Contribution of Meteorological Downscaling to Skill and Precision of Seasonal Drought Forecasts

Ryan A. Zamora aThe Johns Hopkins University, Baltimore, Maryland

Search for other papers by Ryan A. Zamora in
Current site
Google Scholar
PubMed
Close
,
Benjamin F. Zaitchik aThe Johns Hopkins University, Baltimore, Maryland

Search for other papers by Benjamin F. Zaitchik in
Current site
Google Scholar
PubMed
Close
,
Matthew Rodell bHydrological Sciences Laboratory, NASA Goddard Space Flight Center, Greenbelt, Maryland

Search for other papers by Matthew Rodell in
Current site
Google Scholar
PubMed
Close
,
Augusto Getirana bHydrological Sciences Laboratory, NASA Goddard Space Flight Center, Greenbelt, Maryland

Search for other papers by Augusto Getirana in
Current site
Google Scholar
PubMed
Close
,
Sujay Kumar bHydrological Sciences Laboratory, NASA Goddard Space Flight Center, Greenbelt, Maryland

Search for other papers by Sujay Kumar in
Current site
Google Scholar
PubMed
Close
,
Kristi Arsenault cScience Applications International Corporation, McLean, Virginia
dHydrological Sciences Laboratory, NASA Goddard Space Flight Center, Greenbelt, Maryland

Search for other papers by Kristi Arsenault in
Current site
Google Scholar
PubMed
Close
, and
Ethan Gutmann eNational Center for Atmospheric Research, Boulder, Colorado

Search for other papers by Ethan Gutmann in
Current site
Google Scholar
PubMed
Close
Full access

Abstract

Research in meteorological prediction on subseasonal to seasonal (S2S) time scales has seen growth in recent years. Concurrent with this growth, demand for seasonal drought forecasting has risen. While there is obvious synergy between these fields, S2S meteorological forecasting has typically focused on low-resolution global models, whereas the development of drought can be sensitive to the local expression of weather anomalies and their interaction with local surface properties and processes. This suggests that downscaling might play an important role in the application of meteorological S2S forecasts to skillful forecasting of drought. Here, we apply the generalized analog regression downscaling (GARD) algorithm to downscale meteorological hindcasts from the NASA Goddard Earth Observing System global S2S forecast system. Downscaled meteorological fields are then applied to drive offline simulations with the Catchment Land Surface Model to forecast U.S. Drought Monitor–style drought indicators derived from simulated surface hydrology variables. We compare the representation of drought in these downscaled hindcasts with hindcasts that are not downscaled, using the North American Land Data Assimilation System Phase 2 (NLDAS-2) dataset as an observational reference. We find that downscaling using GARD improves hindcasts of temperature and temperature anomalies but that the results for precipitation are mixed and generally small. Overall, GARD downscaling led to improved hindcast skill for total drought across the contiguous United States, and improvements were greatest for extreme (D3) and exceptional (D4) drought categories.

© 2021 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Ryan A. Zamora, ryan.zamora@jhu.edu

Abstract

Research in meteorological prediction on subseasonal to seasonal (S2S) time scales has seen growth in recent years. Concurrent with this growth, demand for seasonal drought forecasting has risen. While there is obvious synergy between these fields, S2S meteorological forecasting has typically focused on low-resolution global models, whereas the development of drought can be sensitive to the local expression of weather anomalies and their interaction with local surface properties and processes. This suggests that downscaling might play an important role in the application of meteorological S2S forecasts to skillful forecasting of drought. Here, we apply the generalized analog regression downscaling (GARD) algorithm to downscale meteorological hindcasts from the NASA Goddard Earth Observing System global S2S forecast system. Downscaled meteorological fields are then applied to drive offline simulations with the Catchment Land Surface Model to forecast U.S. Drought Monitor–style drought indicators derived from simulated surface hydrology variables. We compare the representation of drought in these downscaled hindcasts with hindcasts that are not downscaled, using the North American Land Data Assimilation System Phase 2 (NLDAS-2) dataset as an observational reference. We find that downscaling using GARD improves hindcasts of temperature and temperature anomalies but that the results for precipitation are mixed and generally small. Overall, GARD downscaling led to improved hindcast skill for total drought across the contiguous United States, and improvements were greatest for extreme (D3) and exceptional (D4) drought categories.

© 2021 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Ryan A. Zamora, ryan.zamora@jhu.edu

1. Introduction

Droughts are one of the most damaging climate phenomena, impacting agriculture, water resources, ecosystems, and human health at local to regional scale. Economic impacts are also significant. In the United States, the cost of drought to the economy since 1980 has been estimated to be at least $249 billion, accounting for costs from 26 unique events (NCEI 2020). Early warning of drought has the potential to reduce these costs, through improved preparedness and more rapid impact response (Wilhite and Svoboda 2000), yet most seasonal drought prediction systems are currently working only in a research capacity, with few dynamically based systems applied operationally to drought preparation and management.

Research and interest in subseasonal to seasonal (S2S) meteorological prediction has risen over the past two decades, as is evident in the appearance of multiple intermodel comparison projects and operational forecast ensembles. These include the North American Multimodel Ensemble (NMME; Kirtman et al. 2014), the S2S Prediction Project (Vitart et al. 2017), and the Subseasonal Experiment (SubX; Pegion et al. 2019). Objectives of these efforts include testing the predictability of various climate phenomena and application to regional climate outlooks (Thober et al. 2015; Sossa et al. 2017; Vitart 2017; Kim et al. 2019a). The bulk of these efforts has focused on global dynamically based prediction systems, with local applications for weather hazards, agricultural impacts, and water resource management (among other sectors) generally mediated by local expert interpretation rather than by a formal customization or downscaling of the global forecast models.

Recent research on S2S time scales has spanned a number of different subjects, such as the Madden–Julian oscillation (Jenney et al. 2019; Kim et al. 2019b; Wang et al. 2020), aerosols (Benedetti and Vitart 2018), and tropical cyclones (Robertson et al. 2020). As a number of key impacts of S2S climate variability involve or are caused by changes in hydrology, interest in hydrological S2S prediction has also increased (Wood et al. 2002, 2004; Shukla et al. 2014; Wanders and Wood 2016; Arsenault et al. 2020; DeAngelis et al. 2020; Pendergrass et al. 2020). This increased interest has also resulted in a number of operational dynamically based drought monitoring and prediction systems such as The U.S Drought Monitor (https://droughtmonitor.unl.edu/), the Climate Prediction Center Monthly and Seasonal Drought Outlook (https://www.cpc.ncep.noaa.gov/products/Drought/), and the Drought Monitoring and Prediction System (Luo and Wood 2007). Hydrological S2S forecasts generally rely on atmospheric S2S forecasts to provide meteorological forcing data. As the simulated hydrological impacts of meteorological conditions can be highly sensitive to spatial and temporal resolution (Berne et al. 2004), this raises the question of whether it is important to downscale the output of global S2S forecast systems before applying them for hydrological prediction. The answer to this question depends on the characteristics of the modeling system, the climate and hydrology of the regions of interest, and on the applications context—forecasts focused on problems of hydropower reservoir management, for example, might have different needs from those focused on agricultural drought.

Since 2012, the NASA Goddard Space Flight Center Hydrological Sciences Laboratory has collaborated with the U.S. National Drought Mitigation Center (NDMC) to provide drought monitoring products informed by the Gravity Recovery and Climate Experiment (GRACE) and GRACE-Follow On (GRACE-FO) satellite missions. In this system, GRACE-derived estimates of terrestrial water storage anomalies are assimilated into the Catchment Land Surface Model (CLSM) to provide continuous monitoring of drought and wet anomalies in near-surface soil moisture, root zone soil moisture, and shallow groundwater (Houborg et al. 2012; Li et al. 2019). This information is disseminated publicly by NDMC and is also provided directly to authors of the U.S. Drought Monitor as an input for their expert-informed drought maps. In monitoring mode, this system draws meteorological data from the North American Land Data Assimilation System Phase 2 (NLDAS-2; Xia et al. 2012). Recently, the system has been enhanced to provide seasonal forecasts in addition to drought and wetness monitoring (Getirana et al. 2020). In forecast mode, the system draws meteorology from the NASA Global Earth Observation System (GEOS) S2S forecast ensemble (Molod et al. 2012). The objective is to provide season-ahead forecasts of the same drought/wetness indicators that are offered in monitoring mode.

Here, we assess the importance of downscaling GEOS forecast meteorological fields for use in this system. Where Getirana et al. (2020) bias corrected but did not downscale GEOS forcing, we test an implementation of the forecast system that applies the generalized analog and regression downscaling (GARD; Gutmann et al. 2021, manuscript submitted to J. Hydrometeor.) tool to GEOS surface meteorological fields. GARD implements a hybrid analog–regression approach to downscaling, in which multiple input predictor variables can be used to estimate each variable at each grid cell. First, an analog approach is used to select a group of analog days from the training period for each day to be predicted. These analog days are then used to train a multivariable regression, and this regression is applied to predict a downscaled value on the day of interest. In comparison with commonly used statistical disaggregation techniques like bias correction and spatial disaggregation (BCSD; Wood et al. 2004), GARD allows the downscaling process to be informed by multiple variables at multiple scales, potentially taking advantage of skill in an atmospheric forecast even when the model’s prediction of a particular target variable (e.g., local precipitation) has large errors. In the context of seasonal forecasts, the use of GARD is novel in that it offers the potential to improve a dynamically based seasonal forecast by correcting for systematic errors in the large-scale forecast related to both the spatial placement and temporal variability of weather processes.

Given the large uncertainty in S2S atmospheric forecasts, and the potential for errors in meteorological inputs to have nonlinear impacts on hydrological simulations, skillful downscaling of a global forecast system could be very important when forecasting drought indices. This paper evaluates this sensitivity for the case of an operational drought and wetness monitoring system for the United States.

