Journal of Geophysical Research: Atmospheres

RESEARCH ARTICLE Forecasting Rapid Drought Intensification Using the Climate

10.1029/2018JD028880
Forecast System (CFS)
Key Points: D. J. Lorenz1 , J. A. Otkin2 , M. Svoboda3 , C. R. Hain4 , and Y. Zhong2
• We develop a statistical, probabilistic
flash drought forecasting method 1
Center for Climatic Research, University of Wisconsin-Madison, Madison, WI, USA, 2Space Sciences and Engineering Center,
using recent observations and
subseasonal climate model forecasts Cooperative Institute for Meteorological Satellite Studies, University of Wisconsin-Madison, Madison, WI, USA, 3National
• Due to modest climate model Drought Mitigation Center, University of Nebraska-Lincoln, Lincoln, NE, USA, 4Earth Science Branch, NASA Marshall Space
forecast skill, recent observations Flight Center, Huntsville, AL, USA
dominate the overall skill of the
forecasting method
• Perfect model experiments show Abstract In this study, a statistical method is developed to generate probabilistic forecasts of U.S. Drought
that short forecast lead times are
most important for the upper Monitor (USDM)-depicted drought intensification over two-, four-, and six-week time periods using recent
Midwest and western United States observations and forecast model output from the Climate Forecasting System (CFS). The predictors used
include weekly anomalies in precipitation, potential evapotranspiration, dew point depression, and soil
moisture computed over different time lags. A comparison between the baseline skill obtained using recent
Correspondence to: observations only and the skill obtained by adding CFS forecast fields as predictors shows that the inclusion of
D. J. Lorenz, CFS model output leads to only a very modest increase in skill (about 14% increase in variance explained
[email protected]
over the central and eastern United States). An analysis of this result reveals that the small increase in skill is
due to limited skill in the CFS forecasts themselves, rather than to a time delay in the USDM response to
Citation: conditions on the ground. Perfect model experiments also show that not all forecast lead times are equally
Lorenz, D. J., Otkin, J. A., Svoboda, M.,
important. For example, in the upper Midwest and western United States, the first two weeks account for at
Hain, C. R., & Zhong, Y. (2018).
Forecasting rapid drought least two thirds of the total realizable skill for a four-week forecast.
intensification using the Climate
Forecast System (CFS). Journal of Plain Language Summary Among the most damaging droughts are those that develop very
Geophysical Research: Atmospheres, 123, rapidly because they provide less time to prepare or make decisions. In this study, we develop a
8365–8373. https://doi.org/10.1029/
methodology to forecast these rapidly evolving flash droughts using information from a combination of
2018JD028880
recent weather observations and seasonal climate model forecasts.
Received 1 MAY 2018
Accepted 6 JUL 2018
Accepted article online 14 JUL 2018 1. Introduction
Published online 16 AUG 2018
Drought is a natural form of climate variability that has destructive impacts on agriculture (Hatfield et al.,
2011; Mallya et al., 2013; Zhang et al., 2014), water resources (Shukla et al., 2015; van Dijk et al., 2013), human
health (Pappagianis, 1994; Park et al., 2005; Stanke et al., 2013), and natural ecosystems (Bond et al., 2008;
Bréda et al., 2006; Humphries & Baldwin, 2003; van Dijk et al., 2013). Because of the wide range of different
impacts, precisely defining drought has proven to be difficult (Hayes et al., 2011; Vicente-Serrano et al.,
2012). In the United States, one of the most comprehensive drought indices is the U.S. Drought Monitor
(USDM; Svoboda et al., 2002). The USDM is created each week through expert synthesis of numerous data
sources, including surface streamflow, soil moisture, rainfall anomalies, temperature anomalies, and crop
and range conditions. In the USDM, each location is classified into one of six categories ranging from wet
to dry: no drought, abnormally dry (D0), and moderate (D1), severe (D2), extreme (D3), and exceptional
(D4) droughts.
Droughts have a large range of intensities, durations, and rates of evolution. Among the most damaging
droughts are those that develop very rapidly (e.g., Ford & Labosier, 2017; Hunt et al., 2014; Mozny et al.,
2012; Otkin et al., 2013, 2018; Svoboda et al., 2002) because they provide less time for stakeholders to imple-
ment proactive measures in a timely manner. Such flash droughts have traditionally received less attention in
drought forecasts, which tend to focus on seasonal time scales. Hence, there is a pressing need for drought
warning systems on subseasonal time scales with frequent updates. A promising approach to subseasonal
drought forecasting is the use of short-term temporal tendencies in drought indices to identify regions with
increased likelihood of drought. For example, Otkin et al. (2013, 2015) produced skillful USDM forecasts using a
newly proposed Rapid Change Index, which is based on the temporal tendency of the Evaporative Stress Index
©2018. American Geophysical Union. (ESI; Anderson et al., 1997; Anderson, Kustas, & Norman, 2007; Anderson et al., 2011). Similarly, Ford et al. (2015)
All Rights Reserved. show that anomalies in observed soil moisture often precede the development of drought in Oklahoma.

