Agricultural and Forest Meteorology 237–238 (2017) 105–122

Contents lists available at ScienceDirect

Agricultural and Forest Meteorology

journal homepage: www.elsevier.com/locate/agrformet

Meteorological drought forecasting for ungauged areas based on

machine learning: Using long-range climate forecast and remote
sensing data
Jinyoung Rhee a , Jungho Im b,∗
a
Climate Research Department, APEC Climate Center, Busan, Republic of Korea
b
School of Urban and Environmental Engineering, Ulsan National Institute of Science and Technology (UNIST), Ulsan, Republic of Korea

a r t i c l e i n f o a b s t r a c t

Article history: A high-resolution drought forecast model for ungauged areas was developed in this study. The Standard-
Received 28 September 2016 ized Precipitation Index (SPI) and Standardized Precipitation Evapotranspiration Index (SPEI) with 3-, 6-,
Received in revised form 31 January 2017 9-, and 12-month time scales were forecasted with 1–6-month lead times at 0.05 × 0.05◦ resolution. The
Accepted 6 February 2017
use of long-range climate forecast data was compared to the use of climatological data for periods with
Available online 10 February 2017
no observation data. Machine learning models utilizing drought-related variables based on remote sens-
ing data were compared to the spatial interpolation of Kriging. Two performance measures were used;
Keywords:
one is producer’s drought accuracy, defined as the number of correctly classified samples in extreme,
Drought forecasting
Machine learning
severe, and moderate drought classes over the total number of samples in those classes, and the other
Climate forecast data is user’s drought accuracy, defined as the number of correctly classified samples in drought classes over
Remote sensing the total number of samples classified to those classes. One of the machine learning models, extremely
Spatial interpolation randomized trees, performed the best in most cases in terms of producer’s accuracy reaching up to 64%,
while spatial interpolation performed better in terms of user’s accuracy up to 44%. The contribution of
long-range climate forecast data was not significant under the conditions used in this study, but further
improvement is expected if forecast skill is improved or a more sophisticated downscaling method is
used. Simulated decreases of forecast error in precipitation and mean temperature were tested: the sim-
ulated decrease of forecast error in precipitation improves drought forecast while the decrease of forecast
error in mean temperature does not contribute much. Although there is still some room for improvement,
the developed model can be used for drought-related decision making in ungauged areas.
© 2017 Elsevier B.V. All rights reserved.

1. Introduction studies. In this study, we adopted the definition of the Intergovern-

Droughts have caused significant losses and damages. Defining
the types of drought helps to monitor droughts and to build strate-
gies for the preparation and response to droughts (Ward et al., 2006;
Harou et al., 2010). Wilhite and Glantz (1985) addressed the con-
ceptual definition of drought as well as the operational definition
of drought. The operational definition includes the onset, severity,
termination, and frequency of drought. They also defined meteo-
rological, agricultural, hydrological, and socio-economic droughts
based on their review of more than 150 studies. More recently,
Mishra and Singh (2010) reviewed the definitions of previous
∗ Corresponding author.
E-mail address: ersgis@unist.ac.kr (J. Im).
mental Panel on Climate Change Special Report (IPCC SREX, 2012):
“a period of abnormally dry weather long enough to cause a serious
hydrological imbalance.”
Real-time or near real-time drought monitoring is invaluable for
timely response to drought events. Drought early warning systems
help to perform drought assessment and monitoring, and to make
appropriate decisions in a timely manner by providing drought
information, and to ultimately reduce drought damages. Drought
forecasts, if feasible, can provide drought information in advance,
enabling the reduction of drought impacts by securing appropriate
resources and planning effective allocation of them.
There are not many existing drought forecasting systems,
but a variety of methods have been studied to provide drought
forecast data to be used for drought-related decision making.
Mishra and Singh (2011) listed the components of drought