2. Data

GARD, like most statistical downscaling methods, requires a reliable high-resolution “observational” dataset with a sufficiently long record that includes the meteorological forcing variables being downscaled, along with a training dataset that is representative of the spatially coarse forecast data that need to be downscaled. In our case, we need to downscale the nine surface meteorological variables required by the CLSM: air temperature at 2 m above the surface (T2M), U and V wind components at 10 m above the surface (U10M and V10M, respectively), specific humidity at 2 m above the surface (Q2M), surface pressure (PS), surface downward longwave and shortwave radiation (LWS and SLRSF, respectively), total precipitation (PRECTOT), and convective precipitation (CNPRCP).

a. Observational dataset

The observational dataset we use in this study is NLDAS-2. The data have a fine spatial grid (0.125°), frequent temporal resolution (hourly), and start in 1979, with near-real-time data released to present day. NLDAS-2 is not a purely observational dataset, in that it is derived from a combination of models and observations, but we refer to it as the observational dataset here because of its role in the downscaling process. Nevertheless, its long record, high resolution, and extensive history of evaluation and successful use in applications make it an appealing dataset to use as a high-resolution training dataset for our downscaling purposes. For our downscaling technique, we need 6-hourly, daily averaged, and climatological averaged daily resolution datasets; here we take hourly observations sampled at 0, 6, 12, and 18 h to make our 6-hourly dataset, and average these data to generate a daily dataset for the period 1982–2017 (to match the temporal availability of our forecasting dataset). Using the average of 6-h intervals rather than hourly NLDAS data introduces some error to the diurnal cycle, and this error will be time-zone dependent. The approach is used to be consistent across NLDAS and GEOS. We then create a long-term daily averaged climatological dataset by averaging each day of the year over the entire period.

b. Forecasting dataset

We use the NASA Goddard Earth Observing System S2S forecast ensemble (GEOS-S2S) for our forecasting and training dataset (Molod et al. 2012; Borovikov et al. 2019). Here we use the original version of the system (version 1) to study performance over a long hindcast record. This dataset operates 11 ensembles (4 unperturbed lagged forecasts initialized at 5-day intervals, and 7 perturbed simultaneous forecasts) per forecast month and has an ~1° global spatial grid. The available ensembles are generated from perturbed and unperturbed initial states with start dates closest to the start of the month. GEOS has recently been updated, with some modifications to the model and ensemble strategy.

Downscaling an ensemble forecasting product allows us the opportunity to apply the downscaling technique to multiple forecast realizations using a single observational dataset. Each ensemble member is treated as a unique dataset, downscaled individually, and is only averaged post analysis for purposes of plotting and presentation. For this study, we focus on 3-month hindcasts initialized on 1 May of each year. GEOS forecasts run for 9 months, but we focus on the first 3 months because meteorological forecast skill degrades significantly over time. The surface meteorological forcing fields taken from this dataset have a daily temporal resolution and match those of the observation-based dataset and span the period 1982–2017. For consistency with the overall project (Getirana et al. 2020), we chose to limit the results shown here to the 2003–17 period.

3. Methods

a. Generalized analog regression downscaling

We utilize the National Center for Atmospheric Research (NCAR) GARD algorithm for our downscaling purposes. As its name suggests, GARD has capabilities of downscaling using an analog approach, regression approach, or a combination of the two, which we utilize in our study. It requires three input datasets for downscaling: an observational dataset taken to be “truth” (in our case NLDAS-2), a training dataset that is compared with our observational dataset (historical GEOS hindcasts), and a prediction dataset that we wish to downscale (GEOS for the time period to be downscaled).

The observational and training datasets must have identical time steps for the algorithm to create a link between the coarse- and fine-resolution data points; this is necessary for finding analog values for the predictand variable. The algorithm examines a single data point in the prediction dataset for analog values of that grid cell in the training dataset (and its paired corresponding value in the observational dataset) and performs a regression on this subset to determine its downscaled value. Importantly, the regression applied in this process can be multivariable, such that information from multiple aspects of the atmospheric forecast can inform the estimate of the meteorological value of interest. This results in a regression only being applied using observation and training data pairs that are analogous to the value we want to downscale; for example, if we want to downscale a temperature value where weather conditions are cold and have low winds, the downscaled value is only predicted using days that share these conditions. This process is repeated on a gridcell by gridcell basis for each data point in the prediction dataset, resulting in gridded predicted values with a resolution of the observational dataset.

The regression applied within GARD is a generalized linear model of the form
y^(C,X)=c0+c1x1++cpxp+e,
where C = c1, …, cp are the regression coefficients, c0 is the intercept, X = x1, …, xp are the predictor variables, e is the residuals of the model, and y^ represents the desired predictand. Similarly, if a logistic regression is needed such as in cases in which predicting the occurrence of a threshold being exceeded [e.g., the occurrence of precipitation (values greater than 0), or the occurrence of high-temperature days (values greater than 95°F, i.e., 35°C, in a tropical region)], a logistic regression is applied within GARD of the form
p^(C,X)=11+e(c0+c1x1++cpxp),
where p^ is the predicted probability of exceedance, and the other variables are as in Eq. (1). For a more detailed description of GARD and its configurations, see Gutmann et al. (2021, manuscript submitted to J. Hydrometeor.).

The results GARD produces are fundamentally different than only correcting biases in the mean of the target variable. While part of the technique inherently does accomplish a bias correction, the analog and regression approach used in GARD can also inform estimates of temporal variability (anomalies) in the downscaled variable. In addition, techniques such as BCSD inherently preserve the large-scale patterns in the forecast, while only correcting for a climatological bias, GARD permits the other variables (e.g., wind direction) to influence the downscaling process, and can thus result in significantly different fields at both a large scale, and a small scale.

b. GARD configuration

There are various options that can be set in the downscaling algorithm to change its function. A full review of GARD configuration options and sensitivities is beyond the scope of this paper (see Gutmann et al. 2021, manuscript submitted to J. Hydrometeor.), but we will address the options most relevant to our implementation of this technique. While GARD has the ability to downscale using a pure regression or pure analog approach, we use the hybrid analog regression to fully utilize both features in the code.

The number of analogs chosen will affect the output of the regression: choosing too few analogs will yield a regression from a biased sample, whereas using too many reduces the value of using analog–regression over regression alone. Therefore, we ran a number of experiments, varying only the number of analogs to find a suitable number, and found that 300 (~10% of the training dataset) worked for our purposes. In addition, we set the option to weigh each analog by how close its value (inverse-square distance) is to the current predictand.

Selecting the variables to include in the regression is also consequential. Theoretically, the more information you have on the large-scale circulation of your domain, the more accurate the resultant regression would be. Ideally this would include variables such as geopotential height or upper-level winds; however, a major tenet of our project is to create a self-contained system that can be used operationally for downscaled seasonal drought/flood forecasting without placing overwhelming data burdens on the forecaster. Thus, we limit the variables used in multivariate regressions to the set of nine variables (Table 1) we are downscaling. While a formal study on the choice of predictors (particularly by region) would be an interesting research topic, it is beyond the scope of this paper, and we use the variables found most important in our testing.

Table 1.

The nine downscaled variables and their associated input predictors used for multivariate regression. The level of all variables is at or near the surface.

Table 1.

Total and convective precipitation need to be handled in a distinct way, due to zero inflation in records of 6-hourly precipitation. GARD offers an option to apply a transform to the input data, perform the downscaling, and then apply the inverse transform on the output data; for precipitation, we use a cube root transformation. Furthermore, a logistic regression is applied to compute the probability of occurrence of precipitation, which we can then use in postprocessing.

Output precipitation-based variables have a conditional bias owing to the distribution of precipitation being non-Gaussian. GARD optionally outputs the probability of exceeding set threshold values (in the case of precipitation this value is 0, indicating no precipitation); in this scenario, this is analogous to the probability of precipitation (PoP). Using this in conjunction with the residuals of the regression used to compute the mean precipitation amounts, we can then use a spatial–temporal autocorrelation technique (Clark and Slater 2006; Gutmann et al. 2021, manuscript submitted to J. Hydrometeor.) to postprocess precipitation. In conjunction with the PoP field, a spatio-temporally correlated random field is used to stochastically determine if precipitation occurs or not while maintaining realistic spatiotemporal variability.

In addition, the temporal resolution of our downscaled forecast dataset is daily and needs to be temporally disaggregated to 6-hourly as an additional step prior to input to LIS. We achieve this by first creating a 6-hourly long term mean of our observational dataset, followed by a transformation of these data into 6-hourly anomalies of the long-term daily mean (thus capturing the diurnal cycle). We then apply these anomalies to our daily output forecast datasets to generate temporally disaggregated 6-hourly forecast data. The type of anomalies used can be either additive (shifting the values) or multiplicative (applying a ratio to the values), and each is chosen depending on the type of variable to be disaggregated; for this study, we use multiplicative anomalies for precipitation variables (convective and total) and use additive anomalies for other variables.

c. Land surface modeling

We use NASA’s Land Information System (LIS; Kumar et al. 2006) as a framework for our land surface modeling, which can easily be used with each of our datasets and has been used extensively in a number of studies (e.g., Seneviratne et al. 2010; Zaitchik et al. 2010; Clark et al. 2015; Kumar et al. 2016; McNally et al. 2017). Within LIS, we use the CLSM (Koster et al. 2000) with our downscaled meteorological variables as a forcing dataset. This experiment could be repeated with a number of different land surface models within LIS; however, an extensive study comparing the differences between these is beyond the scope of this paper. CLSM was originally chosen for the drought/wetness monitoring application because it simulates groundwater storage changes, which is essential for GRACE data assimilation. It is used in drought/wetness forecasting application, and hence in this study, for the same reason.