LORENZ ET AL. 8365

 21698996, 2018, 16, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2018JD028880 by Morocco Hinari NPL, Wiley Online Library on [20/09/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Atmospheres 10.1029/2018JD028880

Recently, Lorenz et al. (2017a, 2017b) developed a probabilistic statistical forecasting methodology that uses
recent anomalies in precipitation, ESI, and modeled soil moisture to predict the probability of USDM inten-
sification occurring over subseasonal time scales. In this paper, we extend the results of Lorenz et al. (2017a,
2017b) by including forecasts of drought-related fields from the North American Multi-Model Ensemble
(NMME; Kirtman et al., 2014). The NMME has been used to forecast future drought (Yuan and Wood,
2013; Mo and Lyon, 2015; Thober et al., 2015); however, this guidance has typically been applied to seasonal
time scales. In this paper, the Climate Forecast System version 2 (CFSv2) component of the NMME will be
used to predict drought intensification at two-, four-, and six-week time scales through an extension of
the statistical methodology of Lorenz et al. (2017a, 2017b). One advantage of this methodology is that
the current and past observations can be combined with CFS forecasts into a single USDM intensification
probability that takes into account both sources of predictability. We begin by discussing the statistical
methodology, and the data and model inputs. Second, baseline USDM predictions using recent observa-
tions only are compared with predictions that also include fields from the CFS. Next, we analyze the source
of skill and the potential for further improvement. Finally, we summarize the results and discuss the highest
priority directions for future work that will likely have the most positive impact on forecasting changes in the
USDM or other drought metrics.

2. Methodology
A detailed description of the statistical methodology to predict USDM-depicted drought intensification is
given in Lorenz et al. (2017a, 2017b). In this paper, forecast fields from the CFS are used as predictors to
improve the accuracy of the USDM predictions in Lorenz et al. (2017b). A summary of the most relevant parts
of the methodology are given here.
Because the USDM drought depiction is represented by discrete values, a straightforward methodology such
as standard linear regression is not appropriate. Therefore, we use logistic regression to predict a binary vari-
able that takes the value 1 if the USDM is more intense in n weeks and 0 otherwise. In this paper, n is two, four,
or six weeks. The predictors for the logistic regression include weekly composites of precipitation, potential
evapotranspiration (PET), dew point depression, and topsoil moisture (0–10 cm). All predictors are standar-
dized prior to the logistic regression. Instead of using only one predictor for each variable, we implement a
range of different weekly time lags as separate predictors. For example, in Lorenz et al. (2017b), precipitation
accumulation over the most recent week as well as weekly composites for one, two, and three weeks prior is
used as separate individual predictors. In this paper, this same set of current and past weekly composites is
used (i.e., weeks 0, 1, 2 and 3, where negative corresponds to the past); however, we now add additional
weekly composites of future conditions as forecasted by the CFS. For an n week USDM forecast, we use n
additional predictors per variable: one for each week from 1 to n in the future. For example, for the two-week
USDM forecasts, each variable has six total predictor time lags for 3, 2, 1, 0, 1, and 2 weeks relative to the
present time. This scheme allows the data to choose the optimal weighting in time, or equivalently, the data
chooses the optimal time scale for averaging each predictor.
In this study, as in Lorenz et al. (2017b), standard logistic regression is modified so that the regression coeffi-
cients multiplying the predictors are constrained to be nonnegative using Non-Negative Logistic Regression
(NNLR). The advantage of NNLR is found in its regularization properties (Meinshausen, 2013; Slawski & Hein,
2013), which penalize excessive complexity (i.e., large positive and negative coefficients whose effects mostly
cancel), and therefore, NNLR almost always shows better skill on independent data compared to standard lin-
ear regression. Moreover, NNLR is easy to apply for drought prediction because the sign of the true, physically
based coefficients is known a priori. Note that if the physically based coefficient is negative, simply multiply
the predictor by 1. With many predictors, NNLR regression will result in nonzero weights for a subset of the
predictors and the rest of the weights will be exactly zero. Hence, NNLR is able to discard unimportant or
redundant variables. See Lorenz et al. (2017a, 2017b) for more details.
A key finding of Lorenz et al. (2017b) is that including information on how far the current USDM state is from
next higher or lower drought category helps improve skill significantly. This USDM state variable is used as an
additional predictor for the logistic regression in this paper as well. The USDM state variable is calculated in
exactly the same way as in Lorenz et al. (2017b) except it is extended through the year 2016. We do not dis-
cuss this aspect of the methodology further because it is independent of the use of CFS forecasts, which is the

LORENZ ET AL. 8366

 21698996, 2018, 16, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2018JD028880 by Morocco Hinari NPL, Wiley Online Library on [20/09/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Atmospheres 10.1029/2018JD028880

subject of this paper. Also, we use the (cross-validated) climatological intensification probability as a predictor
as well (Lorenz et al., 2017b). This predictor is useful if the USDM is more likely to intensify during a particular
time of year compared to other times. All skill scores shown below are cross validated: we incrementally leave
one year out of the fit and then test predictions on that year.

