Exploring seasonal and regional relationships between the Evaporative Stress Index and surface weather and soil moisture anomalies across the United States

Abstract. This study uses correlation analyses to explore relationships between the
satellite-derived Evaporative Stress Index (ESI) – which depicts
standardized anomalies in an actual to reference evapotranspiration (ET)
fraction – and various land and atmospheric variables that impact
ET. Correlations
between the ESI and forcing variable anomalies calculated over sub-seasonal
timescales were computed at weekly and monthly intervals during the growing
season. Overall, the results revealed that the ESI is most strongly
correlated to anomalies in soil moisture and 2 m dew point depression.
Correlations between the ESI and precipitation were also large across most of
the US; however, they were typically smaller than those associated with soil
moisture and vapor pressure deficit. In contrast, correlations were much
weaker for air temperature, wind speed, and radiation across most of the US,
with the exception of the south-central US where correlations were large for
all variables at some point during the growing season. Together, these
results indicate that changes in soil moisture and near-surface atmospheric
vapor pressure deficit are better predictors of the ESI than precipitation
and air temperature anomalies are by themselves. Large regional and seasonal
dependencies were also observed for each forcing variable. Each of the
regional and seasonal correlation patterns were similar for ESI anomalies
computed over 2-, 4-, and 8-week time periods; however, the maximum
correlations increased as the ESI anomalies were computed over longer time
periods and also shifted toward longer averaging periods for the forcing
variables.


Various options are available for monitoring ET during the growing season. For example, direct measurements of ET can be obtained using flux tower networks such as AmeriFlux and FLUXNET (Baldocchi et al., 2001); however, their utility for largescale monitoring is limited by their poor spatial sampling. High-resolution ET datasets can be generated using sophisticated land surface models such as those included in the North American Land Data Assimilation System (NLDAS) (Xia et al., 2012a, b). Though these datasets are spatially and temporally continuous, their accuracy will depend on the accuracy of the land surface 5 models and the precipitation, atmospheric, soil property, and vegetation datasets that drive them (Beljaars et al., 1996). Highresolution ET datasets can also be generated using satellite observations by linking instantaneous ET rates to observables such as vegetation cover fraction and land surface temperature. Because ET estimates derived from infrared and visible satellite observations can only be computed when clouds do not obscure the surface, more complete domain coverage can be obtained by compositing clear-sky ET estimates over longer time periods (Anderson et al., 2013). Satellite-derived ET datasets covering 10 regional and global domains can be obtained from a variety of sources, such as the MODIS Global Evapotranspiration Project (Mu et al., 2011), the Global Land Evaporation Amsterdam Model (Martens et al., 2017), and the Atmosphere Land Exchange Inverse (ALEXI) model used to compute the Evaporative Stress Index (ESI) (Anderson et al., 2007a(Anderson et al., , b, 2011. Prior studies have shown that the ESI, representing standardized anomalies in the actual-to-reference ET ratio, can provide early warning of drought development because vegetation often curtails its water usage before visible signs of moisture stress 15 become evident in the vegetation (Otkin et al., 2013;Anderson et al., 2013). Though the ESI has primarily been used to monitor agricultural and ecological drought conditions (Anderson et al., 2007b(Anderson et al., , 2011Otkin et al., 2013), and is well suited for the early detection of rapid onset flash drought events (Anderson et al., 2013;Otkin et al., 2015aOtkin et al., , 2016Otkin et al., , 2018, it can also be used to identify regions with healthy vegetation as inferred by higher than average ET rates. As such, it provides useful information about vegetation health under both favorable and unfavorable growing conditions and has been shown to have high correlations 20 to agricultural crop yields (Anderson et al., 2016a, b;Otkin et al., 2016). Furthermore, Otkin et al. (2014Otkin et al. ( , 2015a have shown that unusually rapid decreases in the ESI provide useful information about the likelihood of drought development over the next 1-2 months, presumably due to soil moisture memory and its impact on vegetation. More recently, Lorenz et al. (2017a, b) developed a hybrid-statistical method that combines information from the ESI with precipitation and soil moisture anomalies to predict changes in the U.S. Drought Monitor (Svoboda et al., 2002) over sub-seasonal time scales. Their method had some 25 forecast skill and was shown to provide useful forecasts, especially during flash drought events.
Given the drought monitoring capabilities of the ESI and the desire within the agricultural and natural resources communities for sub-seasonal drought intensification forecasts during the growing season (Otkin et al., 2015b), it is prudent to explore adaptation of the statistical method developed by Lorenz et al. (2017a, b) so that it can be used to predict changes in the ESI rather than the U.S. Drought Monitor because the ESI is a more direct measure of vegetation health. Such efforts would align 30 with the increasing interest within the forecasting community to produce sub-seasonal forecasts that can fill the gap between medium-range weather forecasts and seasonal forecasts (Vitart et al., 2017). As a first step in this process, this study uses correlation analyses to examine relationships between the ESI and various land surface and atmospheric variables that control ET on sub-seasonal time scales. The study explores regional and sub-seasonal changes in the strengths of the correlations using a version of the ESI that covers the contiguous U.S. (CONUS) with 4-km horizontal grid spacing. This study augments 35 3 Hydrol. Earth Syst. Sci. Discuss., https://doi.org/10.5194/hess-2018-108 Manuscript under review for journal Hydrol. Earth Syst. Sci. Discussion started: 5 April 2018 c Author(s) 2018. CC BY 4.0 License. prior analyses by Anderson et al. (2013) that examined correlations between the ESI and various soil moisture, precipitation, and vegetation datasets over seasonal time scales. Information from the correlation analyses will inform efforts to develop sub-seasonal forecasts depicting changes in the ESI or similar quantities. The paper is organized as follows. Section 2 contains descriptions of the atmospheric, soil moisture, and ET datasets used during this study. Results from correlation analyses are shown in Section 3, with conclusions and a discussion provided in Section 4. 5 2 Data and Methodology