Although GEOS is a global product, our model domain strictly covers the contiguous United States (CONUS; to the extents of the NLDAS-2 product). As we are conducting a comparative study of three experiments using distinct meteorological forcing fields, it is important that the initial conditions of each are similar. While the focus of this research is on the most recent decades, to properly create drought/wetness indicators (based on climatological soil moisture percentiles), we require a longer record than our datasets provide. Because of this, we opt to use the Princeton meteorological reanalysis dataset (Sheffield et al. 2006), which encompasses 1948–2014, as the meteorological forcing to spinup the CLSM states. We initialize our NLDAS-2 meteorological forcing–based runs using the resulting initial conditions to generate our observational hydrological data. However, because the climatologies differ between the datasets, we rescale the NLDAS-2 meteorological fields to match Princeton’s monthly climatology for our run period using simple scaling factors (Houborg et al. 2012). We then use these same initial conditions to initialize both the original GEOS and downscaled GEOS hindcasts. There are a number of output hydrological variables from this experiment; however, we will focus on only three of them here: soil moisture, evapotranspiration, and runoff.

d. Drought indicators

The U.S. Drought Monitor (USDM; Svoboda et al. 2002) is widely available to the public via its website (https://droughtmonitor.unl.edu/) hosted by the NDMC at the University of Nebraska–Lincoln. It is used in a number of different sectors for decision-making including but not limited to declaring drought disaster situations, prescribing federal aid for farmers, performing state and local level drought assessments, and informing farmer planting and management decisions. The USDM provides a weekly operational product that is generated by a team of authors including one lead (the position rotates). The team subjectively combines information from a number of sources, including modeled climatological conditions, accumulated precipitation and other compiled indices, satellite-based measurements, local observations, and input from state climatologists, other experts, and a network of volunteer observers. The observers receive a draft of the map ahead of time and provide feedback and suggestions based on the knowledge they have on the conditions at their location.

The different categories of drought can be determined using a variety of different variables and indices, such as the Palmer drought severity index, the standard precipitation index, and modeled soil moisture or streamflow percentiles. For the research presented here, we use root zone soil moisture percentiles on account of their relevance to agriculture, a primary sector of interest for drought monitoring. These percentiles are created by comparing the local soil moisture with its climatology; lower percentiles indicate drier-than-normal conditions for that location and time of year, while higher values indicate wetter-than-normal conditions. The five categories the NDMC uses to identify drought are D0 (abnormally dry; 21%–30%), D1 (moderate drought; 11%–20%), D2 (severe drought; 6%–10%), D3 (extreme drought; 3% –5%), and D4 (exceptional drought; 0%–2%).

To compile the spatial drought indicators and simplify the analysis, we aggregate the results to distinct regions of the United States. Although we could have regionalized by state, USGS hydrologic unit code (HUC), or by applying a statistical regionalization technique, we chose to use Bukovsky regions (Bukovsky 2011). These are a collection of 29 regions (with 13 larger compound groupings) that have been delineated based on ecoregions; that is, they respond similarly to variations in temperature and precipitation, or share an important regional climate feature. To use this resource with our datasets, we interpolated the masks (defined on a 0.5° latitude/longitude grid) to the resolution of NLDAS-2, 0.125°. For the purposes of our research, we focused on five compound regions that cover the majority of the CONUS: the Central United States, Eastern United States, Southern United States (hereinafter United States will be shortened to U.S. for brevity), Desert, and Mountain West. While this does exclude some regions of the CONUS (viz., parts of California and Michigan), we believe that using only compound regions simplified the presentation of the analysis and was sufficient for our research. The boundaries and locations of each of these compound regions can be seen in Fig. 1.

Fig. 1.
Fig. 1.

Composite Bukovsky regions used to examine changes among similar ecological regions.

Citation: Journal of Hydrometeorology 22, 8; 10.1175/JHM-D-20-0259.1

4. Results

a. Downscaled meteorological fields

We begin with an examination of the downscaled meteorological fields from a climatological perspective (i.e., averaged over 2003–17 for each month individually). While each of the nine variables provided in Table 1 is used as input to the land surface model, here we focus on temperature and precipitation, given their central importance to local climate and hydrology. In viewing Figs. 39, it is helpful to remember that blues indicate more skill in GARD relative to GEOS while reds indicate the opposite. In addition, each of these figures include an inset table showing the long-term monthly regional averages of the variables shown.

Figure 2 shows an example of monthly averaged 2-m temperature and total precipitation for NLDAS, GEOS, and GARD. These plots were created by averaging the initial 31 days of the forecast for May in the year 2004. This figure showcases the resolution differences between observations and the raw downscaled data. This emphasizes a key characteristic of downscaling the original GEOS forecasts: coarser-resolution datasets have difficulty in resolving topography at high elevations. This is particularly important for hydrological forecasts in regions with significant snowpack, and it can also influence water balance simulations via temperature impacts on potential evaporation and snowmelt.

Fig. 2.
Fig. 2.

Monthly averaged (left) 2-m temperature (K) and (right) total precipitation(mm) from (top) NLDAS, (middle) GEOS, and (bottom) GARD for an example month (May 2004).

Citation: Journal of Hydrometeorology 22, 8; 10.1175/JHM-D-20-0259.1

Downscaled temperature shows greater detail over the mountains, resolving elevation contrasts, and generally captures many of the topographic features found in observations. Precipitation is a much more difficult meteorological phenomenon to forecast, such that GARD shows significant differences from NLDAS in a number of regions. The overall distribution of precipitation statistics in GARD does, however, match NLDAS more closely than GEOS does, reflecting the ability of the downscaling algorithm to replicate higher-intensity precipitation at sub-GEOS resolution. Notably, because we use multivariate regression through downscaling, GARD precipitation is a result of information from many variables (total precipitation, convective precipitation, temperature, and surface pressure; see Table 1). This fact, in addition to the decrease in accuracy of precipitation forecasts overtime (e.g., Koster et al. 2004), results with some regions having different precipitation patterns between GARD and GEOS.

Figure 3 shows the climatological difference between downscaled forecasts (which we will refer to as FGARD) and raw forecasts (which we will refer to as FGEOS) in absolute bias of 2-m temperature for May (top panel), June (middle panel), and July (bottom panel), where bias is calculated for each dataset relative to NLDAS-2. This figure highlights where the greatest departure from NLDAS-2 between the two datasets occurs, regardless of whether the value is under or over predicted. Warmer colors are indicative of areas where FGARD has a larger absolute bias (greater departure from the true value), whereas cooler colors highlight areas of larger absolute bias in FGEOS. Note that FGEOS is linearly interpolated to match the spatial resolution of FGARD and NLDAS-2. One of the more noticeable features in this figure is the dominance of large FGEOS absolute bias in temperature when compared with FGARD, particularly in the mountains and the Great Lakes region. In each month, there are select regions where FGARD has a larger absolute bias than FGEOS, but these regions are in the minority and the difference in biases is of smaller magnitude.

Fig. 3.
Fig. 3.

Long-term mean (2003–17) difference between FGARD and FGEOS in absolute bias of 2-m temperature (K) for May, June, and July. Bias is calculated for each dataset relative to NLDAS-2. The inset table shows region averages.

Citation: Journal of Hydrometeorology 22, 8; 10.1175/JHM-D-20-0259.1

Figure 4 is similar to Fig. 3 but for long-term averaged total precipitation (values shown in mm). In contrast to temperature, absolute precipitation bias varies greatly between the two datasets. There are many regions where FGARD has a larger absolute bias, and in May FGEOS appears less biased overall. The spatial distribution of signed bias of FGEOS (not shown) shows that FGEOS generally underpredicts precipitation in the Central and Southern U.S., while overpredicting elsewhere. However, these regions of over/underpredicting the observed value are not necessarily distributed similarly in FGARD, leading to a less clear result than was seen with temperature. Although there are some consistencies in absolute bias difference over time, there does not appear to be a general consensus in which one forecast has better results regionally than the other.

Fig. 4.
Fig. 4.

As in Fig. 3, but for total precipitation (mm).

Citation: Journal of Hydrometeorology 22, 8; 10.1175/JHM-D-20-0259.1

The majority of the results to be presented in this research use ensemble means for brevity, but we also include a measure for probabilistic forecasts using the individual ensemble members. The Brier score (BS) is a measure of the error in a probabilistic forecast of a binary event occurring. For our purposes, we wanted to capture the probability of a high-temperature month (≥95% percentile) and a low-precipitation month (≤5% percentile); these would be strong meteorological indicators for drought. The Brier skill score (BSS) measures the accuracy of a probabilistic forecast relative to a reference forecast (climatology). A perfectly accurate forecast has a BSS of 1, the reference forecast has a value of 0, and values lower than that (to −∞) represent forecasts that are worse than the reference.

Table 2 shows the BSS for each month and the overall forecasting period for a high-temperature month and a low-precipitation month. Boldface values indicate which forecast (FGARD or FGEOS) has a higher BSS for a given period. For the high-temperature metric, both forecasts have BSS that is slightly below zero—slightly worse than, and statistically indistinguishable from, a reference forecast. This is a disappointing result for a seasonal forecast, although domain-wide score does average across regions with better or worse performance relative to climatology. While the overall performance is poor, FGARD does show improved BSS relative to FGEOS, indicating that the downscaling method had a positive influence on forecast skill. For our low-precipitation metric, FGARD does have BSS values higher than the reference forecast (indicated by BSS > 0) for the low-precipitation event in May, June, and over the forecasting period. While these values are near zero, they show improvement over FGEOS, which does not exhibit skill higher than the reference forecast for this event.

Table 2.

Brier skill scores (BSS) for capturing the probability of a high-temperature month (≥95% percentile) and low-precipitation month (≤5% percentile) for each month and for the overall forecasting period. Boldface values indicate which forecast (FGARD or FGEOS) has a better BSS for a given time period. A BSS of 1 indicates perfect skill for a probabilistic forecast.

Table 2.

Comparing the differences in absolute bias in the two models is helpful for determining how well they can approximate the mean of the observed dataset. In forecast applications, however, predicting the anomaly is often more important than matching climatology. GARD is a particularly appealing downscaling method when considering forecasts of anomalies, given its ability to use multiple predictor variables to inform the prediction of each value (in contrast to simple disaggregation methods). This means that a GARD-informed forecast can produce anomaly time series that have systematically different variability from the model being downscaled, rather than simply scaling the model anomalies to match the statistical distribution of the local record.

