Increasing importance of temperature as a contributor to the spatial extent of streamflow drought

The daily streamflow time series were downloaded for the period 1981–2018 from the USGS website (https://waterdata.usgs.gov/nwis) using the R-package dataRetrieval (De Cicco et al 2018). Areal precipitation (mm) and mean daily temperature (°C) for the same period were computed using the Daymet dataset which provides gridded, observation-based estimates of daily precipitation and temperature at a 1-km spatial resolution (Thornton et al 2012). Snow-water-equivalents (SWE; mm) and soil moisture values (mm) for the period 1981–2014 were derived from a modeled data set by Newman et al (2015) who used calibrated lumped implementations of the Snow-17 snow accumulation and ablation model and the Sacramento Soil Moisture Accounting model (SacSMA; Burnash et al 1973) to derive a consistent set of hydro-meteorological variables.

Streamflow droughts at individual sites are extracted using a variable threshold-level approach suitable for regions with a seasonal streamflow regime (Van Loon and Laaha 2015, Heudorfer and Stahl 2017) at the 15th flow percentile (figure 1(A)). The use of a variable instead of a fixed threshold leads to the identification of droughts defined as streamflow anomalies rather than low flows. Please note that such anomalies can also be detected in winter when streamflow anomalies may not have direct societal impacts. The daily time series is smoothed over a moving window of 30 days prior to event extraction to avoid identifying dependent events (Tallaksen and Hisdal 1997, Van Loon and Laaha 2015). The variable threshold is composed by the 15th flow percentile for each day of the year determined within a moving window of ±15 days around the day of interest. We only include events with a minimum duration of 30 days to avoid the consideration of minor droughts. The drought extraction procedure results in a first quartile of 18, a median of 20, and a third quartile of 23 events identified per catchment. These events are spread across seasons as a result of using a variable threshold, which depends on flow seasonality.

Drought spatial extent is determined at a monthly scale as the percentage of catchments affected by drought during a certain month (figure 1(B)). Alternatively, spatial extent could be defined by area-weighting the affected catchments, which does, however, not change the main conclusions of this study. Based on this drought spatial extent time series, we define spatially large drought events as events affecting at least 20% of the catchments in the dataset. However, the drought-affected area does not necessarily need to be contiguous. The duration of these large events is determined as the time elapsing between the start of the event defined as the time of the rise of the extent time series above the threshold of 0.2 and the end of the event when the time series falls below that threshold again. The main date of occurrence is determined as the month with the largest drought extent. We rank the large spatial events according to their bivariate, joint probabilities in terms of event duration and extent determined by their empirical copula (the most severe event is assigned the highest rank; Deheuvels 1979, Genest and Favre 2007).

To evaluate changes in the monthly time series of drought spatial extent, we apply the non-parametric Mann–Kendall test (Mann 1945). In addition, we compare the distributions of drought spatial extent for the two periods 1981–1999 and 2000–2018 for the three value ranges $\lt$0.1, 0.1–0.2 and $\gt$0.2 using the two-sided Kolmogorov–Smirnov test (Smirnov 1939).

To analyze the importance of different hydro-meteorological contributors to drought spatial extent, we introduce a contributor overlap measure defined as the percentage of catchments under hydrological drought simultaneously affected by precipitation drought, temperature anomaly, SWE deficit, or soil moisture deficit. The higher the overlap of a hydro-meteorological contributor with the area under hydrological drought, the more important is the contributor to explain drought spatial extent. An overlap of 1 (0) means that 100% (0%) of the stations under hydrological drought are affected by a deficit in the contributor considered. Precipitation (P) droughts are defined in the same way as streamflow droughts, using a variable threshold, and based on daily precipitation time series. Temperature (T) anomalies are determined as above threshold events using monthly temperature time series and a variable threshold at the 85% quantile. SWE and soil moisture (SM) deficits are similarly determined using a below-threshold approach on monthly SWE and soil moisture time series, respectively, with a variable threshold at the 15% quantile. In addition to pure overlap time series, we look at overlap ratios for T/P to assess how the relative importance of these two contributors changes. Denominators of zero were replaced by 0.001.