Evaporative Stress Index
The ESI shows standardized anomalies in a reference ET fraction (ET/ET ref ), where ET is the actual ET flux and ET ref is a reference ET flux computed using a Penman-Monteith formulation (Allen et al., 1998). Using a reference ET helps minimize the impact of the seasonal cycle in net radiation at the land surface when assessing anomalies in ET. As discussed in Anderson The actual ET flux is estimated using the ALEXI model (Anderson et al., 2007a(Anderson et al., , 2011. ALEXI computes the ground, latent, 15 and sensible heat fluxes for bare soil and vegetated components of the land surface using land surface temperatures retrieved from satellite thermal infrared imagery and the Norman et al. (1995) two-source energy balance model. The partitioning of the surface energy budget into its constituent components is achieved through use of vegetation cover fraction estimates derived from the MODIS leaf area index product (Myneni et al., 2002). For each satellite pixel, the total surface energy budget is computed using the observed increase in land surface temperatures from 1.5 h after local sunrise until 1.5 h before local noon. 20 The atmospheric boundary layer growth model developed by McNaughton and Spriggs (1986) is used to provide closure for the energy balance equations, with temperature profiles in the lower troposphere used by the model obtained from the Climate Forecast System Reanalysis (CFSR) (Saha et al., 2010). The ALEXI model is run daily on a 4-km resolution grid covering CONUS using land surface temperature and insolation estimates derived from the Geostationary Operational Environmental Satellite (GOES) imager. The reader is referred to Anderson et al. (2007aAnderson et al. ( , 2013 for a complete description of the ALEXI 25 model.
Because ET estimates derived from satellite thermal infrared imagery can only be computed for pixels that remain entirely clear during the morning integration period, the resultant daily ET datasets often have extensive data gaps. This issue is partially remedied by compositing the clear-sky ET estimates and corresponding reference ET fluxes over multi-week time periods.
Standardized ET fraction anomalies, expressed as pseudo z-scores normalized to a mean of 0 and a standard deviation of 1, 30 are then computed at weekly intervals using data composited over 2, 4, and 8 week time periods. The mean and standard deviations for each week and compositing period are computed separately for each grid point using data from 2001-2015.
Positive (negative) ESI anomalies depict above (below) normal reference ET fractions that typically correspond to better (worse) than average vegetation health and higher (lower) than average soil moisture content.