Figure 5 shows the long-term mean FGARDFGEOS difference in normalized root-mean-square error (NRMSE) of 2-m temperature anomaly relative to NLDAS-2. Root-mean-square error (RMSE) is a measure of the spread in the prediction errors (accuracy) when compared with the observational dataset, while NRMSE is RMSE divided by the mean of the observed dataset. We choose to express this value as a percentage by multiplying by 100, so in this context it is a measure of the percentage error in the variation of temperature over time. Since Fig. 5 is the difference of this metric between the raw and downscaled datasets, positive values (redder colors) indicate regions where FGARD has a larger NRMSE and negative values (bluer colors) are indicative of larger a NRMSE in FGEOS. This figure shows that the 2-m temperature anomaly error is larger in FGARD in the first month of the forecast, and larger in FGEOS in the following 2 months. Furthermore, while there are some regions with consistently lower or higher NRMSE values comparatively, even the largest differences are no greater than ~0.5°C, showing that skill in reproducing temperature anomalies is largely similar.

Fig. 5.
Fig. 5.

Long-term mean (2003–17) difference (%) between FGARD and FGEOS in normalized root-mean-square error (NRMSE) of 2-m temperature anomaly for May, June, and July. NRMSE is calculated for each dataset relative to NLDAS-2. The inset table shows region averages.

Citation: Journal of Hydrometeorology 22, 8; 10.1175/JHM-D-20-0259.1

Likewise, Fig. 6 shows a similar metric as in Fig. 5 but for the anomaly in total precipitation. As was seen for the mean value in absolute bias in Fig. 4, the spatial distribution in precipitation is much noisier when compared with temperature; however, the regions where each dataset has favorably smaller differences are clustered. During May, FGARD has a larger NRMSE temperature anomaly throughout much of the Southern and Central United States. In contrast, FGARD has consistently smaller error throughout the Eastern and Southern U.S. in June. It is important to note that the largest NRMSE difference values shown here are an order of magnitude smaller than the typical RMSE errors found when FGEOS or FGARD is compared with NLDAS-2, so downscaling seems to have relatively little impact on this metric when compared with general forecast error. That being said, Fig. 6 shows many favorable regions for FGARD, particularly in June and July, over many Eastern and Mountain West regions.

Fig. 6.
Fig. 6.

As in Fig. 5, but for total precipitation anomaly.

Citation: Journal of Hydrometeorology 22, 8; 10.1175/JHM-D-20-0259.1

We emphasize that GARD is a downscaling framework that allows for substantial user flexibility in the choice of predictor variables and number of analogs. As such, the results obtained here are not the only results that could be obtained when applying GARD to these datasets. It is possible that better results for some regions could be obtained with more extensive testing or with inclusion of predictors from outside our set of nine surface meteorological variables. Our assessment was that these results are, however, adequate for our purposes when compared with other combinations of GARD settings, and we proceed with them for the remainder of the paper.

b. Downscaled hydrological fields

Using the downscaled forecast meteorological variables in conjunction with the CLSM within the LIS framework, we are able to generate high-resolution hydrological forecasts. The FGEOS results shown here are linearly interpolated (as previously), but also bias corrected within LIS using a standard environmental lapse-rate (ELR) adjustment to forcing fields, [as found in Cosgrove et al. (2003)] and slope-aspect correction [as found in Kumar et al. (2013)]. We examine three key hydrological variables: root zone soil moisture, surface runoff, and evapotranspiration, which are heavily influenced by many of the meteorological variables discussed previously (Kato et al. 2007).

Figure 7 shows the climatological FGARDFGEOS difference in normalized root-mean-square error of root zone soil moisture anomaly, calculated relative to a retrospective Catchment LSM simulation that uses NLDAS-2 meteorological forcing (hereinafter referred to as “retrospective”). The scale of the differences depicted here are ~1 order of magnitude less than the individual RMSE values, showing that on a climatological scale the errors are spatially similar between the two datasets (with some differences). FGARD shows lower NRMSE values in much of the Southeastern United States and desert regions during each month and becomes dominantly favorable during the third month of the forecast nationally. However, there do exist regions where the FGEOS has a smaller relative NRMSE, particularly near the Texas–Oklahoma border and parts of the upper Central U.S. As soil moisture percentiles (discussed in more depth in the next section) are a key indicator of drought, regional biases in the skill of determining long-term mean soil moisture anomalies are important in drought forecasting.

Fig. 7.
Fig. 7.

Long-term mean (2003–17) difference between FGARD and FGEOS in normalized root-mean-square error (NRMSE) of root zone soil moisture anomaly (%) for May, June, and July. NRMSE is calculated for each dataset relative to retrospective (NLDAS-2 forced LIS/CLSM). The inset table shows region averages.

Citation: Journal of Hydrometeorology 22, 8; 10.1175/JHM-D-20-0259.1

A large portion of the precipitation that does not seep into the ground and contribute to the terrestrial water storage will flow downhill along the surface as runoff. Higher-resolution data should be able to better capture these fluxes, which are integral to the water cycle. Figure 8 is similar to Fig. 7, but for the anomaly of surface runoff. When compared with root zone soil moisture, there appear to be greater localized differences in the regions showing greater NRMSE in FGEOS and FGARD; however, most of the region shows only small differences (less than or equal to 15%) between the two. In contrast, many areas (particularly in Desert regions in June, and Mountain West in July) show much lower NRMSE runoff anomalies in FGARD.

Fig. 8.
Fig. 8.

As in Fig. 7, but for surface runoff anomaly.

Citation: Journal of Hydrometeorology 22, 8; 10.1175/JHM-D-20-0259.1

Evapotranspiration (ET) is an important process to evaluate when considering drought; it is a measure of the flux of water from the surface to the atmosphere through soil moisture evaporation and plant transpiration. Being able to properly identify and predict changes in ET are important for the agricultural sector in decision making on irrigation. Furthermore, it has been reported that ET is important in detecting flash droughts (Koster et al. 2019), which can be expensive and agriculturally and ecologically devastating events. Figure 9 is similar to Figs. 7 and 8, but for the anomaly of evapotranspiration. Much of the United States shows smaller NRMSE in FGARD; however, this is not true in some regions, such as Texas and the Southern Plains. Much of the Southern and Eastern U.S. and the West have favorable values for FGARD in each of the 3 months.

Fig. 9.
Fig. 9.

As in Fig. 7, but for evapotranspiration anomaly.

Citation: Journal of Hydrometeorology 22, 8; 10.1175/JHM-D-20-0259.1

The evolution of relative performance of FGARD and FGEOS for evapotranspiration exemplifies a tendency toward relatively smaller error in FGARD in the latter half of the forecast (shown here particularly in July), which is also seen in other variables. This tendency can be explained in part by changes in long-term mean temperature (Figs. 3 and 5), precipitation (Figs. 4 and 6), and radiation fluxes (not shown). It is also likely that this reflects the dominant control of precipitation over soil moisture early in the forecast, transitioning to influence of surface energy balance terms later in the forecast, as precipitation forecast skill degrades. We will explore this phenomenon further as we examine drought indicators in the following section.

c. Monthly averaged drought indicators

Here we evaluate how well FGARD (relative to FGEOS) forecasts each of the USDM drought categories individually and cumulatively over the 2003–17 period relative to retrospective. There are a few ways to represent this, of which we will be exploring two here: monthly mean and RMSE percentage of area in drought. Using the regions shown in Fig. 1, we can examine the percentage of each region that is experiencing drought. Figure 10 shows the monthly average of this measure for each USDM drought category in each month, year, and dataset over the CONUS region. Each month/year combination in the plot includes three bars: retrospective, FGEOS, and FGARD.

Fig. 10.
Fig. 10.

Monthly average percentage of the CONUS in drought determined by root zone soil moisture percentiles for 2003–17 by month (May, June, and July). For each year and month, retrospective is the left bar, the FGEOS hindcasts are the center bar, and FGARD hindcasts are represented by the right bar. Data for 2006 and 2012 are emphasized (green outline).

Citation: Journal of Hydrometeorology 22, 8; 10.1175/JHM-D-20-0259.1

We will examine recent notable drought years specifically in the following section. It is evident from Fig. 10, however, that both FGEOS and FGARD are capable of capturing interannual variations in the percentage of area in total drought (expressed as total bar height), with some exceptions. These results reflect the combined contribution of forecast skill and, importantly, skill in initial surface conditions. We note that while not all regions experienced extreme or exceptional drought based on observations during the period of record, there is interest to examine the ability to forecast weak drought in off years. Further, there does not appear to be a general dry or wet bias in either dataset when compared with retrospective. One promising feature is how well FGARD does in correctly determining the percentage of area in total drought when FGEOS overpredicts (frequently seen in June between 2003 and 2010) and underpredicts (frequently seen in July 2010–17). Moreover, this aspect is especially true not only for the total drought percentage, but for the exceptional drought category (D4), meaning that FGARD is more capable in resolving drought severity.

There are some months for which FGARD degrades skill relative to FGEOS in predicting the percentage of area in all drought classes. These periods are often when FGEOS is not skilled in capturing drought either, but FGARD does even worse. This underscores an important tenet of our research: while downscaling can and does improve our results overall, we should not expect an unskilled forecast to be saved solely through downscaling. In addition, we have included this figure for different regions in the appendix (Figs. A1 and A2) and will address some of the regional differences later in this section.

Figure 11 shows a scatterplot of the data shown in Fig. 10 (monthly average percentage of the CONUS in drought) by drought category and month. Retrospective values are shown along the x axis and the associated forecast values are shown on the y axis for FGEOS (red) and FGARD (blue). Values along the 1:1 diagonal represent strong correlation, whereas values above (below) this line represent over (under) prediction. The coefficients of determination r2 between the forecast and retrospective values are also included on the bottom right of each panel. As lower drought severity is more common, the top-right section of each panel commonly has lower drought categories, while the bottom-left section has more severe drought categories. The spread of the data in both forecasts is low in May and becomes increasingly wider (chiefly in lower drought categories) at longer lead times. Overall, FGARD shows stronger correlations with observation when compared with FGEOS (although both are high), particularly in the most severe (D3 and D4) drought categories.

Fig. 11.
Fig. 11.