3. Observed and CFS Forecast Data

The weekly USDM classifies conditions into six categories based on expert synthesis of numerous data
sources (Svoboda et al., 2002). The six categories from wettest to driest are no drought, abnormally dry
(D0), moderate drought (D1), severe drought (D2), extreme drought (D3), and exceptional drought (D4)
and are based on a ranking percentile classification scheme.
The predictors in this study have been modified from Lorenz et al. (2017b): the predictor variables are
restricted to those available as CFS NMME output. For example, the ESI (Anderson, Norman, et al., 2007) is
a satellite-based drought indicator, based on a surface energy balance model (ALEXI; Anderson et al., 1997)
driven by changes in land surface temperature that is statistically independent to the CFS, and therefore, it
is not used (except for the USDM state predictor; see section 2). The predictors chosen for this study are
weekly mean precipitation, PET, dew point depression, and soil moisture in the top 10 cm of the soil profile.
The PET is calculated using the Penman-type algorithm in Mahrt and Ek (1984). The USDM and all predictors
are gridded to the same 0.12 × 0.12° resolution grid using the methods in Lorenz et al. (2017a).

3.1. Current and Past Data

As shown in Otkin et al. (2014, 2015) and Lorenz et al. (2017b) past anomalies in soil moisture, ET, and preci-
pitation have skill in predicting future USDM intensification. Therefore, we continue to use current and past
anomalies in these variables while also adding CFS-derived predictors of future conditions. Precipitation data
are obtained from the Climate Prediction Center’s gridded analysis of daily precipitation reports from
National Weather Service reporting stations and cooperative observers (Higgins et al. 2000). Soil Moisture
(total water mass per area from 0- to 10-cm depth) is taken from the North American Land Data
Assimilation System (NLDAS; Mitchell et al., 2004; Xia et al., 2012). The soil moisture is the ensemble average
value over three NLDAS models: (1) NOAH, (2) MOSAIC, and (3) VIC because this has been shown to more
accurately depict drought conditions (Xia et al., 2014). The top-soil moisture is used instead of the total col-
umn soil moisture because we found it is more highly correlated with USDM intensification (not shown). PET
and dew point depression are taken from the CFS Reanalysis (Saha et al., 2010).

3.2. CFS Forecasts

In this study, the CFS is used because more variables are available in the CFS than other NMME models. Using
lagged correlation analysis between the USDM time tendency and various predictor fields, we found that soil
moisture, PET, and dew point depression were much more highly correlated with USDM change than the
more widely available air temperature fields. This is also consistent with recent studies by Otkin et al.
(2018) and Ford and Labosier (2017), both of whom found stronger relationships between drought develop-
ment and soil moisture, PET, and dew point depression. These variables are only available for the CFS and
therefore only that model is used. The CFS NMME fields are the same as those that are used for current
and past data: precipitation, PET, dew point depression, and 0–10 cm soil moisture.
From 1982 to 2011, four CFS forecasts were initialized at 0, 6, 12, and 18 UTC on every fifth day. After 1999,
four CFS NMME forecasts were made every day; however, we only use forecasts from every fifth day so that
a longer climatology can be used to help reduce model bias. The CFS forecasts are composited into weekly
mean values from Tuesday morning to Tuesday morning (Tuesday morning is the data cutoff time for the
USDM). The data from each of the four model runs produced each day are then averaged together to form
a four-member ensemble-mean. To remove any bias in the mean forecast state, CFS data from 1982 to
2016 are used to form a climatological forecast for each initialization date and forecast lead time. This clima-
tology is then subtracted from the forecasts. Next, values are normalized by the standard deviation to form
standardized forecast anomalies. The longest USDM forecast presented here is for six weeks; therefore, the
maximum CFS lead time is also six weeks plus up to four days to account for the five-day frequency of CFS
forecasts used here.

LORENZ ET AL. 8367

 21698996, 2018, 16, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2018JD028880 by Morocco Hinari NPL, Wiley Online Library on [20/09/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Atmospheres 10.1029/2018JD028880

4. Results
We begin by focusing on the four-week intensification forecasts and use
the Brier Skill Score (BSS) as our forecast skill metric. The climatological
intensification probability is used as the reference Brier score. The BSS
using only current and past data (Figure 1a) shows highest skill in the
north-central United States, which is consistent with Lorenz et al.
(2017b). Adding four weeks of CFS NMME forecast data to this baseline
forecast, however, leads to only very minor changes in USDM forecast skill
(Figure 1b). As we will show shortly, the skill does tend to increase almost
everywhere; however, the change is minor. There are two possible causes
for the small effect of CFS forecast data: (1) a large portion of the USDM
variability has a delayed response to conditions on the ground; therefore,
very recent changes (i.e., last several weeks) do not yet impact the USDM,
and (2) the CFS forecasts themselves have relatively little skill over these
time scales (four weeks). The first possible cause is consistent with the fact
that current and past information do have skill predicting future USDM
intensification (Figure 1a and Lorenz et al., 2017b and Otkin et al., 2014,
2015), but to test this more rigorously a more quantitative approach is
needed. To this end, we will test the second possible cause by substituting
future observations for CFS forecasts in the prediction scheme. We call this
experiment the perfect CFS forecast experiment. The BSS for the perfect
CFS forecast (Figure 1c) is substantially larger than the actual CFS-based
forecasts (Figure 1b); therefore, it appears that the USDM does respond
in real time to conditions on the ground and that poor CFS forecast accu-
racy is the primary reason for the small improvements in BSS going from
Figures 1a and 1b.