The contributor overlap measure is computed over the whole study domain (CONUS) to determine the overall importance of different hydro-meteorological contributors on drought spatial extent. In addition, it is computed for nine eco-regions with similar regional climatology (Bukovsky regions; Bukovsky 2011) to identify regionally important contributors. Furthermore, we perform a correlation analysis of regional contributor overlap with physiographical and climatic catchment characteristics as provided by the CAMELS dataset (Addor et al 2017) to identify catchment characteristics that might be related to the strength of contributor overlap. The following catchment characteristics are considered: latitude, longitude, catchment area, elevation, mean precipitation, mean potential evapotranspiration, aridity, snow fraction, mean discharge, baseflow index, runoff ratio, soil porosity, soil conductivity, sand fraction, silt fraction, porosity, permeability, and forest cover.

The overlap time series for the four hydro-meteorological variables are used in a trend analysis to determine changes in the importance of different variables as contributors to drought spatial extent. We use the non-parametric Mann–Kendall test (Mann 1945) to compute p-values and the Sen's slope estimator to determine the direction of change (Sen 1968). The results of the trend analysis are mapped per Bukovsky region.

We vary the drought threshold at individual sites (t = 0.1, 0.15, 0.2) and the areal percentage threshold when defining large spatial events (p = 0.15, 0.2, 0.25, 0.3) to investigate the sensitivity of threshold choices on the number of spatial events, event duration and spatial extent. The number of large spatial events extracted lies around 25 if a drought threshold at the 15% quantile or higher and an areal percentage threshold lower than 20% is chosen (figure S1, stacks.iop.org/ERL/16/024038/mmedia). An increase in thresholds results in the selection of fewer events. Event duration and extent also depend on the thresholds chosen with extents hardly exceeding 0.5 even for low drought thresholds. A drought threshold at the 15% flow quantile and an areal percentage of 20% were chosen for the final analysis resulting in 30 spatially large drought events.

At the monthly scale, drought spatial extent varies considerably over time ranging from near zero to a maximum of $\sim$40%, and shows a modest ($\sim$1%/decade) but statistically significant (p-value = 0.00063) increasing trend (figure 2(A)). This increase in spatial extent with time can mainly be attributed to increases in spatial extents at lower extent ranges (i.e. events with $\lt$10% coverage; $\lt$0.1; p-value = 0.0062), while the distributions at higher ranges do not show statistically significant changes (0.1–0.2 and $\gt$0.2, p-values =  0.11436 and 0.92303) (figure 2(b)). In other words: the extent of small spatial events is increasing, while there is little evidence for an increase in the extent of the most geographically extensive events. These changes were assessed by comparing events during the period 2000–2018 to 1981–1999.

Zoom In Zoom Out Reset image size

Within the spatial extent time series, we identify 30 spatially large events (extent $\gt$0.2) with durations of 1–13 months occurring throughout the year (figure 2(b)). The large events generally appear to cluster in time with several large events occurring in the periods 1986–1992, 1998–2003, 2006–2009, and 2010–2018. We find the most severe of these spatial events in terms of extent and duration were the events in 1988 (start: 1988/02, end: 1989/02, duration: 13 months, max. extent: 0.386); 2002 (start: 2002/05, end: 2003/01, duration: 9 months, max. extent: 0.353), 2001 (start: 2001/08, end: 2002/03, duration: 8 months, max. extent: 0.362), 2007 (start: 2007/05, end: 2008/01, duration: 9 months, max. extent: 0.337), and 1981 (start: 1981/01, end: 1981/04, duration: 4 months, max. extent: 0.435). The 1988 event and the events in the early 2000s were also identified as spatially extensive in a model-based study by Andreadis et al (2005).

We now consider the importance of hydro-meteorological contributors in governing the strength of drought spatial extent. To do so, we introduce contributor anomaly overlap as a measure of association, which describes the percentage of catchments in streamflow drought simultaneously affected by a precipitation drought/deficit (P), positive temperature anomaly (T), snow-water-equivalent (SWE) or soil moisture deficit (SM). We define both the meteorological forcings (P and T) and modulating hydrologic storages (SM and SWE) as potential contributors to streamflow drought extent, while recognizing that variability in SM and SWE is driven by variability in P and T in advance of their impact on streamflow. We look at the covariation of each potential contributor with monthly spatial streamflow drought extent to assign temporally proximal driving roles to all four variables. If streamflow drought extent shows a high overlap within a specific month with SWE or SM deficits, we treat these as contributors to streamflow drought. These storage deficits may have been driven in turn by P deficits or above average T, which in our analysis would not be identified as contributors if that influence occurred prior to the month under consideration. By including storages as a distinct driving factor, we are able to highlight their role in modulating the spatial coherence of streamflow drought and to implicitly consider the lagged influence of the climatic contributors precipitation and temperature.