Scatterplot of monthly average percentage of the CONUS in drought determined by root zone soil moisture percentiles for 2003–17 by drought category and month. Retrospective values are shown along the x axis, and the associated forecast values are shown on the y axis for FGEOS (red) and FGARD (blue). Values along the 1:1 diagonal represent strong correlation, and overlaid text shows coefficients of determination r2 for each forecast.

Citation: Journal of Hydrometeorology 22, 8; 10.1175/JHM-D-20-0259.1

Table 3 shows r2 for monthly average percentage of regions in drought determined by root zone soil moisture percentiles for 2003–17 by category and month. Boldface values indicate which forecast (FGARD or FGEOS) has a larger correlation (relative to retrospective) for a given region. This table summarizes the results found in Fig. 11 but includes each of the subregions of our domain. The correlations by drought category show a higher correspondence in FGARD over all regions with the exception of the Central U.S., where FGEOS is higher (although the differences between them in this region are generally small; <0.1). In addition, some regions (CONUS and Desert) have a slightly higher relative correlation in FGEOS at lower drought categories (D0 and D1). Examining the correlations by month (panel b), shows stronger correlations in FGARD in all regions and months, except for the Central U.S. in June. Overall, this table highlights the Eastern U.S. and Central U.S as respectively one of the most or least favorable regions in FGARD.

Table 3.

Coefficients of determination r2 for monthly average percentage of regions in drought determined by root zone soil moisture percentiles for 2003–17 by category and month. Boldface values indicate which forecast (FGARD or FGEOS) has a stronger correlation (relative to retrospective) for a given region.

Table 3.

We examine the regional contrasts in Fig. 12, which summarizes the results discussed previously using RMSE of the percentage of area in drought for the CONUS (top panel), Central (middle panel), and Eastern (bottom panel) U.S. regions. Here we calculate the difference (FGARDFGEOS) in RMSE from daily values for each category, forecast year, and climatological month. We have also included the cumulative drought category, which corresponds to <30% in root zone soil moisture percentiles (bottom row of each table). Blue cells (positive values) are indicative of lower FGARD error, and red cells (negative values) indicate lower FGEOS error. This depiction reinforces some of the key results of the previous figures, namely, the varying skill FGARD has in the Central versus Eastern U.S., doing exceedingly well at capturing the most severe droughts in CONUS and Eastern regions, and, more often than not, showing improvement in its ability to capture the percentage of area experiencing drought.

Fig. 12.
Fig. 12.

Difference in root-mean-square error values (FGARDFGEOS derived daily) of the percentage of area in drought by category for each forecast year and climatological month for the (top) CONUS, (middle) Central U.S., and (bottom) Eastern U.S. Units of values shown are in root zone soil moisture percentiles. Blue cells are indicative of smaller FGARD RMSE, and red cells indicate smaller FGEOS RMSE.

Citation: Journal of Hydrometeorology 22, 8; 10.1175/JHM-D-20-0259.1

Looking at the right side of the figure, we can see how well FGARD does in each climatological month. Overall, it shows improvement for each category and month with the exception of May, where for a few categories FGEOS performed better in each region. This follows with our general finding that precipitation controls the soil moisture forecast in the first month, but evapotranspiration (driven by temperature and solar radiation) controls more in months 2 and 3. Downscaling cannot help much in overcoming a poor precipitation forecast, but, because GARD is more skilled in downscaling temperature, FGARD begins to outperform FGEOS in months 2 and 3.

In addition to comparing the percentage of each region experiencing drought among the datasets, we also want to present an evaluation of the detection skills of FGARD and FGEOS. We examined a number of metrics such as probability of detection and false-alarm rate, along with threat score (also called the critical success index), which measures a forecast’s ability to detect observational features on a gridcell by gridcell basis. Figure 13 shows the threat score difference for drought severity categories by week and region. Threat score (TS) is a ratio between hits (number of grid cells correctly detecting an observed drought category) and the sum of hits, misses (number of grid cells not detecting a drought category that was observed), and false alarms (number of grid cells detecting a drought category that was not observed). The values for the threat score can range from 0 to 1, with 1 being the best score. As we are depicting the difference in TS, positive values are indicative of better FGARD performance and negative values are indicative of better FGEOS performance.

Fig. 13.
Fig. 13.

Threat score difference (FGARDFGEOS) in drought severity indices for long-term (2003–17) root zone soil moisture by week and region. Categories shown are for cumulative (all drought categories) and D3_D4 (extreme and exceptional categories only).

Citation: Journal of Hydrometeorology 22, 8; 10.1175/JHM-D-20-0259.1

The data shown in Fig. 13 are weekly averaged (rather than monthly) to better depict changes in skill as the forecast progresses. Over the CONUS, FGARD TS is better than FGEOS at the very beginning and tail end of the forecasts. The Central U.S has one of the weakest FGARD results for both cumulative and D3–D4 joint categories (values are rarely positive for FGARD), which is consistent with our previous findings. A majority of the regions have favorable FGARD scores in the last few weeks of the forecast overall, and in the Eastern U.S. FGARD almost always outperforms FGEOS for D3–D4 scores.

d. Spatial drought indicators in recent years

As mentioned previously, the USDM incorporates data from a variety of sources. One source is the groundwater and soil moisture drought indicator product derived from LIS/CLSM driven by NLDAS-2 forcing with GRACE data assimilation (GRACE-DA). This product is generated and released weekly by scientists at NASA Goddard Space Flight Center. It is directly comparable to our results (Fig. 14) because both sets of simulations use LIS/CLSM; the differences being our usage of forecast data (as opposed to NLDAS-2 meteorological inputs) for forcing and our lack of data assimilation. The main difference between the LIS/CLSM soil moisture drought indicators used here for evaluation and the GRACE-DA drought indicators is that the former are derived from an open loop (no data assimilation) LIS/CLSM simulation. While we have run similar experiments using data assimilation, we have elected not to include that in this study, as Getirana et al. (2020) have already evaluated the effects of GRACE data assimilation on drought/wetness forecasting, and it would complicate the analysis of our results.

Fig. 14.
Fig. 14.

Spatial drought indicators (root zone soil moisture percentiles) for retrospective, FGEOS, and FGARD datasets. Plots are shown for 30, 60, and 90 days after initial 1 May forecast for 2006.

Citation: Journal of Hydrometeorology 22, 8; 10.1175/JHM-D-20-0259.1

There are a number of notable U.S. summer droughts within our experiment period (2003–17), which can be used as benchmarks for our downscaled forecast drought indicators. In this section, we focus on two such events, the summer droughts of 2006 and 2012. The drought in 2006 was very costly, especially for those in the agricultural sector. Soil moisture exhaustion lead to severely damaged pastures, loss of crops, a number of wildfires, and other devastating effects (Kogan and Guo 2015). Figure 14 depicts spatial drought indicator (root zone soil moisture percentile) forecasts at 30, 60, and 90 days after the forecast initialization for FGEOS and FGARD in comparison with drought indicators derived from the retrospective (open-loop, NLDAS-2 forced LIS/CLSM) simulation for the summer of 2006. The color scale follows the USDM categories for drought but also includes the upper percentiles (typically used to diagnose floods/water rich locations).

The retrospective simulation shows a growing drought throughout much of the U.S. (particularly in the Central region), with a large number of areas experiencing exceptional drought in the North and Northwest U.S. by the end of July. At 30 days into the forecast, FGEOS has overpredicted the drought in the Eastern and parts of the Southern U.S. and under predicted drought in the Central regions. At 60 days after, the plots are largely similar with a few regional exceptions (e.g., Indiana, Ohio, and Louisiana). However, at 90 days, while FGEOS is able to capture much of the areas experiencing drought, the severities are not captured well. The FGARD seems to underpredict the drought at 30 days, but generally improves by day 60. At 90 days after initialization, FGARD also captures much of the area experiencing drought, as well as the severity in parts of the northern Central U.S. Like FGEOS, FGARD predicts an overly extreme drought in Louisiana, Mississippi, and Alabama, but it corrects that problem in Florida.

The summer U.S. drought of 2012 was one of the worst droughts, in terms of both severity and extent, since the early 1930s. It is estimated that the total cost of this drought was roughly $33.9 billion (NCEI, https://www.ncdc.noaa.gov/billions/) and resulted in widespread harvest failure for a number of crops. There are a number of characteristics that led to this event, including one of the largest summer rainfall deficits ever recorded, and summer temperatures among the highest on record to date. However, one of the most notable aspects of this event was the difficulty that many climate models had in predicting it, due in part to its sudden onset (PaiMazumder and Done 2016).

Figure 15 is similar to Fig. 14 but for the year 2012. The NLDAS-2-forced LIS/CLSM plots show widespread drought throughout the U.S. at the beginning of June and increasing severity in the following two months. At 30 days after initialization, FGEOS captured some of the drought in the western and eastern United States, but overpredicted its severity; moreover, it missed the drought completely in the Central U.S. At 60 days after initialization, it is unable to capture the drought, forecasting wet conditions in some of the most severe locations. At 90 days after, it does capture much of the drought through the Central U.S. but still does not capture the severity well in many regions. Using the downscaled data, not much is improved at day 30, but at 60 days after initialization FGARD captures much more of the drought than FGEOS does (although both of them significantly underpredict extent). At 90 days after initialization, FGARD shows a higher skill in detecting the extent of the drought, while still underpredicting the severity in the Central U.S. Thus, while we can see improvement through downscaling, there is still substantial room for progress.

Fig. 15.
Fig. 15.

As in Fig. 14, but for 2012.

Citation: Journal of Hydrometeorology 22, 8; 10.1175/JHM-D-20-0259.1

5. Summary and conclusions

While seasonal forecasting has seen great advances in recent decades, many seasonal ensemble forecast products have relatively coarse resolutions. Downscaling has the potential to improve these forecasts by increasing their resolution and, in the case of GARD, by leveraging forecast information from multiple fields to inform the estimate of each variable of interest. In this application, we find that GARD downscaling provides clear advantages over raw GEOS forecasts in bias correcting local temperature. Using GARD shows mixed results in downscaling precipitation. The fact that GARD can significantly improve or degrade a forecast at regional scales stems from the fact that it attempts to inform estimates with information from multiple variables. Given the degrees of freedom available to optimize GARD for each variable, it is likely that better performance could be obtained by customizing GARD settings on a regional basis. Results might also be improved by including upper-air meteorological variables that can provide information on weather patterns. Furthermore, the regression approach in GARD might perform better with a different training dataset. Because there is no expectation for a seasonal forecast to match daily precipitation values, it is likely to revert toward climatology of the analog days selected. This approach was used here to keep the training data as consistent with the forecast data as possible.