Because BSS is a squared measure analogous to the fraction variance

explained, it is useful to average the BSS spatially to get a general,
smoothed sense of the effect of the CFS forecasts. Averaged over the
domain east of 105°W longitude, the BSS for the no CFS, CFS, and perfect
CFS are 10.7%, 12.1%, and 37.7%, respectively. The CFS forecasts lead to a
modest increase in variance explained compared to the baseline (from
10.7% to 12.1%). With a perfect CFS forecast, however, the fraction of var-
iance explained more than triples. Therefore, the majority of the USDM
variance is responding in real time as far as these four-week forecasts
are concerned. Therefore, it is assumed that the limited CFS forecast accu-
racy over subseasonal time scales is the primary cause for the small
improvements, although there is some portion of USDM variability that
Figure 1. (a) Brier Skill Score (BSS) for the baseline four-week USDM intensi- is determined by current and past conditions.
fication forecasts using current and past data only (no CFS forecasts). (b) As in
(a) but with CFS forecasts fields added as predictors. (c) As in (b) but withExpanding the analysis to include the two- and six-week predictions,
Figure 2 shows the change in BSS going from no-CFS (i.e., baseline) fore-
future observations substituted for CFS forecasts. This panel shows potential
predictability possible from a perfect CFS forecast. casts to CFS-based forecasts (Figures 2a–2c) and the change in BSS going
from the CFS-based forecasts to the perfect CFS forecasts (Figures 2d–2f).
The top, middle, and bottom panels show the results for the two-, four-, and six-week forecasts, respectively.
Hence, Figure 2b is the difference between Figures 1b and 1a, and Figure 2e is the difference between
Figures 1c and 1b.

The effect of the CFS forecasts on the USDM skill is small but consistently positive (Figures 2a–2c). The
improvement is largest and most robust for the four-week predictions and least consistently positive for
the six-week predictions. The two-week improvements are almost as much as the four-week improvements
but nevertheless tend to be slightly smaller. The fact that the six-week predictions tend to be less accurate
than the 4-week forecasts on average indicates that the CFS forecast skill degrades at longer lead times.
The two-week predictions, however, do not follow this rule. We believe that the two-week forecasts are

LORENZ ET AL. 8368

 21698996, 2018, 16, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2018JD028880 by Morocco Hinari NPL, Wiley Online Library on [20/09/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Atmospheres 10.1029/2018JD028880

Figure 2. (a) The change in BSS going from the baseline (i.e. no CFS) two-week forecasts to the CFS-based forecasts. (b) As
in (a) but for the four-week forecasts. (c) As in (a) but for the 6 week forecasts. (d) The change in BSS going from the
CFS-based two-week forecasts to the perfect CFS-based (i.e., future observations) forecasts. (e) As in (d) but for the four-
week forecasts. (f) As in (d) but for the six-week forecasts.

less accurate than the four-week forecasts because the precise timing of the intensification, in both the
USDM and the CFS forecasts, matters more for the two-week predictions. For example, for the four-week
forecast, the USDM can intensify during any of the following four weeks. The two-week forecasts, on the
other hand, have a smaller time window where the intensification can occur and so the precise timing is
more important.
The effect of perfect CFS forecasts (e.g., with actual observations used as the predictors) on USDM intensifica-
tion forecasts in shown in Figures 2d–2f. As expected from Figure 1, the improvements in BSS are much larger
here than for the actual CFS forecasts. In this case, the improvement increases as the forecast lead time
increases from two to six weeks. In this case, the CFS skill degradation at longer forecast lead times is not
an issue. Instead, the timing issue described in the previous paragraph implies that the USDM is easier to fore-
cast at longer lead times because less precision in the timing of the USDM intensification is necessary.
The relative weight of the different predictor variables and predictor lags is shown in Figure 3. Because all
variables are standardized before the logistic regression, the relative size of the regression coefficients (i.e.,
weights) quantifies the relative importance of the different predictors and time lags. To reduce noise, the
regression coefficients are averaged over the domain east of 105°W (longitude of Denver, CO). Because of