http://dx.doi.org/10.1016/j.agrformet.2017.02.011
0168-1923/© 2017 Elsevier B.V. All rights reserved.
106 J. Rhee, J. Im / Agricultural and Forest Meteorology 237–238 (2017) 105–122
forecasting as hydro-meteorological variables, drought indices,
large-scale climate indices, methodologies, and output. The hydro-
meteorological variables used for drought forecasting can be
determined according to the type of drought of interest; precipi-
tation is the most important variable for meteorological drought,
while soil moisture and reservoir level are essential for agricul-
tural and hydrological droughts, respectively. Drought indices such
as the Standardized Precipitation Index (SPI; McKee et al., 1993)
and the Palmer Drought Severity Index (PDSI; Palmer, 1965), or
the combination of any related variables, are used as indicators for
drought conditions. Large-scale climate variabilities such as the El
Nino-Southern Oscillation (ENSO) and Arctic Oscillation (AO) may
be used for long-range drought forecast. Methodologies of drought
forecasting include regression models (e.g., Leilah and Al-Khateeb,
2005), time series models (e.g., Han et al., 2010), probability models
(e.g., Cancelliere et al., 2007), artificial neural network models (e.g.,
Morid et al., 2007), and hybrid models (e.g., Mishra et al., 2007), and
each method produces outputs to quantify the drought condition
determining the initiation and termination of drought, the nature
of severity, and the probability of occurrence, among other values
(Mishra and Singh, 2011).
Long-range climate forecast data for variables such as precip-
itation, air temperature, and relative humidity with lead times
up to 6 months can be used for drought forecasting. The sea-
sonal drought outlook of the National Oceanic and Atmospheric
Administration (NOAA) of the US has already used 3-month
long-range climate forecast data as well as the 1-month precipi-
tation and temperature forecast of the Climate Prediction Center
(CPC), short-term climate forecast data of the Weather Predic-
tion Center, soil moisture model results, the probability of the
termination and reduction of PDSI, climatology, and initial con-
ditions to produce seasonal drought outlook results (CPC, 2016).
Drought forecast data based on long-range climate forecasts are
reprocessed sometimes to be easily understood by end users;
Steinemann (2006) developed the Forecast Precipitation Index
(FPI) so that decision makers can easily use drought forecast
results after conducting a questionnaire survey on water resources
managers to bridge the gap between climate science and social
decisions.
Each method introduced above has its own advantages to pre-
dict drought-related variables. In order to forecast drought indices
that can be calculated based on hydro-meteorological variables
from climate model outputs, the use of long-range climate forecast
is a good choice since it is not necessary to use any proxy vari-
able or to depend on any indirect method. Mo and Lyon (2015)
recently performed meteorological drought prediction based on
SPI using the North American Multi-Model Ensemble (NMME) and
the Global Precipitation Climatology Center (GPCC) precipitation
data at 1 × 1◦ (about 100 × 100 km near the equator). This resolu-
tion is too coarse though to provide detailed drought information
for decision makers. Weather station data can replace the GPCC
precipitation. In this case, there should be a model to provide
information for ungauged areas between the locations of weather
stations.
In this study, a high resolution meteorological drought fore-
cast model was developed to provide drought forecast information
based on SPI and Standardized Precipitation-Evapotranspiration
Index (SPEI) for ungauged areas. The high resolution (0.05 × 0.05◦ )
of the developed model enables the provision of detailed data for
regional and local uses and ultimately the reduction of drought
impact. The objectives of this study are to (1) develop a drought
forecast model based on the combination of remote sensing and
long-range forecast data using machine learning for ungauged
areas, and (2) provide improved ranges of drought forecast in case
of the improvement of forecasting skill of the long-range forecast
data.
2. Study area and data
2.1. Study area
The study area is South Korea, located in northeast Asia (Fig. 1).
The area of South Korea is about 100,284 km2 . About 19% of the
area is composed of rice paddies and other crop lands, and about
64% consists of forests (KOSIS, 2016a). The annual average temper-
ature of the five main cities, Seoul, Incheon, Gangneung, Mokpo,
and Busan ranges from 11.4–14.1 ◦ C, and average annual precipi-
tation ranges from 1108–1445 mm (KOSIS, 2016b). The temporal
scope of this study is from January 2003 to August 2015 and one to