From the perspective of seasonal forecasting, capturing anomalies is generally more important than replicating climatological mean values. For this reason, our evaluation of FGARD largely focused on skill in capturing meteorological anomalies. Results are mixed. Overall, downscaling did improve representation of meteorological anomalies, but that was not the case in all regions—especially for precipitation—or for all times, and differences between FGARD and FGEOS were generally small relative to the difference between both and the NLDAS-2 reference dataset. FGARD tended to outperform FGEOS in the Eastern U.S. and for the final month of the forecast period, but in the Central U.S. downscaling with GARD degraded some fields.

Our evaluation of the resulting drought indicators showed that FGARD increased skill relative to FGEOS in predicting drought across the CONUS and in many subregions. Moreover, FGARD generally better detected and resolved extreme (D3) and exceptional (D4) drought. This is an encouraging finding, as if it can be replicated using a wider variety of datasets and in regions outside the United States, it would suggest that downscaling meteorological forecasts can be advantageous when forecasting major droughts at S2S time scales. Performance in forecasting soil moisture drought generally followed performance in forecasting meteorology: for regions and months where the downscaled forecast improved precipitation and/or temperature estimates, it improved soil moisture drought predictions as well.

We conclude that over the CONUS and during the period of this study, downscaling with GARD improved forecasts more often and in more regions than it degraded them, and downscaling appears to have particular promise for capturing the most severe droughts. For the majority of subregions in our study, it appears to be preferable to implement a drought forecast system that uses FGARD over one that simply uses FGEOS. Further, the value of downscaling with an advanced, customizable tool like GARD can vary with location, climate, and degree of customization. At the same time, we recognize that this implementation of GARD did not improve performance consistently for all regions, and that the improvement that downscaling offers can overcome only a small portion of forecast error.

Acknowledgments

This study was funded by NASA’s Applied Sciences Program–Water Resources Award 13-WATER13-0030 and the GRACE and GRACE Follow On Science Team. The National Center for Atmospheric Research is sponsored by the National Science Foundation. The NLDAS-2 meteorological dataset was acquired as part of the activities of NASA’s Science Mission Directorate and is archived and distributed by the Goddard Earth Sciences (GES) Data and Information Services Center (DISC) (https://earthdata.nasa.gov/about/daacs/daac-ges-disc). LIS is open source (https://lis.gsfc.nasa.gov).

APPENDIX

Regional Results

Figures A1 and A2 repeat the comparative-performance analysis given in Fig. 10, but for the central and eastern U.S. subregions, respectively. We have included these regions specifically for two reasons: to evaluate how well different subregions capture the categories of drought and because these regions demonstrate interesting contrasts when examining the long-term average RMSE of temperature, precipitation, soil moisture, and the hydrologic flux anomalies (as shown in the previous figures).

The Central U.S. subregion is the largest composite region that we examined, covering the North, Central, and South Plains, as well as the Prairie regions of the United States. It shows some of the weakest improvement when compared with other regions. In Fig. A1, particularly in the third month of the forecast (July), there are some years (e.g. 2006, 2010, 2012, 2014, and 2015) for which FGARD degrades skill relative to FGEOS in predicting the cumulative drought. FGARD has noticeably lower skill in 2012 in predicting the percentage of drought area in the Central U.S. relative to other years.

In contrast, FGARD consistently outperforms FGEOS in the Eastern U.S. (Fig. A2). Its worst months are in the third months of the hindcasts for 2004, 2013, and 2015 for which it overpredicted the areas of the weaker drought categories (D0, D1, and D2). However, FGARD in almost every case accurately captures the low area percentages of the extreme (D3) and exceptional (D4) drought categories in this region, based on retrospective, whereas FGEOS overpredicts those percentages.

Fig. A1.
Fig. A1.

Monthly average percentage of the Central U.S. in drought determined by root zone soil moisture percentiles for 2003–17 by month (May, June, and July). For each year and month, retrospective is the left bar, the FGEOS hindcasts are the center bar, and FGARD hindcasts are represented by the right bar.

Citation: Journal of Hydrometeorology 22, 8; 10.1175/JHM-D-20-0259.1

Fig. A2.
Fig. A2.

As in Fig. A1, but for the Eastern U.S.

Citation: Journal of Hydrometeorology 22, 8; 10.1175/JHM-D-20-0259.1

REFERENCES

  • Arsenault, K. R., and Coauthors, 2020: The NASA hydrological forecast system for food and water security applications. Bull. Amer. Meteor. Soc., 101, E1007E1025, https://doi.org/10.1175/BAMS-D-18-0264.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Benedetti, A., and F. Vitart, 2018: Can the direct effect of aerosols improve subseasonal predictability? Mon. Wea. Rev., 146, 34813498, https://doi.org/10.1175/MWR-D-17-0282.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Berne, A., G. Delrieu, J.-D. Creutin, and C. Obled, 2004: Temporal and spatial resolution of rainfall measurements required for urban hydrology. J. Hydrol., 299, 166179, https://doi.org/10.1016/j.jhydrol.2004.08.002.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Borovikov, A., R. Cullather, R. Kovach, J. Marshak, G. Vernieres, Y. Vikhliaev, B. Zhao, and Z. Li, 2019: GEOS-5 seasonal forecast system. Climate Dyn., 53, 73357361, https://doi.org/10.1007/s00382-017-3835-2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Bukovsky, M. S., 2011: Masks for the Bukovsky regionalization of North America. Regional Integrated Sciences Collective, Institute for Mathematics Applied to Geosciences, National Center for Atmospheric Research, accessed 26 February 2021, http://www.narccap.ucar.edu/contrib/bukovsky/.

  • Clark, M. P., and A. G. Slater, 2006: Probabilistic quantitative precipitation estimation in complex terrain. J. Hydrometeor., 7, 322, https://doi.org/10.1175/JHM474.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Clark, M. P., and Coauthors, 2015: A unified approach for process-based hydrologic modeling: 1. Modeling concept. Water Resour. Res., 51, 24982514, https://doi.org/10.1002/2015WR017198.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Cosgrove, B. A., and Coauthors, 2003: Real-time and retrospective forcing in the North American Land Data Assimilation System (NLDAS) project. J. Geophys. Res., 108, 8842, https://doi.org/10.1029/2002JD003118.

    • Search Google Scholar
    • Export Citation
  • DeAngelis, A. M., H. Wang, R. D. Koster, S. D. Schubert, Y. Chang, and J. Marshak, 2020: Prediction skill of the 2012 U.S. Great Plains flash drought in Subseasonal Experiment (SubX) models. J. Climate, 33, 62296253, https://doi.org/10.1175/JCLI-D-19-0863.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Getirana, A., M. Rodell, S. Kumar, H. K. Beaudoing, K. Arsenault, B. Zaitchik, H. Save, and S. Bettadpur, 2020: GRACE improves seasonal groundwater forecast initialization over the United States. J. Hydrometeor., 21, 5971, https://doi.org/10.1175/JHM-D-19-0096.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Houborg, R., M. Rodell, B. Li, R. Reichle, and B. F. Zaitchik, 2012: Drought indicators based on model-assimilated Gravity Recovery and Climate Experiment (GRACE) terrestrial water storage observations. Water Resour. Res., 48, W07525, https://doi.org/10.1029/2011WR011291.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Jenney, A. M., K. M. Nardi, E. A. Barnes, and D. A. Randall, 2019: The seasonality and regionality of MJO impacts on North American temperature. Geophys. Res. Lett., 46, 91939202, https://doi.org/10.1029/2019GL083950.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kato, H., M. Rodell, F. Beyrich, H. Cleugh, E. V. Gorsel, H. Liu, and T. P. Meyers, 2007: Sensitivity of land surface simulations to model physics, land characteristics, and forcings, at four CEOP sites. J. Meteor. Soc. Japan, 85A, 187204, https://doi.org/10.2151/jmsj.85A.187.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kim, H., M. A. Janiga, and K. Pegion, 2019a: MJO propagation processes and mean biases in the SubX and S2S reforecasts. J. Geophys. Res. Atmos., 124, 93149331, https://doi.org/10.1029/2019JD031139.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kim, H., J. H. Richter, and Z. Martin, 2019b: Insignificant QBO-MJO prediction skill relationship in the SubX and S2S subseasonal reforecasts. J. Geophys. Res. Atmos., 124, 12 65512 666, https://doi.org/10.1029/2019JD031416.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kirtman, B. P., and Coauthors, 2014: The North American multimodel ensemble: Phase-1 seasonal-to-interannual prediction; phase-2 toward developing intraseasonal prediction. Bull. Amer. Meteor. Soc., 95, 585601, https://doi.org/10.1175/BAMS-D-12-00050.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kogan, F., and W. Guo, 2015: 2006–2015 mega-drought in the western USA and its monitoring from space data. Geomatics Nat. Hazards Risk, 6, 651668, https://doi.org/10.1080/19475705.2015.1079265.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • 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 80924 822, https://doi.org/10.1029/2000JD900327.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Koster, R. D., and Coauthors, 2004: Realistic initialization of land surface states: Impacts on subseasonal forecast skill. J. Hydrometeor., 5, 10491063, https://doi.org/10.1175/JHM-387.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Koster, R. D., S. D. Schubert, H. Wang, S. P. Mahanama, and A. M. DeAngelis, 2019: Flash drought as captured by reanalysis data: Disentangling the contributions of precipitation deficit and excess evapotranspiration. J. Hydrometeor., 20, 12411258, https://doi.org/10.1175/JHM-D-18-0242.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kumar, S. V., and Coauthors, 2006: Land information system: An interoperable framework for high resolution land surface modeling. Environ. Modell. Software, 21, 14021415, https://doi.org/10.1016/j.envsoft.2005.07.004.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kumar, S. V., C. D. Peters-Lidard, D. Mocko, and Y. Tian, 2013: Multiscale evaluation of the improvements in surface snow simulation through terrain adjustments to radiation. J. Hydrometeor., 14, 220232, https://doi.org/10.1175/JHM-D-12-046.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kumar, S. V., and Coauthors, 2016: Assimilation of gridded grace terrestrial water storage estimates in the North American land data assimilation system. J. Hydrometeor., 17, 19511972, https://doi.org/10.1175/JHM-D-15-0157.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Li, B., and Coauthors, 2019: Global GRACE data assimilation for groundwater and drought monitoring: Advances and challenges. Water Resour. Res., 55, 75647586, https://doi.org/10.1029/2018WR024618.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Luo, L., and E. F. Wood, 2008: Use of Bayesian merging techniques in a multimodel seasonal hydrologic ensemble prediction system for the eastern United States. J. Hydrometeor., 9, 866884, https://doi.org/10.1175/2008JHM980.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • McNally, A., and Coauthors, 2017: A land data assimilation system for sub-Saharan Africa food and water security applications. Sci. Data, 4, 170012, https://doi.org/10.1038/sdata.2017.12.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Molod, A., L. Takacs, M. Suarez, J. Bacmeister, I. S. Song, and A. Eichmann, 2012: The GEOS-5 atmospheric general circulation model: Mean climate and development from MERRA to Fortuna. NASA Tech. Memo. NASA/TM-2012-104606/Vol 28, 115 pp., https://ntrs.nasa.gov/citations/20120011790.