LORENZ ET AL. 8369

 21698996, 2018, 16, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2018JD028880 by Morocco Hinari NPL, Wiley Online Library on [20/09/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Atmospheres 10.1029/2018JD028880

the nonnegativity constraint in the logistic regression (see section 2), there
is no possibility of large negative coefficients canceling large positive coef-
ficients in the area average.
For the CFS-based forecasts, precipitation has the largest weight by far;
however, some of this is due to the fact that the PET, dew point depression,
and soil moisture are more strongly correlated with each other than they
are to precipitation. In other words, the PET, dew point depression, and soil
moisture variables are somewhat redundant, and therefore, the aggregate
weight of these three variables is perhaps a more relevant measure of their
importance. The precipitation weight is largest at zero lag, which means
that the current precipitation is a better predictor of USDM intensification
than the CFS forecast precipitation. The weight for the one-week forecast
(+1) has about the same weight as one-week previous ( 1). Beyond two
weeks, there is almost no precipitation weight presumably because the
skill of precipitation forecasts in the CFS degrades substantially at longer
forecast times. The PET, dew point depression, and soil moisture weights
are largest from lag 0 to lag +1, and all these predictors have almost no
weight from past lags ( 3 to 1). The fact that the peak weight is shifted
toward positive lags compared to the precipitation suggests that CFS fore-
casts of these variables are more skillful, which is consistent with McEvoy
et al. (2016).

For the perfect CFS forecasts, the weights increase because USDM
forecast skill improves and the weights shift toward positive lags. For
precipitation, the peak weight is at lag +1 and the current week precipita-
tion (lag 0) is smaller than all future weights except possibly lag +4. Unlike
precipitation, PET, dew point depression, and soil moisture have essen-
tially zero weight for current and past time lags. The soil moisture weight
Figure 3. (a) Regression coefficients (i.e., predictor weights) averaged over
peaks at two weeks, and the PET and dew point depression have
the domain east of 105°W for the four-week CFS-based intensification uniformly small weights over all positive time lags. Also interesting is
forecasts. All predictors are standardized prior to the analysis, and therefore, the fact that the small weights for PET and dew point depression suggest
the relative size of the weights tracks the relative importance of the that they are unimportant once the future soil moisture is known with pre-
predictor. Positive lags are the weights for future CFS forecasted fields and cision. The weights for the two and six-week forecasts are similar to the
negative/zero lags are for past and current observations. (b) As in (a) but for
the perfect CFS-based (i.e., future observations) forecasts.
four-week forecasts (not shown).

The spatial structure of the regression weights is shown in Figure 4. To help

summarize the results, the weights are summed over time lag for each pre-
dictor. Next, we add the weights of the nonprecipitation variables together (dew point depression, PET, and
soil moisture) to better compare their aggregate importance to the more dominant precipitation. For the CFS
forecasts, the precipitation dominates in the Midwest and is also very important through much of the central
and eastern United States (Figure 4a). The other three variables are most important in the western United
States where together they sometimes have more weight than precipitation (for example, the four corners
region; Figure 4b). For the perfect model experiments, the precipitation is by far the most dominant predictor
east of 100°W (Figure 4c). It is only in the western United States that the other predictors become comparable
to or more important than precipitation.

The weights in Figure 3b suggest that the most important forecast times are lags +1 and +2 weeks, which is
encouraging because two-week CFS skill is much more attainable than that at longer lead times. To quantify
the effect of perfect one-week skill on the four-week USDM forecasts, we ran an additional experiment where
we include only one week of future observations and no information at greater time lags (weeks 2, 3, and 4).
We also repeated this experiment but for perfect two-week skill. The change in BSS relative to the CFS based
forecasts is shown in Figures 5a and 5b. These maps should be compared with the change in BSS using the
full four-week of future observations (Figure 2e). For the eastern United States the four-week perfect is sub-
stantially bigger than the one- or two-week perfect; however, as one moves west the difference becomes less
big. To better compare these experiments, the ratio between the one- and four-week experiments is shown

LORENZ ET AL. 8370

 21698996, 2018, 16, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2018JD028880 by Morocco Hinari NPL, Wiley Online Library on [20/09/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Atmospheres 10.1029/2018JD028880

Figure 4. (a) Precipitation weights summed over all time lags for the four-week CFS forecasts. (b) As in (a) but for the sum of
dew point depression, PET, and soil moisture. (c and d) Same as (a) and (b) but for the perfect model forecasts.

Figure 5. (a) As in Figure 2e but for only one week of future observations (i.e., no predictors for two, three, and four weeks
into future) instead of four weeks of future observations. (b) As in (a) but for two weeks of future observations. (c) The ratio
of impact of one week of future observations (a) to the impact of four weeks of future observations (Figure 2e) given in
percent. The BSS has been smoothed in order help focus on the large-scale features. (d) Same as (c) but for the ratio of two
to four weeks of future observations.