six months of lead time were used.
The methodology of this study may be applied to any region
with appropriate input variables related to drought. South Korea is
located in the mid-latitude region with partial influence of large-
scale atmosphere-ocean interactions such as ENSO, and the forecast
skill of the long-range climate forecast data is also known to be
lower compared to tropical regions (e.g., Wang et al., 2001; Wang
et al., 2004). The study area was thus selected to examine the per-
formance of machine learning models under the limited skill of the
long-range forecast data. The study area will be expanded to areas
with distinct influence of large-scale atmosphere-ocean interac-
tions such as Southeast Asia in the following studies.
There have been serious historical droughts in South Korea.
The events in 1939, 1968, 1978, and 1982 were considered as
the most extreme drought events before the 1990s (Sim, 2009).
After the 1990s, the area experienced a nation-wide drought dur-
ing 1994–1995 (Park and Schubert, 1997). The 2001 and 2008–2009
drought events were also recorded as devastating ones especially
in Kangwon Province located in the northeast of South Korea where
more than 50,000 residents experienced drinking water shortage
(Sim, 2009). More recently, a short-term spring drought occurred
in 2012 resulting in less than 30% of normal precipitation, and a
long-term drought initiated in 2013 continued to 2015, especially
in the northwest part of South Korea.
2.2. Data
2.2.1. Automatic synoptic observation system (ASOS) data
Drought conditions in this study were measured using drought
indices, and the reference drought index values were calculated
from the ASOS data. Monthly precipitation and temperature data
from 61 ASOS stations in the study area (Fig. 1) were obtained
from the Korea Meteorological Administration for the period
of January 1973–August 2015. Potential evapotranspiration data
were obtained from mean air temperature using the Thornth-
waite method (Thornthwaite, 1948). Despite that this method
could be misleading during winter seasons in the study area
because it assumes zero potential evapotranspiration for below
zero degree Celsius temperatures, the physically-based methods
such as Penman-Monteith or Hargreaves could not be used because
the minimum data requirements are the minimum and maximum
temperature while long-range climate forecast only provides mean
temperature.
2.2.2. Drought indicators
Many drought indices have been developed since the mid-20th
century. The most widely used drought indices include the PDSI
developed in the 1960s and the SPI developed in the 1990s. The
Standardized Precipitation Evapotranspiration Index (SPEI), which
was developed by Vicente-Serrano et al. (2010) considering not
only the precipitation used for SPI but also atmospheric moisture
demand, has also been used in many studies. Usually drought index
values are calculated at weather stations locations from time series
 J. Rhee, J. Im / Agricultural and Forest Meteorology 237–238 (2017) 105–122
Fig. 1. Study area of South Korea.
Table 1
Drought condition classifications for SPI and SPEI (McKee et al., 1993).
Classification Index Value
Extremely wet (EW) ≥2.00
Very wet (VW) 1.50–1.99
Moderately wet (MW) From 1.00 to 1.49
Near normal (NN) From 0.99 to −0.99
Moderate drought (MD) From −1.00 to −1.49
Severe drought (SD) From −1.50 to −1.99
Extreme drought (ED) ≤−2.00
data of meteorological variables to assess and monitor drought
conditions.
While many drought indices exist, each with their own advan-
tages and disadvantages (Mishra and Singh, 2010), we used SPI
and SPEI in this study. The SPI was recommended for meteorologi-
cal drought monitoring by the World Meteorological Organization
(WMO) (Hayes et al., 2011). SPEI can be calculated using similar
processes with SPI, but it considers both precipitation and poten-
tial evapotranspiration, and reflects water availability in a more
realistic way.
2.2.2.1. Standardized Precipitation Index. SPI can be calculated by
accumulating precipitation over the desired time scale (e.g., 3
months for 3-month SPI, 6 months for 6-month SPI), fitting the
precipitation data to an appropriate probability distribution such
as gamma or pearson type-III (Guttman, 1999), and converting it
to a gaussian probability distribution for the standardization. The
gamma probability distribution was used in this study as it is an
appropriate distribution for fitting precipitation data in the South
Korea region (Lee and Kim, 2012).
Since it uses precipitation data only, it can be applied in many
areas that lack data for various variables. Spatial and temporal com-
parisons are also possible due to the standardization. Since the time