Figure 3 illustrates the overlap measure for the five largest events. The 1981 event mainly affected the eastern part of the US, a large part of which was simultaneously affected by precipitation drought and soil moisture deficit (figure 3(a)). The 1988 event affected a similar region but warm temperature anomalies are more prominent than precipitation deficits (especially in the north; figure 3(b)). In 2001, basins along the west coast and in the Rocky Mountains were jointly affected by streamflow drought with catchments along the east coast (figure 3(c)). Precipitation deficits show high overlap in the east, while soil moisture deficits are more prominent in the Rocky Mountains and temperature anomalies are more prominent in the southwest. Temperature anomalies and soil moisture deficits were also important during the 2002 event, which affected the eastern US simultaneously with the Pacific Northwest and the Rocky Mountains (figure 3(d)). Temperature anomalies were also important during the 2007 event, which affected mainly the eastern and southern portions of the US (figure 3(e)).

Zoom In Zoom Out Reset image size

Across all events, the importance of different hydro-meteorological variables as contributors to drought spatial extent varies substantially (figure 3(f)). While a subset of events do appear to have one primary hydro-meteorological contributor (e.g. 1981: precipitation deficits), streamflow drought is more often associated with a range of underlying contributors that vary by region (e.g. 2002: warm temperature anomalies in the east and soil moisture deficits in the west). That the relative importance of different hydro-meteorological contributors varies on an event-by-event basis is consistent with earlier studies (e.g. for the Pacific Northwest in Bumbaco and Mote 2010).

Soil moisture deficits are the single contributor with the highest mean explanatory power for drought extent (mean overlap ca. 50%) meaning that regions affected by streamflow drought are often simultaneously affected by soil moisture deficits. The direct importance of precipitation deficits and temperature anomalies, on the other hand, varies more widely across events with overlaps ranging from near zero to as high as 80%. The importance of temperature as a contributor during the month of streamflow drought occurrence varies on a seasonal basis, and is relatively low during the cool season (late autumn through early spring) but often quite high during the warm season (late spring through early autumn). The seasonal importance of temperature as a contributor to drought spatial extent corroborates earlier findings showing that temperature strongly influences other drought characteristics such as duration (Southwestern US; Woodhouse et al 2009). SWE deficits have only limited explanatory power for drought spatial extent for the US as a whole but can be important regionally—particularly in the Rocky Mountains where snow water storage represents a large fraction of the water balance.

The importance of individual hydro-meteorological variables for drought spatial extent not only varies by event but also by region as shown by our correlation analysis of regional contributor overlap with catchment characteristics (figure S2). Precipitation droughts are generally important contributors to streamflow drought extent in the eastern US, while they are less important in high-elevation regions with strong snow influences. Temperature is an important contributor in arid and non-forest catchments, while SWE is important at higher latitudes and more generally in places with higher snow fraction. Soil moisture deficits are especially important in lower-elevation regions and in the eastern US.

Over the full CONUS, the importance of precipitation as a contributor to drought spatial extent remains relatively stable over time for all events (figure 4(a), p-value = 0.2753) but decreases for the large events (figure 4(b), p-value: 0.00627). In contrast, temperature becomes more important across all events as a contributor to spatial extent (figure 4(c), p-value: 0.00000). However, this increase is weaker and not statistically significant for the large events alone because the really large events are driven by a combination of precipitation and temperature (figure 4(d), p-value: 0.6121). The strong increase in the relative importance of temperature, combined with the more weakly decreasing relative importance of precipitation, yields a large and statistically robust increase in the ratio of T to P influence (T/P) (figures 4(i) and (j); p-values: 0.00000, and 0.2515). The importance of both SWE and soil moisture remains relatively stable across all events (figures 4(e) and (g), p-values: 0.03961 and 0.06682) though it decreases for large events (figures 4(f) and (h), p-values: 0.0.20404 and 0.00000).