LORENZ ET AL. 8371

 21698996, 2018, 16, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2018JD028880 by Morocco Hinari NPL, Wiley Online Library on [20/09/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Atmospheres 10.1029/2018JD028880

in Figure 5c. To help show the important large-scale features, the fields are smoothed in space prior to com-
puting the ratio. If all future time lags are equally important, then one expects that the ratio of the one- to the
four-week experiment will be 25%, which is denoted by the absence of shading in the figure. For the northern
and western United States the ratio is larger than 25% suggesting that the first week carries more weight than
later weeks for these regions. Similarly, Figure 5d shows the ratio of the two- to the four-week experiment. In
this case a ratio of 50% implies equal importance, and therefore, ratios near this value have no shading in the
figure. In this case, the first two weeks carry more weight than the later times over an even broader region of
the United States. Focusing on the northern plains where skill is very good (Figure 1c), one sees that the first
two-weeks account for over two thirds of the total realizable skill (Figure 5d). The larger than expected role of
the one- and two-week forecasts suggests that substantial drought forecasting skill is attainable through
improved precipitation forecasts in the CFS.

5. Conclusions
This paper explores the use of CFS NMME forecasts to improve statistical forecasts of USDM intensification at
two, four, and six weeks in the future. First, the baseline USDM forecast skill using current and past anomalies
in precipitation, PET, dew point depression, and soil moisture is computed. Next, future CFS forecasts of the
above predictors are added to the statistical forecasting scheme. While adding CFS forecast data improves
the drought forecasting skill for much the central and eastern United States, the improvements are very mod-
est (about a 14% increase in variance explained on average). Next, we explored whether the modest increase
in USDM skill is due to the modest skill of the CFS forecasts themselves or whether the USDM forecast skill is
more-or-less already determined by current and past conditions because the USDM has a delayed response
to conditions on the ground. In an experiment using future observations rather than CFS forecasts as predic-
tors, it is found that USDM skill more than triples, which suggests that the CFS forecasts are responsible for
the small improvement in skill. Further experiments show that the first two weeks of the CFS forecast have
a larger than expected impact on the four-week USDM forecasts (for the northern and western United
States), suggesting that significant improvement in drought forecasting is attainable.
The relative importance of short-term (i.e., one- and two-week) long-term information even in a perfect fore-
cast scenario suggests that perhaps the most fruitful monthly USDM forecasts would be based on standard
long-range weather forecasting models rather than using output from climate models. Weather forecasting
models have the added benefit of higher resolution and more ensemble members. In addition, the above
results suggest that the USDM forecast skill presented here is limited by the five-day interval between succes-
sive CFS forecasts used here. This means that the latest initialization time is typically not optimal given the
weekly interval of the USDM. For example, sometimes the one-week composite forecast used here is actually
based on a 5- to 12-day CFS forecast. Repeating the analysis including more recent daily CFS reforecasts or,
alternatively, using the Global Ensemble Forecast System (GEFS) might yield substantial improvements.

Acknowledgments References
This work was supported by funds
provided by the NOAA Climate Program Anderson, M. C., Hain, C. R., Wardlow, B., Mecikalski, J. R., & Kustas, W. P. (2011). Evaluation of drought indices based on thermal remote
Office’s Modeling, Analysis, Predictions, sensing of evapotranspiration over the continental U.S. Journal of Climate, 24, 2025–2044.
and Projections (MAPP) program under Anderson, M. C., Kustas, W. P., & Norman, J. M. (2007). Upscaling flux observations from local to continental scales using thermal remote
grant NA14OAR4310226. The data used sensing. Agronomy Journal, 99(1), 240–254.
are listed in the references. Anderson, M. C., Norman, J. M., Diak, G. R., Kustas, W. P., & Mecikalski, J. R. (1997). A two-source time-integrated model for estimating surface
fluxes using thermal infrared remote sensing. Remote Sensing of Environment, 60(2), 195–216. https://doi.org/10.1016/S0034-
4257(96)00215-5
Anderson, M. C., Norman, J. M., Mecikalski, J. R., Otkin, J. A., & Kustas, W. P. (2007). A climatological study of evapotranspiration and moisture
stress across the continental U.S. based on thermal remote sensing: 1. Model formulation. Journal of Geophysical Research, 112, D10117.
https://doi.org/10.1029/2006JD007506
Bond, N. R., Lake, P. S., & Arthington, A. H. (2008). The impacts of drought on freshwater ecosystems: An Australian perspective. Hydrobiologia,
600(1), 3–16. https://doi.org/10.1007/s10750-008-9326-z
Bréda, N., Huc, R., Granier, A., & Dreyer, E. (2006). Temperate forest trees and stands under severe drought: A review of ecophysiological
responses, adaptation processes and long-term consequences. Annals of Forest Science, 63(6), 625–644. https://doi.org/10.1051/
forest:2006042
Ford, T. W., & Labosier, C. F. (2017). Meteorological conditions associated with the onset of flash drought in the eastern United States.
Agricultural and Forest Meteorology, 247, 414–423. https://doi.org/10.1016/j.agrformet.2017.08.031
Ford, T. W., McRoberts, D. B., Quiring, S. M., & Hall, R. E. (2015). On the utility of in situ soil moisture observations for flash drought early
warning in Oklahoma, USA. Geophysical Research Letters, 42, 9790–9798. https://doi.org/10.1002/2015GL066600
Hatfield, J. L., Boote, K. J., Kimball, B. A., Ziska, L. H., Izaurralde, R. C., Ort, D., et al. (2011). Climate impacts of agriculture: Implications for crop
production. Agronomy Journal, 103(2), 351–370. https://doi.org/10.2134/agronj2010.0303