  • NCEI, 2020: Billion-dollar weather and climate disasters. Accessed 15 June 2020, https://www.ncdc.noaa.gov/billions/.

  • PaiMazumder, D., and J. M. Done, 2016: Potential predictability sources of the 2012 U.S. drought in observations and a regional model ensemble. J. Geophys. Res. Atmos., 121, 12 58112 592, https://doi.org/10.1002/2016JD025322.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Pegion, K., and Coauthors, 2019: The Subseasonal Experiment (SubX): A multimodel subseasonal prediction experiment. Bull. Amer. Meteor. Soc., 100, 20432060, https://doi.org/10.1175/BAMS-D-18-0270.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Pendergrass, A. G., and Coauthors, 2020: Flash droughts present a new challenge for subseasonal-to-seasonal prediction. Nat. Climate Change, 10, 191199, https://doi.org/10.1038/s41558-020-0709-0.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Robertson, A. W., F. Vitart, and S. J. Camargo, 2020: Subseasonal to seasonal prediction of weather to climate with application to tropical cyclones. J. Geophys. Res. Atmos., 125, e2018JD029375, https://doi.org/10.1029/2018JD029375.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Seneviratne, S. I., T. Corti, E. L. Davin, M. Hirschi, E. B. Jaeger, I. Lehner, B. Orlowsky, and A. J. Teuling, 2010: Investigating soil moisture–climate interactions in a changing climate: A review. Earth-Sci. Rev., 99, 125161, https://doi.org/10.1016/j.earscirev.2010.02.004.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • 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, 30883111, https://doi.org/10.1175/JCLI3790.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Shukla, S., A. McNally, G. Husak, and C. Funk, 2014: A seasonal agricultural drought forecast system for food-insecure regions of East Africa. Hydrol. Earth Syst. Sci., 18, 39073921, https://doi.org/10.5194/hess-18-3907-2014.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Sossa, A., B. Liebmann, I. Bladé, D. Allured, H. H. Hendon, P. Peterson, and A. Hoell, 2017: Statistical connection between the Madden–Julian oscillation and large daily precipitation events in West Africa. J. Climate, 30, 19992010, https://doi.org/10.1175/JCLI-D-16-0144.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Svoboda, M., and Coauthors, 2002: The Drought Monitor. Bull. Amer. Meteor. Soc., 83, 11811190, https://doi.org/10.1175/1520-0477-83.8.1181.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Thober, S., R. Kumar, J. Sheffield, J. Mai, D. Schäfer, and L. Samaniego, 2015: Seasonal soil moisture drought prediction over Europe using the North American Multi-Model Ensemble (NMME). J. Hydrometeor., 16, 23292344, https://doi.org/10.1175/JHM-D-15-0053.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Vitart, F., 2017: Madden-Julian oscillation prediction and teleconnections in the S2S. Quart. J. Roy. Meteor. Soc., 143, 22102220, https://doi.org/10.1002/qj.3079.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Vitart, F., and Coauthors, 2017: The Subseasonal to Seasonal (S2S) prediction project database. Bull. Amer. Meteor. Soc., 98, 163173, https://doi.org/10.1175/BAMS-D-16-0017.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wanders, N., and E. F. Wood, 2016: Improved sub-seasonal meteorological forecast skill using weighted multi-model ensemble simulations. Environ. Res. Lett., 11, 094007, https://doi.org/10.1088/1748-9326/11/9/094007.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wang, J., H. Kim, D. Kim, S. A. Henderson, C. Stan, and E. D. Maloney, 2020: MJO teleconnections over the PNA region in climate models. Part II: Impacts of the MJO and basic state. J. Climate, 33, 50815101, https://doi.org/10.1175/JCLI-D-19-0865.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wilhite, D. A., and M. D. Svoboda, 2000: Drought early warning systems in the context of drought preparedness and mitigation. In Early warning systems for drought preparedness and drought management, World Meteorological Organization Tech. Doc. AGM-2, WMO/TD 1037, 1–21, http://www.wamis.org/agm/pubs/agm2/agm02.pdf.

  • Wood, A. W., E. P. Maurer, A. Kumar, and D. P. Lettenmaier, 2002: Long-range experimental hydrological forecasting for the eastern United States. J. Geophys. Res., 107, 4429, https://doi.org/10.1029/2001JD000659.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wood, A. W., L. R. Leung, V. Sridhar, and D. P. Lettenmaier, 2004: Hydrologic implications of dynamical and statistical approaches to downscaling climate model outputs. Climatic Change, 62, 189216, https://doi.org/10.1023/B:CLIM.0000013685.99609.9e.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Xia, Y., and Coauthors, 2012: Continental-scale water and energy flux analysis and validation for the North American Land Data Assimilation System project phase 2 (NLDAS-2): 1. Intercomparison and application of model products. J. Geophys. Res., 117, D03109, https://doi.org/10.1029/2011JD016048.

    • Search Google Scholar
    • Export Citation
  • 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, W06507, https://doi.org/10.1029/2009WR007811.

    • Crossref
    • Search Google Scholar
    • Export Citation
Save
  • Arsenault, K. R., and Coauthors, 2020: The NASA hydrological forecast system for food and water security applications. Bull. Amer. Meteor. Soc., 101, E1007E1025, https://doi.org/10.1175/BAMS-D-18-0264.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Benedetti, A., and F. Vitart, 2018: Can the direct effect of aerosols improve subseasonal predictability? Mon. Wea. Rev., 146, 34813498, https://doi.org/10.1175/MWR-D-17-0282.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Berne, A., G. Delrieu, J.-D. Creutin, and C. Obled, 2004: Temporal and spatial resolution of rainfall measurements required for urban hydrology. J. Hydrol., 299, 166179, https://doi.org/10.1016/j.jhydrol.2004.08.002.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Borovikov, A., R. Cullather, R. Kovach, J. Marshak, G. Vernieres, Y. Vikhliaev, B. Zhao, and Z. Li, 2019: GEOS-5 seasonal forecast system. Climate Dyn., 53, 73357361, https://doi.org/10.1007/s00382-017-3835-2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Bukovsky, M. S., 2011: Masks for the Bukovsky regionalization of North America. Regional Integrated Sciences Collective, Institute for Mathematics Applied to Geosciences, National Center for Atmospheric Research, accessed 26 February 2021, http://www.narccap.ucar.edu/contrib/bukovsky/.

  • Clark, M. P., and A. G. Slater, 2006: Probabilistic quantitative precipitation estimation in complex terrain. J. Hydrometeor., 7, 322, https://doi.org/10.1175/JHM474.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Clark, M. P., and Coauthors, 2015: A unified approach for process-based hydrologic modeling: 1. Modeling concept. Water Resour. Res., 51, 24982514, https://doi.org/10.1002/2015WR017198.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Cosgrove, B. A., and Coauthors, 2003: Real-time and retrospective forcing in the North American Land Data Assimilation System (NLDAS) project. J. Geophys. Res., 108, 8842, https://doi.org/10.1029/2002JD003118.