Zoom In Zoom Out Reset image size

Trend analyses for the nine climatic regions reveal substantial regional differences in the monthly overlap time series for the different hydro-meteorologic contributors (figure 5). Precipitation overlap decreases over most regions except the Great Plains (figure 5(a)), while temperature overlap increases in most regions except for portions of the southeast (figure 5(b))—resulting in an overall increase of the importance of temperature relative to precipitation (increase in T/P overlap ratio in all regions except the Great Plains, figure 5(e)). The increase of the importance of temperature relative to precipitation is especially pronounced across the inter-mountain west and Pacific Southwest but is also strong across the eastern US. Changes in SWE deficit overlap are mostly small except in the Pacific Northwest, where we note a substantial increase in SWE deficit overlap with drought spatial extent (figure 5(c)). Finally, the importance of soil moisture as an explanatory variable for drought extent decreases in most regions, with the strongest decreases found across the eastern US (figure 5(d)).

Zoom In Zoom Out Reset image size

Increasing spatial extent of streamflow droughts—as we have identified in the present study in the US and has been previously identified in Great Britain (Rudd et al 2019)—have substantial implications for their associated socioeconomic and environmental impacts. An increase in drought extent, for instance, implies increases in the probability that neighboring or upstream-downstream catchments co-experience drought (Patterson et al 2013). Such an increase in regional drought hazard makes water management considerably more challenging. Inter-basin transfers (Gupta and van der Zaag 2008) may no longer be an option, and water contributions from water-abundant upstream regions to dependent downstream regions may be reduced if upstream and downstream regions co-experience drought (Viviroli et al 2020). For example, Southern California, home to roughly 25 million people, sources water originating in both the north and south Sierra mountain ranges, as well as from the upper Colorado River basin, a strategy which ideally hedges against the risk of co-varying droughts in all source regions (Record et al 2016). A decrease in the possibility of such transfers and contributions may increase the severity of drought impacts and drought risk as potentially more people, ecosystems, and industries are affected. The simultaneous occurrence of drought in several basins and regions may therefore expose weaknesses in existing water management policies and increase the need for coordination among regions from both water supply and demand perspectives.

High temperatures can intensify drought events and support their propagation from one to another region through land-atmospheric feedbacks (e.g. Miralles et al 2019). Our findings show that the importance of temperature as a contributor to drought is not limited to soil moisture droughts (Ault 2020, Williams et al 2020a) but extends to the spatial extent of streamflow droughts particularly during the warm season (late spring through early autumn). The impact of temperature on drought and therefore drought extent is twofold: In winter, increased temperatures decrease snow accumulation, which can lead to time-lagged streamflow deficits later in the year (Bumbaco and Mote 2010). In summer, high temperatures increase evaporative demand which can reduce streamflow directly through in-channel evaporation and indirectly through reduced soil moisture inputs (Luo et al 2017, Dai et al 2018).

The increasing importance of temperature as a contributor to drought spatial extent suggests that future temperature increases might not only lead to increases in soil moisture drought spatial extents (Sheffield and Wood 2008, Dai 2013, Lu et al 2019) and streamflow drought frequencies (Strzepek et al 2010) but related to these also to spatial streamflow drought extents. In relatively moist and cool regions such as the Pacific Northwest, where a lack of snowpack has historically been an important contributor to hydrological drought (Bumbaco and Mote 2010), temperature may be especially influential. Indeed, a decrease in Pacific Northwest snowpack has already been observed as temperature has warmed over the past few decades (Mote et al 2018). In more arid regions, such as the Great Plains and the interior Southwest, temperature affects drought extent primarily through an increase in evaporative demand (Vicente-Serrano et al 2020). Here, too, a temperature driven climate change signal has already been identified in drought trends during the late 21st century (Cook et al 2015, Martin et al 2020). Indeed, temperature changes may be more directly translated into changes in drought spatial extent than precipitation changes as they are more spatially coherent (i.e. virtually the entire Earth is warming, but regional precipitation trends are far more heterogeneous; Wuebbles et al 2014, Cook et al 2020).