LORENZ ET AL. 8372

 21698996, 2018, 16, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2018JD028880 by Morocco Hinari NPL, Wiley Online Library on [20/09/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Journal of Geophysical Research: Atmospheres 10.1029/2018JD028880

Hayes, M., Svoboda, M., Wall, N., & Wildhalm, M. (2011). The Lincoln declaration on drought indices: Universal meteorological drought index
recommended. Bulletin of the American Meteorological Society, 92(4), 485–488. https://doi.org/10.1175/2010BAMS3103.1
Higgins, R. W., Shi, W., Yarosh, E., & Joyce, R. (2000). Improved United States precipitation quality control system and analysis. NCEP/Climate
Prediction Center Atlas 7, NOAA (40 pp.). Retrieved from http://www.cpc.ncep.noaa.gov/research_papers/ncep_cpc_atlas/7/toc.html
Humphries, P., & Baldwin, D. S. (2003). Drought and aquatic ecosystems: An introduction. Freshwater Biology, 48(7), 1141–1146. https://doi.
org/10.1046/j.1365-2427.2003.01092.x
Hunt, E., Svoboda, M., Wardlow, B., Hubbard, K., Hayes, M. J., & Arkebauer, T. (2014). Monitoring the effects of rapid onset of drought on
non-irrigated maize with agronomic data and climate-based drought indices. Journal of Agricultural and Forest Meteorology, 191, 1–11.
https://doi.org/10.1016/j.agrformet.2014.02.001
Kirtman, B. P., Min, D., Infanti, J. M., Kinter, J. L. III, Paolino, D. A., Zhang, Q., et al. (2014). The North American multimodel ensemble: Phase-1
seasonalvto-interannual prediction; phase-2 toward developing intraseasonal prediction. Bulletin of the American Meteorological Society,
95(4), 585–601.
Lorenz, D. J., Otkin, J. A., Svoboda, M., Hain, C. R., Anderson, M. C., & Zhong, Y. (2017a). Predicting US Drought Monitor (USDM) states using
precipitation, soil moisture, and evapotranspiration anomalies. Part I: Development of a non-discrete USDM index. Journal of
Hydrometeorology, 18(7), 1943–1962. https://doi.org/10.1175/JHM-D-16-0066.1
Lorenz, D. J., Otkin, J. A., Svoboda, M., Hain, C. R., Anderson, M. C., & Zhong, Y. (2017b). Predicting US Drought Monitor (USDM) states using
precipitation, soil moisture, and evapotranspiration anomalies. Part II: Intraseasonal drought intensification forecasts. Journal of
Hydrometeorology, 18(7), 1963–1982. https://doi.org/10.1175/JHM-D-16-0067.1
Mahrt, L., & Ek, M. (1984). The influence of atmospheric stability on potential evaporation. Journal of Climate and Applied Meteorology, 23(2),
222–234. https://doi.org/10.1175/1520-0450(1984)023%3C0222:TIOASO%3E2.0.CO;2
Mallya, G., Tripathi, S., Kirshner, S., & Govindaraju, R. S. (2013). Probabilistic assessment of drought characteristics using hidden Markov
model. Journal of Hydrologic Engineering, 18(7), 834–845. https://doi.org/10.1061/(ASCE)HE.1943-5584.0000699
McEvoy, D. J., Huntington, J. L., Mejia, J. F., & Hobbins, M. T. (2016). Improved seasonal drought forecasts using reference evapotranspiration
anomalies. Geophysical Research Letters, 43, 377–385. https://doi.org/10.1002/2015GL067009
Meinshausen, N. (2013). Sign-constrained least squares estimation for high-dimensional regression. Electronic Journal of Statistics, 7(0),
1607–1631. https://doi.org/10.1214/13-EJS818
Mitchell, K. E., Lohmann, D., Houser, P. R., Wood, E. F., Schaake, J. C., Robock, A., et al. (2004). The multi-institution North American Land Data
Assimilation System (NLDAS): Utilizing multiple GCIP products and partners in a continental distributed hydrological modeling system.
Journal of Geophysical Research, 109, D07S90. https://doi.org/10.1029/2003JD003823
Mo, K. C., & Lyon, B. (2015). Global meteorological drought prediction using the North American multi-model ensemble. Journal of
Hydrometeorology, 16, 1409–1424.
Mozny, M., Trnka, M., Zalud, Z., Hlavinka, P., Nekovar, J., Potop, V., & Virag, M. (2012). Use of a soil moisture network for drought monitoring in
the Czech Republic. Theoretical and Applied Climatology, 107(1–2), 99–111. https://doi.org/10.1007/s00704-011-0460-6
Otkin, J. A., Anderson, M. C., Hain, C., Mladenova, I., Basara, J., & Svoboda, M. (2013). Examining flash drought development using the thermal