scale is determined by the user, SPI with a specific time scale can
be used for monitoring appropriate types of drought as the drought
condition becomes aggravated. The mean and standard deviation
of SPI are zero and one respectively. The drought classes suggested
by McKee et al. (1993) were used in this study (Table 1).
107
Based on the two-parameter gamma distribution, 3-, 6-, 9-, and
12-month SPI values for 61 ASOS stations were calculated as ref-
erence data. The thirty-year period from 1981 to 2010 was used as
the calibration period for parameter estimation. In order to com-
pare with the machine learning models, SPI values using long-range
forecast were also calculated for the weather station locations and
then spatially interpolated.
2.2.2.2. Standardized Precipitation Evapotranspiration Index. SPEI
uses the same drought classification as SPI (Table 1). SPEI is based
on a conceptual water balance where the difference between pre-
cipitation and evapotranspiration equals the summation of runoff,
groundwater, and soil storage. The standardization enables spatial
and temporal comparisons of the index. In this study, the R SPEI
package (version 1.6; http://cran.r-project.org/web/packages/SPEI/
) was used. Potential evapotranspiration was also obtained based
on the Thornthwaite method when calculating SPEI. As with SPI,
3-, 6-, 9-, and 12-month SPEI values for 61 ASOS stations were cal-
culated as reference data using the thirty-year period from 1981 to
2010 as the calibration period for parameter estimation. In order
to compare SPEI with the machine learning models, SPEI values
using long-range forecast were calculated for the weather station
locations and then spatially interpolated.
2.2.3. Remote sensing data related to drought
Drought-related variables can be monitored using remote sens-
ing data (e.g., AghaKouchak et al., 2015; McVicar and Jupp, 1998).
Since remote sensing data provide spatially and temporally con-
tinuous information, the use may compensate the limitation of
point-based observation data for hydro-meteorological variables
as well as enable drought monitoring in ungauged areas. Hydro-
meteorological variables include precipitation, potential or actual
evapotranspiration, soil moisture, runoff, reservoir level, stream-
flow, and temperature. Precipitation is the most suitable variable
for meteorological drought monitoring. Streamflow or reservoir
level data may be used for monitoring hydrological drought.
Soil moisture provides valuable information to assess agricultural
drought. Precipitation and evapotranspiration can be considered
as drivers controlling drought conditions, while other variables of
108 J. Rhee, J. Im / Agricultural and Forest Meteorology 237–238 (2017) 105–122
Table 2
Remote sensing data used in this study.
Product Temporal Spatial
Resolution Resolution
TRMM 3B43 Monthly 0.25 × 0.25◦
GPM IMERG Monthly 0.1 × 0.1◦
GPM IMERG Daily 0.1 × 0.1◦
MCD43C4 16-day (Produced 0.05 × 0.05◦
every 8 days with
16 days of
acquisition)
MYD11C3 Monthly 0.05 × 0.05◦
MYD13C2 Monthly 0.05 × 0.05◦
streamflow and reservoir level are response variables determined
by the drought. Both types of variables are useful for effective
drought studies. In this study, precipitation (PRCP), daytime land
surface temperature (LST DAY), nighttime land surface temper-
ature (LST NIGHT), potential evapotranspiration (PET) estimated
using mean temperature, the Normalized Difference Vegetation
Index (NDVI), and the Normalized Difference Water Index (NDWI)
were used (Table 2).
Tropical Rainfall Measuring Mission (TRMM; Huffman et al.,
2007) data developed by National Aeronautics and Space Admin-
istration (NASA) and Japan Aerospace Exploration Agency (JAXA)
as well as Global Precipitation Measurement (GPM; Huffman et al.,
2015) data were used for precipitation. The TRMM 3B43 data were
obtained from the Goddard Earth Sciences Data and Information
Service Center (GES DISC), USA; it combines three-hour TRMM 3B42
data, the Climate Anomaly Monitoring System (CAMS)’s global
gridded precipitation, and Global Precipitation Climatology Cen-
ter (GPCC)’s global gridded observed precipitation. Since TRMM’s
descent after July 2014 due to a lack of fuel, GPM Integrated
Multi-satellite Retrievals for GPM (IMERG) obtained from the Pre-
cipitation Measurement Missions (PMM) of NASA, USA were used
from July 2014. Since the monthly data has latency of 4 months at
the time of the data analysis of this study, daily data were merged
into monthly data from May to August 2015.
Daytime and nighttime Land Surface Temperature (LST) data
were obtained from the MYD11C3 Land Surface Temperature and
Emissivity product of the Moderate Resolution Imaging Spectro-
radiometer (MODIS) sensor onboard NASA’s Aqua satellite (Wan,