    • Search Google Scholar
    • Export Citation
  • DeAngelis, A. M., H. Wang, R. D. Koster, S. D. Schubert, Y. Chang, and J. Marshak, 2020: Prediction skill of the 2012 U.S. Great Plains flash drought in Subseasonal Experiment (SubX) models. J. Climate, 33, 62296253, https://doi.org/10.1175/JCLI-D-19-0863.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Getirana, A., M. Rodell, S. Kumar, H. K. Beaudoing, K. Arsenault, B. Zaitchik, H. Save, and S. Bettadpur, 2020: GRACE improves seasonal groundwater forecast initialization over the United States. J. Hydrometeor., 21, 5971, https://doi.org/10.1175/JHM-D-19-0096.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Houborg, R., M. Rodell, B. Li, R. Reichle, and B. F. Zaitchik, 2012: Drought indicators based on model-assimilated Gravity Recovery and Climate Experiment (GRACE) terrestrial water storage observations. Water Resour. Res., 48, W07525, https://doi.org/10.1029/2011WR011291.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Jenney, A. M., K. M. Nardi, E. A. Barnes, and D. A. Randall, 2019: The seasonality and regionality of MJO impacts on North American temperature. Geophys. Res. Lett., 46, 91939202, https://doi.org/10.1029/2019GL083950.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kato, H., M. Rodell, F. Beyrich, H. Cleugh, E. V. Gorsel, H. Liu, and T. P. Meyers, 2007: Sensitivity of land surface simulations to model physics, land characteristics, and forcings, at four CEOP sites. J. Meteor. Soc. Japan, 85A, 187204, https://doi.org/10.2151/jmsj.85A.187.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kim, H., M. A. Janiga, and K. Pegion, 2019a: MJO propagation processes and mean biases in the SubX and S2S reforecasts. J. Geophys. Res. Atmos., 124, 93149331, https://doi.org/10.1029/2019JD031139.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kim, H., J. H. Richter, and Z. Martin, 2019b: Insignificant QBO-MJO prediction skill relationship in the SubX and S2S subseasonal reforecasts. J. Geophys. Res. Atmos., 124, 12 65512 666, https://doi.org/10.1029/2019JD031416.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kirtman, B. P., and Coauthors, 2014: The North American multimodel ensemble: Phase-1 seasonal-to-interannual prediction; phase-2 toward developing intraseasonal prediction. Bull. Amer. Meteor. Soc., 95, 585601, https://doi.org/10.1175/BAMS-D-12-00050.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kogan, F., and W. Guo, 2015: 2006–2015 mega-drought in the western USA and its monitoring from space data. Geomatics Nat. Hazards Risk, 6, 651668, https://doi.org/10.1080/19475705.2015.1079265.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • 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 80924 822, https://doi.org/10.1029/2000JD900327.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Koster, R. D., and Coauthors, 2004: Realistic initialization of land surface states: Impacts on subseasonal forecast skill. J. Hydrometeor., 5, 10491063, https://doi.org/10.1175/JHM-387.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Koster, R. D., S. D. Schubert, H. Wang, S. P. Mahanama, and A. M. DeAngelis, 2019: Flash drought as captured by reanalysis data: Disentangling the contributions of precipitation deficit and excess evapotranspiration. J. Hydrometeor., 20, 12411258, https://doi.org/10.1175/JHM-D-18-0242.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kumar, S. V., and Coauthors, 2006: Land information system: An interoperable framework for high resolution land surface modeling. Environ. Modell. Software, 21, 14021415, https://doi.org/10.1016/j.envsoft.2005.07.004.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kumar, S. V., C. D. Peters-Lidard, D. Mocko, and Y. Tian, 2013: Multiscale evaluation of the improvements in surface snow simulation through terrain adjustments to radiation. J. Hydrometeor., 14, 220232, https://doi.org/10.1175/JHM-D-12-046.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kumar, S. V., and Coauthors, 2016: Assimilation of gridded grace terrestrial water storage estimates in the North American land data assimilation system. J. Hydrometeor., 17, 19511972, https://doi.org/10.1175/JHM-D-15-0157.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Li, B., and Coauthors, 2019: Global GRACE data assimilation for groundwater and drought monitoring: Advances and challenges. Water Resour. Res., 55, 75647586, https://doi.org/10.1029/2018WR024618.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Luo, L., and E. F. Wood, 2008: Use of Bayesian merging techniques in a multimodel seasonal hydrologic ensemble prediction system for the eastern United States. J. Hydrometeor., 9, 866884, https://doi.org/10.1175/2008JHM980.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • McNally, A., and Coauthors, 2017: A land data assimilation system for sub-Saharan Africa food and water security applications. Sci. Data, 4, 170012, https://doi.org/10.1038/sdata.2017.12.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Molod, A., L. Takacs, M. Suarez, J. Bacmeister, I. S. Song, and A. Eichmann, 2012: The GEOS-5 atmospheric general circulation model: Mean climate and development from MERRA to Fortuna. NASA Tech. Memo. NASA/TM-2012-104606/Vol 28, 115 pp., https://ntrs.nasa.gov/citations/20120011790.

  • NCEI, 2020: Billion-dollar weather and climate disasters. Accessed 15 June 2020, https://www.ncdc.noaa.gov/billions/.

  • PaiMazumder, D., and J. M. Done, 2016: Potential predictability sources of the 2012 U.S. drought in observations and a regional model ensemble. J. Geophys. Res. Atmos., 121, 12 58112 592, https://doi.org/10.1002/2016JD025322.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Pegion, K., and Coauthors, 2019: The Subseasonal Experiment (SubX): A multimodel subseasonal prediction experiment. Bull. Amer. Meteor. Soc., 100, 20432060, https://doi.org/10.1175/BAMS-D-18-0270.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Pendergrass, A. G., and Coauthors, 2020: Flash droughts present a new challenge for subseasonal-to-seasonal prediction. Nat. Climate Change, 10, 191199, https://doi.org/10.1038/s41558-020-0709-0.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Robertson, A. W., F. Vitart, and S. J. Camargo, 2020: Subseasonal to seasonal prediction of weather to climate with application to tropical cyclones. J. Geophys. Res. Atmos., 125, e2018JD029375, https://doi.org/10.1029/2018JD029375.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Seneviratne, S. I., T. Corti, E. L. Davin, M. Hirschi, E. B. Jaeger, I. Lehner, B. Orlowsky, and A. J. Teuling, 2010: Investigating soil moisture–climate interactions in a changing climate: A review. Earth-Sci. Rev., 99, 125161, https://doi.org/10.1016/j.earscirev.2010.02.004.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • 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, 30883111, https://doi.org/10.1175/JCLI3790.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Shukla, S., A. McNally, G. Husak, and C. Funk, 2014: A seasonal agricultural drought forecast system for food-insecure regions of East Africa. Hydrol. Earth Syst. Sci., 18, 39073921, https://doi.org/10.5194/hess-18-3907-2014.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Sossa, A., B. Liebmann, I. Bladé, D. Allured, H. H. Hendon, P. Peterson, and A. Hoell, 2017: Statistical connection between the Madden–Julian oscillation and large daily precipitation events in West Africa. J. Climate, 30, 19992010, https://doi.org/10.1175/JCLI-D-16-0144.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Svoboda, M., and Coauthors, 2002: The Drought Monitor. Bull. Amer. Meteor. Soc., 83, 11811190, https://doi.org/10.1175/1520-0477-83.8.1181.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Thober, S., R. Kumar, J. Sheffield, J. Mai, D. Schäfer, and L. Samaniego, 2015: Seasonal soil moisture drought prediction over Europe using the North American Multi-Model Ensemble (NMME). J. Hydrometeor., 16, 23292344, https://doi.org/10.1175/JHM-D-15-0053.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Vitart, F., 2017: Madden-Julian oscillation prediction and teleconnections in the S2S. Quart. J. Roy. Meteor. Soc., 143, 22102220, https://doi.org/10.1002/qj.3079.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Vitart, F., and Coauthors, 2017: The Subseasonal to Seasonal (S2S) prediction project database. Bull. Amer. Meteor. Soc., 98, 163173, https://doi.org/10.1175/BAMS-D-16-0017.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wanders, N., and E. F. Wood, 2016: Improved sub-seasonal meteorological forecast skill using weighted multi-model ensemble simulations. Environ. Res. Lett., 11, 094007, https://doi.org/10.1088/1748-9326/11/9/094007.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wang, J., H. Kim, D. Kim, S. A. Henderson, C. Stan, and E. D. Maloney, 2020: MJO teleconnections over the PNA region in climate models. Part II: Impacts of the MJO and basic state. J. Climate, 33, 50815101, https://doi.org/10.1175/JCLI-D-19-0865.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wilhite, D. A., and M. D. Svoboda, 2000: Drought early warning systems in the context of drought preparedness and mitigation. In Early warning systems for drought preparedness and drought management, World Meteorological Organization Tech. Doc. AGM-2, WMO/TD 1037, 1–21, http://www.wamis.org/agm/pubs/agm2/agm02.pdf.

  • Wood, A. W., E. P. Maurer, A. Kumar, and D. P. Lettenmaier, 2002: Long-range experimental hydrological forecasting for the eastern United States. J. Geophys. Res., 107, 4429, https://doi.org/10.1029/2001JD000659.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Wood, A. W., L. R. Leung, V. Sridhar, and D. P. Lettenmaier, 2004: Hydrologic implications of dynamical and statistical approaches to downscaling climate model outputs. Climatic Change, 62, 189216, https://doi.org/10.1023/B:CLIM.0000013685.99609.9e.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Xia, Y., and Coauthors, 2012: Continental-scale water and energy flux analysis and validation for the North American Land Data Assimilation System project phase 2 (NLDAS-2): 1. Intercomparison and application of model products. J. Geophys. Res., 117, D03109, https://doi.org/10.1029/2011JD016048.

    • Search Google Scholar
    • Export Citation
  • 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, W06507, https://doi.org/10.1029/2009WR007811.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Fig. 1.

    Composite Bukovsky regions used to examine changes among similar ecological regions.

  • Fig. 2.

    Monthly averaged (left) 2-m temperature (K) and (right) total precipitation(mm) from (top) NLDAS, (middle) GEOS, and (bottom) GARD for an example month (May 2004).

  • Fig. 3.

    Long-term mean (2003–17) difference between FGARD and FGEOS in absolute bias of 2-m temperature (K) for May, June, and July. Bias is calculated for each dataset relative to NLDAS-2. The inset table shows region averages.

  • Fig. 4.

    As in Fig. 3, but for total precipitation (mm).

  • Fig. 5.

    Long-term mean (2003–17) difference (%) between FGARD and FGEOS in normalized root-mean-square error (NRMSE) of 2-m temperature anomaly for May, June, and July. NRMSE is calculated for each dataset relative to NLDAS-2. The inset table shows region averages.

  • Fig. 6.

    As in Fig. 5, but for total precipitation anomaly.

  • Fig. 7.

    Long-term mean (2003–17) difference between FGARD and FGEOS in normalized root-mean-square error (NRMSE) of root zone soil moisture anomaly (%) for May, June, and July. NRMSE is calculated for each dataset relative to retrospective (NLDAS-2 forced LIS/CLSM). The inset table shows region averages.

  • Fig. 8.

    As in Fig. 7, but for surface runoff anomaly.

  • Fig. 9.

    As in Fig. 7, but for evapotranspiration anomaly.