infrared based Evaporative Stress Index. Journal of Hydrometeorology, 14(4), 1057–1074. https://doi.org/10.1175/JHM-D-12-0144.1
Otkin, J. A., Anderson, M. C., Hain, C., & Svoboda, M. (2014). Examining the relationship between drought development and rapid changes in
the Evaporative Stress Index. Journal of Hydrometeorology, 15(3), 938–956. https://doi.org/10.1175/JHM-D-13-0110.1
Otkin, J. A., Anderson, M. C., Hain, C., & Svoboda, M. (2015). Using temporal changes in drought indices to generate probabilistic drought
intensification forecasts. Journal of Hydrometeorology, 16(1), 88–105. https://doi.org/10.1175/JHM-D-14-0064.1
Otkin, J. A., Svoboda, M., Hunt, E. D., Ford, T. W., Anderson, M. C., Hain, C., & Basara, J. B. (2018). Flash droughts: A review and assessment of the
challenges imposed by rapid onset droughts in the United States. Bulletin of the American Meteorological Society, 99, 911–919. https://doi.
org/10.1175/BAMS-D-17-0149.1
Pappagianis, D. (1994). Marked increase in cases of coccidioidomycosis in California: 1991, 1992, and 1993. Clinical Infectious Diseases,
19(Supplement_1), S14–S18. https://doi.org/10.1093/clinids/19.Supplement_1.14
Park, B. J., Sigel, K., Vaz, V., Komatsu, K., McRill, C., Phelan, M., et al. (2005). An epidemic of coccidioidomycosis in Arizona associated with
climatic changes, 1998–2001. The Journal of Infectious Diseases, 191(11), 1981–1987. https://doi.org/10.1086/430092
Saha, S., Moorthi, S., Pan, H. L., Wu, X., Wang, J., Nadiga, S., et al. (2010). The NCEP Climate System Forecast Reanalysis. Bulletin of the American
Meteorological Society, 91(8), 1015–1058. https://doi.org/10.1175/2010BAMS3001.1
Shukla, S., Safeeq, M., AghaKouchak, A., Guan, K., & Funk, C. (2015). Temperature impacts on the water year 2014 drought in California.
Geophysical Research Letters, 42, 4384–4393. https://doi.org/10.1002/2015GL063666
Slawski, M., & Hein, M. (2013). Non-negative least squares for high-dimensional linear models: Consistency and sparse recovery without
regularization. Electronic Journal of Statistics, 7(0), 3004–3056. https://doi.org/10.1214/13-EJS868
Stanke, C., Kerac, M., Prudhomme, C., Medlock, J., & Murray, V. (2013). Health effects of drought: A systematic review of the evidence.
PLoS Currents, 5.
Svoboda, M., LeComte, D., Hayes, M., Heim, R., Gleason, K., Angel, J., et al. (2002). The drought monitor. Bulletin of the American Meteorological
Society, 83(8), 1181–1190. https://doi.org/10.1175/1520-0477-83.8.1181
Thober, S., Kumar, R., Sheffield, J., Mai, J., Schäfer, D., & Samaniego, L. (2015). Seasonal soil moisture drought prediction over Europe using the
North American Multi-Model Ensemble (NMME). Journal of Hydrometeorology, 16(6), 2329–2344.
van Dijk, A. I. J. M., Beck, H. E., Crosbie, R. S., de Jeu, R. A. M., Liu, Y. Y., Podger, G. M., et al. (2013). The millennium drought in southeast Australia
(2001–2009): Natural and human causes and implications for water resources, ecosystems, economy, and society. Water Resources
Research, 49, 1040–1057. https://doi.org/10.1002/wrcr.20123
Vicente-Serrano, S. M., Beguería, S., Lorenzo-Lacruz, J., Camarero, J. J., López-Moreno, J. I., Azorin-Molina, C., et al. (2012). Performance of
drought indices for ecological, agricultural, and hydrological applications. Earth Interactions, 16(10), 1–27. https://doi.org/10.1175/
2012EI000434.1
Xia, Y., Mitchell, K., Ek, M., Sheffield, J., Cosgrove, B., Wood, E., et al. (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.
Journal of Geophysical Research, 117, D03109. https://doi.org/10.1029/2011JD016048
Xia, Y., Sheffield, J., Ek, M. B., Dong, J., Chaney, N., Wei, H., et al. (2014). Evaluation of multi-model simulated soil moisture in NLDAS-2. Journal
of Hydrology, 512, 107–125. https://doi.org/10.1016/j.jhydrol.2014.02.027
Yuan, X., & Wood, E. F. (2013). Multimodel seasonal forecasting of global drought onset. Geophysical Research Letters, 40, 4900–4905.
Zhang, Q., Sun, P., Li, J., Singh, V. P., & Liu, J. (2014). Spatiotemporal properties of droughts and related impacts on agriculture in Xinjiang,
China. International Journal of Climatology, 35(7), 1254–1266. https://doi.org/10.1002/joc.4052

LORENZ ET AL. 8373