1999). The data were obtained from the Earth Observing System
Data and Information System of NASA, USA.
NDVI data were obtained from the MYD13C2 Vegetation Indices
product of the MODIS sensor onboard the Aqua satellite (Huete
et al., 1999). Reflectance values of the red and near infrared regions
are used to monitor the vitality of vegetation in terms of the changes
in chlorophyll and spongy mesophyll content of vegetation (Ke
et al., 2015). The relationship between NDVI and drought, how-
ever, depends on the phenological phase, the type of vegetation,
and the length of drought (Ji and Peters, 2003; Vicente-Serrano
et al., 2013; Yagci et al., 2015; Muriithi et al., 2016). Thus, NDWI,
which was developed by Gao (1996) and shows high correlations
with drought conditions (Rhee et al., 2010), was also used in this
study. It uses two bands, 0.86 and 1.24 ␮m, which show similar lev-
els of scattering by vegetation (Gao, 1996). The 1.24 ␮m channel is
more sensitive to water existing in vegetation compared to other
channels with longer wavelengths because the energy absorption
by water inside vegetation is stronger in this region (Gao, 1996).
NDWI is known to be less sensitive to atmospheric effects com-
pared to NDVI and detects the overall amount of water existing
in layers of leaves (Gao, 1996); it can be very useful for drought
studies along with NDVI (e.g., Anderson et al., 2010; Chakraborty
Variable Data Use Start Data Use End Latency
Date Date
Precipitation Jan 2003 June 2014 Archived past
data
Precipitation July 2014 Apr 2015 4 Months
Precipitation May 2015 Aug 2015 18 h (Late Run)
Bands1-7 Surface Jan 2003 Aug 2015 NA
Reflectance
Land Surface Jan 2003 Aug 2015 NA
Temperature
(Day, Night)
NDVI Jan 2003 Aug 2015 NA
and Shehgal, 2010; Gu et al., 2007). Surface reflectance data were
obtained from the MCD43C4 Nadir Bidirectional Reflectance Distri-
bution Function (BRDF)-Adjusted Reflectance product of the MODIS
sensor (Strahler et al., 1999). This product combines data from both
the Aqua and Terra satellites. Only data from seven bands out of 35
bands of MODIS are provided through this product. The 16-day data
were converted to monthly data using the numbers of days of the
16-day period belonging to each month.
2.2.4. Large-scale climate index
Drought events are sometimes connected to large-scale atmo-
sphere and ocean interactions even in mid-latitude regions. Kang
(1998) examined the statistical correlations between interannual
variations of temperature and precipitation in South Korea and the
Southern Oscillation Index (SOI). High winter temperatures and low
summer temperatures, as well as enhanced summer precipitation,
were observed during El Nino years. Cha (1999) also reported that
there have been low summer temperature and high winter tem-
peratures during El Nino years, as well as increased precipitation.
Kim (2004) performed the Rotated Empirical Orthogonal Function
(REOF) analysis of 500 hPa geopotential height showing the first
ENSO mode and the second Arctic Oscillation (AO) mode.
In this study, the Multivariate ENSO Index (MEI; Wolter and
Timlin, 2011) and Arctic Oscillation Index (AOI; Higgins et al.,
2002) were examined for their contributions to drought forecast-
ing. Among various ENSO indices, the MEI was adopted in this study.
It uses six variables: sea-level pressure, zonal and meridional wind
speed, SST, surface air temperature, and total cloudiness fraction of
the sky. The MEI values are extracted as the first EOF mode. MEI data
from 1950 to August 2015 were obtained from NOAA Earth Sys-
tem Research Laboratory. AOI data from 1950 to August 2015 were
obtained from NOAA CPC. Correlations between the meteorological
drought indices versus MEI as well as AOI were calculated because
existing studies did not provide consistent time lags between large-
scale signals and drought occurrence in the study area (e.g. Kang,
1998; Cha, 1999; Kim, 2004). Although Pearson’s correlations were
generally low (data not shown), a one-year time lag was applied for
both MEI and AOI since it provided statistically significant results
with a 90% confidence level.
2.2.5. Long-range climate forecast data
Precipitation and 2-m temperature data with 1–6 month lead
times were obtained from six individual Global Climate Mod-
els (GCMs) used for APEC Climate Center (APCC) Multi-Model
Ensemble (MME). They are MSC CanCM3 (Meteorological Service
of Canada Climate Centre third generation coupled global climate
model), MSC CanCM4 (Meteorological Service of Canada Climate
Centre fourth generation coupled global climate model), POAMA
(Predictive Ocean Atmosphere Model for Australia), NCEP CFSv2
(National Centers for Environmental Prediction Coupled Forecast
 J. Rhee, J. Im / Agricultural and Forest Meteorology 237–238 (2017) 105–122 109