North American Land Data Assimilation System
The ESI anomalies were compared to modeled soil moisture anomalies computed using data from three NLDAS-2 models (Xia et al., 2012a, b), including the Noah (Ek et al., 2003;Barlage et al., 2010;Wei et al., 2013), Mosaic (Koster and Suarez, 5 1996), and Variable Infiltration Capacity (Liang et al., 1996) models. Each of these land surface models uses discretized forms of the energy and water balance equations to simulate changes in soil moisture content in multiple layers. Though each model uses the same atmospheric and precipitation forcing datasets, the soil moisture response can differ between models because they may use different approximations to treat key processes such as evaporation, drainage, canopy uptake, and vegetation rooting depth. Given these differences, the ensemble average soil moisture is used here to represent the spatial distribution of 10 soil moisture conditions. Xia et al. (2014) has shown that the ensemble average is more accurate than the individual models at depicting drought conditions. Ensemble mean soil moisture analyses generated each day for the topsoil (0-10 cm) and total column (0-200 cm) layers were averaged over 2-, 4-, and 8-week time periods and then standardized anomalies were computed at weekly intervals for each layer (hereafter referred to as TS and TC, respectively) using data from 1979-2015. These datasets are useful for this study because they provide spatially and temporally continuous soil moisture information across the entire 15 CONUS.

Atmospheric variables
The relationships between the ESI and near-surface atmospheric conditions were evaluated using analyses from the CFSR, which is a fully coupled atmosphere-land-ocean modeling system (Saha et al., 2010). Given their importance for driving changes in ET, this study focuses on 2-m temperature, 2-m dew point depression, 10-m wind speed, and downwelling shortwave 20 radiation (hereafter referred to as TEMP, DPD, WSPD, and DSW, respectively). Daily averages were computed for each variable using analyses available every 6 h on a 38 km resolution grid, and then interpolated to the ESI grid using a nearest neighbor approach. Standardized anomalies were computed at weekly intervals for 2-, 4-, and 8-week averaging periods using data from 1979-2015.

25
Standardized Precipitation Index (SPI) (McKee et al., 1993) anomalies were also computed over 4-, 8-, and 12-week time periods to assess the relationship between the ESI and precipitation. The SPI is a normalized variable such that anomalies greater (less) than zero indicate that the observed precipitation for a given location was more (less) than the climatological mean for a given period of time. Gridded precipitation analyses for 1948-2015 were obtained from the Climate Prediction Center gauge-based analysis of daily precipitation reports from cooperative observers and National Weather Service stations 30 (Higgins et al., 2000), with the 0.25 • resolution daily precipitation analyses interpolated to the ESI grid using a nearest neighbor 5 Hydrol. Earth Syst. Sci. Discuss., https://doi.org/10.5194/hess-2018-108 Manuscript under review for journal Hydrol. Earth Syst. Sci. Discussion started: 5 April 2018 c Author(s) 2018. CC BY 4.0 License.
approach. The daily datasets were summed to create 4-, 8-, and 12-week accumulated precipitation amounts prior to computing the SPI. The SPI anomalies were computed over longer time periods than the other variables because Anderson et al. (2011) has shown that ET as a physical process retains some memory of precipitation prior to the start of the compositing period.

Monthly correlation analysis 5
In this section, the relationship between the ESI and various atmospheric and land surface variables is assessed during the warm season using correlation analyses. Figures 1 and 2 show the Pearson correlation coefficients between the 4-wk ESI and the corresponding 4-wk SPI, TS, TC, DPD, TEMP, WSPD, and DSW anomalies at monthly intervals from April to September.
Note that the sign is reversed for the DPD, TEMP, WSPD, and DSW correlations given the expectation that larger (smaller) values for each of these variables will typically be associated with higher (lower) moisture stress and negative (positive) ESI 10 anomalies when assessed over long time periods. The correlations were computed separately for each grid point and month using all of the weekly analyses from 2001-2015 for which the end of the 4-wk period fell within a given month. Using 4-wk periods for all of the datasets allows us to examine the contemporaneous relationships between the ESI and the various forcing variables across the entire U.S. during different portions of the growing season. The 4-wk ESI was chosen for this part of the analysis because it provides a balance between the fast response of the ESI to changing conditions when it is computed over 15 short time periods and the seasonal moisture stress signals contained in longer-term ESI anomalies (Otkin et al., 2013). A more comprehensive assessment of the relationships between the various forcing variables and ESI anomalies computed using different compositing lengths is presented in the regional analysis shown in Sec. 3.2. The results also show that the strengths of these relationships vary during the growing season across this region. For example, the correlations for DSW are largest during the spring and early summer when surface radiation anomalies due to changes in cloud cover influence the timing and vigor of early plant growth and its release of ET, whereas TEMP anomalies are more Finally, the correlations over most of California have a distinct seasonal cycle that is opposite that found in the heavily forested areas across the northern U.S. For example, with the exception of WSPD, all of the variables are strongly correlated 15 (> 0.6) to the ESI during the spring. These correlations then rapidly decrease after June and are generally < 0.2 across most of the state in August and September. This sequence suggests that ET is strongly influenced by the amount of cool season precipitation, which is a key component of the hydrological cycle in the region (Neiman et al., 2008), and its subsequent impact on soil moisture during the first half of the growing season. The results indicate that a wetter (drier) than normal cool season that then leads to wetter (drier) than normal soil moisture conditions during the spring are typically associated with 20 above (below) normal ET. The results also show that anomalies in evaporative demand strongly impact ET during this time period as evidenced by the large correlations between the ESI and the DPD, TEMP, and DSW variables from April to June.

