comprise several concurrent extreme hydroclimatological events and can be further explored with precipitation deficit–temperature coupling methods (Zhang et al. 2015a) or joint probability approaches (Zhang et al. 2015b) based on short time scales (generally one week to one month). Although the climatology patterns of BMME are close to those of AMME, the optimal model can shorten the bias range and reduce the amount of model computation. Therefore, we recommend the use of the BMME (five best model ensemble mean) to replace the AMME (all-model ensemble mean) to investigate flash drought characteristics in humid subtropical basins such as the Gan River basin.
This study may provide useful insight into the role that the downscaled CMIP5–VIC plays in simulating flash droughts in humid subtropical basins in China. Flash drought assessments were completed based on the downscaled CMIP5–VIC simulations in terms of spatial patterns and interannual variability for the period 1961–2005. The findings presented here may also provide a feasible and credible scientific basis for the prediction of future flash droughts, which we plan to further explore in our next study. Our most notable conclusions can be summarized as follows:
The downscaled CMIP5–VIC is capable of simulating the spatial FOC patterns of flash droughts with respect to the benchmarks. The MME FOC results of HWFDs and PDFDs were also in agreement with the benchmarks, albeit with a slight underestimation of HWFDs and overestimation of PDFDs in southern and northern parts of the basin. The effectiveness of the downscaled CMIP5–VIC in simulating flash drought spatial characteristics was demonstrated by Taylor diagrams. According to the loosely scattered distribution on Taylor diagram, the intermodel spread for HWFDs was slightly smaller than that of PDFDs. The rankings of the three indices (STD, RMSE, and R) for HWFDs were similar, but very different for PDFDs.
Almost half of the downscaled CMIP5–VICs displayed a high level of uncertainty in their ability to simultaneously reproduce interannual variability in HWFD and PDFD simulations. The downscaled CMIP5–VIC does not readily replicate interannual variations compared to the benchmarks. However, the benchmarks at most time steps occurred between the 25th and 75th percentiles in the downscaled CMIP5–VIC simulations, which can reasonably simulate long-term average pentads of HWFD and PDFD events.
The five optimal models were MRI-CGCM3, ACCESS1.0, BCC_CSM1.1, Can-ESM2, and CESM1(BGC) according to spatial and temporal characteristics selection standards. Although the climatology patterns of BMME are close to those of AMME, the BMME comes with shorter bias range and less burdensome model computation. Therefore, we recommend the use of the five best models as a replacement for the all-model ensemble to investigate the flash drought characteristics in the Gan River basin, as well as other basins that share similar climatic characteristics.
This study was jointly supported by the National Key R&D Program of China (2017YFA0603804), National Natural Science Foundation (41771069), Jiangsu Natural Science Funds for Distinguished Young Scholar (BK20140047), the Priority Academic Program Development of Jiangsu Higher Education Institutions (PAPD), and the Research and Innovation Project for College Graduates of Jiangsu Province (1344051501007). Climate scenarios used were from the NEX-GDDP dataset prepared by the Climate Analytics Group and NASA Ames Research Center using the NASA Earth Exchange and distributed by the NASA Center for Climate Simulation (NCCS).