Table 3
Six Global Climate Models (GCM) used in this study.

Product Reference Temporal/Spatial Resolution Variable Hindcast Forecast

MSC CanCM3 Scinocca et al. (2008), Verseghy (2000) Monthly, 2.5 × 2.5◦ Precipitation, 2 m Air Jan. 1982–Jan. 2011 Jan. 2013–Aug. 2015
MSC CanCM4 Merryfield et al. (2013) Temperature Feb. 1981–Jan. 2011 Jan. 2013–Aug. 2015
NASA Rienecker et al. (2008), Vernieres et al. (2012) Dec. 1981–Nov. 2012 Jan. 2013–Aug. 2015
NCEP CFSv2 Saha et al. (2014) Mar. 1982–Sep. 2009 Jan. 2013–Aug. 2015
PNU Sun and Ahn (2011) Mar. 1980–Dec. 2012 Jan. 2013–Aug. 2015
POAMA Alves et al. (2003) Jan. 1983–Dec. 2011 Jan. 2013–Aug. 2015

System version 2), PNU (Pusan National University), and NASA
(National Aeronautics and Space Administration) (Table 3). Units
are mm/day and Kelvin (K) for precipitation and temperature,
respectively.
3. Methodology
We focused on the meteorological drought caused by a pro-
longed period of a lack of precipitation. SPI and SPEI with 3-, 6-,
9-, and 12-month time scale were used (SPI3, SPI6, SPI9, SPI12,
SPEI3, SPEI6, SPEI9, and SPEI12) for analyses. Lead times from 1 to
6 months were used. When calculating the index, observation data
were used when available (past period), and the remaining period
was filled with long-range forecast data (future period). This means
that the period with observation data is longer with a shorter lead
time, and more persistency or memory of the index will appear
with a shorter lead time. For the SPI9 forecast with 3-month lead
time as an example, observation data with a 6-month period and
long-range forecast data for the following 3-month period are used.
3.1. Drought category classification accuracy measures
3.2. Drought forecasting at gauge locations
The use of long-range forecast data was explored for drought
forecasting and compared to the baseline using climatology as per-
formed by Mo and Lyon (2015). Instead of using 1 × 1◦ grids, data
for 61 ASOS weather stations were used. For the station locations,
drought index values of SPI and SPEI with 3-, 6-, 9-, and 12-month
time scales were calculated. For each lead time, observation data
were used for the past period, and two methods were used to fill the
future period: one method uses long-range forecast data (F-method
hereafter) and the other method uses climatology (baseline; C-
method hereafter).
The climatology data were obtained from the median value of
100 samples randomly derived from the observation data of the
same month. The long-range forecast data were used in a way that
the percent increment of modeled precipitation or temperature
anomaly is applied to observational climatology data to form the
calibrated (bias-corrected) model data (Quan et al., 2012). Before
being combined with observation data, ensembles of percent incre-
ment of precipitation or temperature from six individual models
were obtained (Quan et al., 2012).

Ma∗ = Ma /Mc (3)