This section assesses changes in the correlations between the ESI and the various forcing variables at weekly intervals from
March to October when different compositing and averaging lengths are used to compute the anomalies for each variable. 25 This analysis will be used to assess relationships between the ESI computed over short-to-intermediate time scales (2, 4, and 8 weeks) and each of the forcing variables computed over similar time periods. To capture regional differences in the relationships, the correlations were computed separately for the western, south central, north central, and eastern U.S. using data from 2001-2015. The outlines for each region are shown in Fig. 3. By assessing the relationships over such large regions, some of the local details discussed in Section 2.1 will be lost; however, this approach makes discussion of the results more   also shows that short-term ET anomalies in these drier areas are fundamentally tied to the availability of soil moisture over short time periods.
To assess how these relationships change when ESI anomalies are computed over longer time periods, Figs. 5 and 6 show correlations between each variable and the 4-and 8-wk ESI, respectively. Comparison to the 2-wk ESI correlations shown in Fig. 4 reveals that the seasonal patterns in the correlations remain similar for each region and variable when the ESI anomalies 25 are computed over longer time periods; however, the maximum correlations shift toward longer averaging periods for most of the variables. This shift shows that the longer duration ESI anomalies are most closely related to atmospheric and land surface anomalies occurring over similarly long time scales, while still having some sensitivity to shorter fluctuations in these variables. The maximum correlation for a given variable also tends to increase as the ESI compositing period increases from 2 to 8 weeks, with the DPD, TS, and TC variables having the largest correlations in each region during most of the growing 30 season regardless of the length of time used to compute the ESI.
An interesting pattern emerges when comparing the correlations for the DPD, TS, and TC variables. Whereas all three variables had their largest correlations to the 2-wk ESI when their anomalies were computed over 2-and 4-wk time periods, their behavior diverges for the 8-wk ESI for which the maximum correlations shift to longer averaging periods for the TS and DPD variables but remain large for TC regardless of the length of time used to compute anomalies for that variable. This pattern occurs in all of the regions, and shows that as the ESI anomalies are computed over longer time periods, they become most closely related to TS and DPD anomalies occurring over similar time scales, but to TC anomalies occurring over all time scales. This behavior is likely due to the tendency for TC soil moisture to change more slowly than DPD and TS soil moistureboth of which are more strongly influenced by synoptic-scale (e.g. weekly) weather features -and thus remain closely related to the ESI over multiple time scales.

5
As was the case with the 2-wk ESI, longer-term ESI anomalies are most strongly correlated to DSW across the central and western U.S. during the first half of the growing season. Likewise, SPI and WSPD correlations continue to be the largest over the south-central U.S, with weaker correlations found elsewhere. The relationship between the ESI and TEMP also remains strong over the south-central U.S., where correlations exceed 0.5 from June until the middle of September. In all of the regions, the TEMP correlations for a given averaging period are routinely smaller than the corresponding DPD correlations during the 10 entire growing season for the 2-, 4-, and 8-wk ESI. This indicates that TEMP anomalies are not a dominant driver of changes in ET; rather, it is near surface humidity that is most important. For example, a period characterized by hot temperatures may not necessarily lead to increased moisture stress (e.g., negative ESI) if it also accompanied by heavy rainfall, which is a common occurrence across the central and eastern U.S. during the summer. Instead, if hot temperatures occur alongside lower dew point temperatures, the resultant increase in the DPD will have a larger impact on ET than the higher TEMP alone. The stronger 15 relationship between the ESI and DPD is consistent with prior work that has shown that stomatal conductance and the release of ET by many plant species is strongly controlled by the vapor pressure deficit (Oren et al., 1999). These results also suggest that drought forecasts that rely upon monthly-to-seasonal temperature outlooks to predict changes in vegetation health may be more accurate if anomalies in near surface humidity are also considered. 20 This study used correlation analyses to explore relationships between the satellite-derived ESI -which depicts anomalies in an actual to reference ET fraction -and a set of land and atmospheric variables that are known to influence ET through their impact on soil moisture and evaporative demand. Overall, the results showed that anomalies in ET as expressed by the ESI are most strongly correlated to anomalies in soil moisture and near-surface humidity (TS, TC, and DPD) regardless of the time period over which the anomalies are computed. Correlations between the ESI and precipitation (SPI) are also relatively large 25 across most of the U.S.; however, they are typically smaller than the TS, TC, and DPD correlations for a given location and time of year. The strong correlations to soil moisture over sub-seasonal time scales are consistent with results from Anderson et al. (2011Anderson et al. ( , 2013 that showed that the ESI is also strongly correlated to soil moisture over seasonal time scales. In contrast, and then increase during the second half of the growing season. The weak relationships to the ESI indicate that there are no dominant drivers of ET during the first half of the growing season in these northern locations; however, ET becomes more strongly coupled to the forcing variables later in the growing season as the regions transition from energy-limited regimes to potentially moisture-limited regimes.

Conclusions and discussion
Large sub-seasonal fluctuations in the correlations are also evident in some of the variables across parts of the U.S. For 5 example, correlations to DSW are large across the central U.S. during the spring and early summer, whereas TEMP anomalies become more important during the second half of the growing season. Over the western U.S., the correlations are much lower for all variables during the climatological peak of the North American Monsoon during July and August when compared to other parts of the year. California also has a distinct seasonal pattern where the correlations are largest during the spring and then rapidly diminish after June. Each of the regional and seasonal patterns were similar for ESI anomalies computed over 10 2-, 4-, and 8-wk time periods; however, the maximum correlations typically increased as the ESI anomalies were computed over longer time periods and also shifted toward longer averaging periods for the forcing variables. This shift shows that ESI anomalies computed over short (long) time periods are most strongly correlated to atmospheric and land surface anomalies occurring over similar time scales, while also having some sensitivity to anomalous conditions occurring over longer (shorter) time periods. 15 Investigation of the monthly and regional correlations also showed that anomalies in the ESI are typically much more strongly correlated to anomalies in the DPD than they are to anomalies in TEMP during the entire growing season across most of the U.S. This indicates that in most situations it is the vapor pressure deficit rather than air temperature that is the most important driver of changes in ET. Likewise, though the correlations for SPI are relatively large, they are still generally smaller than those associated with the soil moisture variables. Together, these results indicate that fluctuations in soil moisture and near-20 surface humidity are better predictors of the ESI than are SPI and TEMP anomalies by themselves. This is consistent with a recent study by Ford and Labosier (2018) that showed that temperature and rainfall departures by themselves were only weakly related to the occurrence of rapid-onset flash droughts, whereas variables that explicitly account for changes in soil moisture content and near-surface humidity were more closely linked to the development of these features. These findings also illustrate that existing monthly-to-seasonal outlooks that tend to focus on predicting anomalies in air temperature and precipitation are 25 insufficient for predicting changes in agricultural or ecological drought conditions. Instead, greater focus should be placed on predicting changes in soil moisture and vapor pressure deficit given their more dominant influence on ET. Indeed, a recent study by Lorenz et al. (2018) has shown that inclusion of vapor pressure deficit and soil moisture predictions from climate models increased the accuracy of sub-seasonal drought intensification forecasts generated using a hybrid statistical method. This is also supported by a study by Seager et al. (2015) that showed that large forest fires are often associated with very large 30 vapor pressure deficits caused by antecedent surface drying and large-scale subsidence.
Competing interests. The authors declare that there are no competing interests with the work performed during this study.