Drought forecasting results can be obtained as a form of drought
index value, or its corresponding category. Since decision making Mcal = Oc (1 + Ma∗ ) (4)
regarding drought is usually done based on drought classes, the
where Ma∗
is the percent increment of model prediction anomaly,
accuracy of drought category forecasting was used when compar-
Ma is model prediction anomaly, Mc is model climatology, Oc is
ing different forecasting models.
observation climatology, and Mcal is calibrated model prediction.
There are several measures in the accuracy assessment of
classifications: overall accuracy, user’s accuracy, and producer’s
3.3. Drought forecasting for ungauged areas
accuracy. The user’s accuracy is used as a measure of the error
of commission that occurs when the samples are classified to a
The use of machine learning models was tested to obtain spa-
class incorrectly, while the producer’s accuracy is used as a mea-
tially distributed drought forecasting data (ML-method hereafter)
sure of the error of omission that occurs when the samples are not
and a spatial interpolation was compared as a baseline method
classified to a class (Jensen, 2005). Two types of cost may occur in
(I-method hereafter). Kriging (gaussian process) was used as the
drought forecasting; these are when a serious drought condition
spatial interpolation baseline method (Cressie, 1993); it has been
was forecasted but actually did not happen, and when a serious
widely used to estimate spatially distributed information in many
drought happened but it was not forecasted beforehand. Consid-
studies. Leave-one-out cross-validation of spatial interpolation was
ering the tremendous economic cost of drought events that have
used to obtain estimated values at weather stations locations for
not been forecasted, it was assumed that the error of omission is
comparison.
more serious than the error of commission and the producer’s accu-
The use of long-range forecast data was also compared to the
racy is more appropriate to use in this study. Overall accuracy was
use of climatology in ungauged areas. Four different combinations
not used because it may be misleading due to the inclusion of all
were tested: F-ML-method (long-range forecast data and machine
wet classes. Thus, producer’s accuracy was used as the main accu-
learning), C-ML-method (climatology data with machine learning),
racy measure for drought category classification, and only drought
classes of extreme drought (ED), severe drought (SD), and moder-
ate drought (MD) were considered (producer’s Drought Accuracy
hereafter):
The number of samples correctly classified to ED, SD, or MD classes
Producer’ s Drought Accuracy = (1)
Total number of samples in ED, SD, or MD classes
User’s accuracy was also used as a supplementary accuracy mea-
sure for drought category classification, to provide false positive
prediction information (user’s Drought Accuracy hereafter):
Then number of samples correctly classified to ED, SD, or MD classes
User’ s Drought Accuracy = (2)
Total number of samples classified to ED, SD, or MD classes
110 J. Rhee, J. Im / Agricultural and Forest Meteorology 237–238 (2017) 105–122
F-I-method (long-range forecast data with spatial interpolation),
and C-I-method (climatology data with spatial interpolation).
Machine learning models and the spatial interpolation model were
first developed using observation remote sensing data assuming
perfect forecasts (PF-ML model and PF-I model hereafter) and the
performance of the models using long-range forecast data or cli-
matological data were tested.
Drought forecasting was performed by machine learning in this
study using the scikit-learn python library. Machine learning mod-
els provide results that are stable against outliers and especially
useful when used with remote sensing data (Mishra and Desai,
2006). Three machine learning models, decision trees (DT), random
forest (RF), and extremely randomized trees (ERT), were tested.
The machine learning models use the remote sensing data intro-
duced in Section 2.2.3 (LST DAY, LST NIGHT, NDVI, NDWI) as well
as large-scale climate indices (MEI, AOI), the long-range forecast
or climatological data combined with remote sensing data (PRCP,
PET), and temporal and topographical information (MONTH, ELEV)
(Fig. 2). The drought-related remote sensing variables were used to
provide initial conditions of ungauged areas and to help estimate
future drought conditions of the areas. The long-range forecast or
climatological data are combined with remote sensing data in the
same way that they are combined with weather station data; only
the observation data in this case are not weather station data but
historical remote sensing data.
Feature scaling was applied to LST DAY, LST NIGHT, NDVI,
NDWI, PRCP, and PET; the times series of these variables are nor-
malized using the maximum and minimum values for each month
and location. This is to separate weather impacts from the environ-
mental component (Kogan, 1995). Seasonality can also be removed
by this normalization. Separate spatial analyses were not per-
formed; data for all 61 weather station locations were pooled and
examined together assuming spatial homogeneity of the study area.
This assumption may not be realistic and it will be examined in fur-
ther studies. Monthly analyses were not performed either. Instead,
MONTH was included as a discrete input variable to the models.
Elevation data (ELEV) was also included as an input variable to use
topographical information about the area.
Only SPI and SPEI values below −0.5 (the value slightly on the
drier side of the Near Normal category) were used for training in
order to avoid the tendency that forecast results are more assigned
to wetter classes. It would be ideal to have training data with evenly
distributed classes, but it is not the case because SPI and SPEI values
follow the gaussian distribution. The full range of data is used for
performance evaluation.
3.3.1. Decision trees
DT is used to develop either a classification or a regression model
and has been widely used in remote sensing applications (Jensen
and Im 2007; Rhee et al., 2008; Im et al., 2012; Lu et al., 2014; Tor-
bick and Corbiere 2015). Since it is a white-box model, the resultant
tree can be explicitly examined. A tree consists of several leaves and
branches. Input data is split at each leaf node in a way to optimize
the parameter of a split function, until all subsets have the same
value of the target variable or the estimation is no longer improved.
The Gini impurity index was used as a split function:

m
 
Gini = fi (1 − fi ) , i ∈ 1, 2, . . ., m
i=1
where m is the number of classes and fi is the fraction of samples
classified as class i.
Sensitivity analysis was performed for the most important
parameters. The maximum depth of the tree (MAX DEPTH) is
important for DT models. The tree will be fully developed unless
MAX DEPTH is set to a certain value, and a fully developed tree
sometimes shows overfitting. MAX DEPTH values of 3, 5, 10, and
the full development were tested and the parameter value showing
the best model performance was selected in each case.
3.3.2. Random forest
RF is based on Classification and Regression Trees (CART), a rep-
resentative decision tree algorithm (Breiman, 2001). While the DT
model uses only one tree, the RF model uses an ensemble of trees to
avoid a well-known problem of CART (i.e., overfitting and sensitiv-
ity to training data configuration), Thus, it gained great popularity
in remote sensing applications during the past decade (Li et al.,
2013; Kim et al., 2014,2015; Han et al., 2015; Park et al., 2016).
Randomness is used during the training processes: during the pro-
cess of bagging (bootstrap aggregating), a random subset of input
vectors, with the same size as the original input vectors, are created
allowing repetition and are used for each tree. Thus input vectors
for each tree are independent but have the same probability distri-
bution. All trees vote for the final class or produce a probabilistic
classification result. The use of feature bagging, where only a sub-
set of input variables is used for each leaf, ensures stability against
outliers and noise.
RF also provides the relative importance of input variables
through the Out-Of-Bag (OOB) errors when a variable is perturbed.
OOB errors are the differences between actual and estimated
values of input vectors that are not used for training. Unlike
cross-validation errors, the OOB process provides unbiased error
estimations without using a part of the training data for estimat-
ing errors (Breiman, 2001). In this study, however, cross validation
was performed for all models leaving each year out to avoid over-
fitting since the characteristics of input vectors may be different
each year and the length of observation data is short. Leave-one
year-out cross validation was used instead of leave-one-out cross
validation because the length of training data is relatively short and
the characteristics of data of a certain year may be different from
those of other years.
The number of trees (NUM TREES) is another important param-
eter for RF and ERT models in addition to the MAX DEPTH of each
tree. Since RF and ERT models use an ensemble of trees, overfitting
is less likely even with fully developed trees; each tree is trained
with a subset of input vectors created to minimize correlations with
subsets of input vectors for other trees. NUM TREES of 10, 50, 100,
and 200 were tested and the parameter value showing the best
model performance was selected in each case.
3.3.3. Extremely randomized trees
ERT is similar to RF, but additional randomness is added to the
way that input vectors are split at each leaf. In ERT models, input
vectors are split in random ways and the split showing the best
result is chosen later.
3.3.4. Cross validation
Relationships between input variables and the target variable
may be different each year, since the target variable is drought cat-
egory and its behavior is not well defined due to the short length
of remote sensing records. In order to avoid overfitting, leave-one-
out cross validation was performed not with each sample but with
each year. It may be called as leave-one year-out cross validation
(5)
in this study. Results of the leave-one year-out cross validation
were compared to 10-fold with shuffle cross validation, leave-
one sample-out cross validation, and leave-one station-out cross
validation, and all other results outperformed the leave-one year-
out cross validation results (data not shown). This indicates that
unknown patterns may appear in the future and the leave-one
year-out cross validation is more appropriate to evaluate model
performance.
 J. Rhee, J. Im / Agricultural and Forest Meteorology 237–238 (2017) 105–122
Fig. 2. Structure of each machine learning model.
Since data from January 2003 to August 2015 were used, 13
models were developed for each method, index, time scale, and
lead time. The performance evaluation in the next section is done
using the average of these models in each case. The relative impor-
tance of input variables was also evaluated using the average of the
results from 13 models in each case.
4. Results and discussion
4.1. Accuracy at gauge locations
Drought Accuracy values were compared between the C-
method and F-method for SPI and SPEI with 3-, 6-, 9-, and 12-month
time scales and lead times of 1–6 months (Figs. 3 and 4). For SPI3 and
SPEI3, only 1- and 2-month lead time forecasts retain observation
data. For SPI6 and SPEI6, 1- to 5-month lead time forecasts retain
observation data. Longer past periods exist for 9- and 12-month
time scales.
In all cases where at least one month of past period is included,
the C-method showed higher producer’s Drought Accuracy com-
pared to the F-method (Fig. 3). This indicates that drought forecasts
based on climatological data performed better than the forecasts
based on long-range forecast data. It may be because the bias-
correction method of the GCM results using the percent increment
of anomaly is too simple; appropriate downscaling methods should
be applied to use long-range forecast data for the study area. In
cases with no observation data included (SPI3 and SPEI3 forecasts
with 3–6-month lead times as well as SPI6 and SPEI6 forecasts with
6-month lead time), Drought Accuracy appeared as zero meaning
no forecast skill for both C-method and F-method. Drought Accu-
racy values larger than zero observed here are mostly due to the
memory (persistency) of the drought index itself.
The F-method showed higher user’s Drought Accuracy values
than the C-method for some cases (SPI6 with shorter lead times,
SPI9, SPEI9, and SPI12; Fig. 4). The user’s Drought Accuracy values
for SPI3 with lead times longer than 4 months are not available
because none of the samples were classified to drought classes. It
should be noted though that user’s Drought Accuracy values can be
misleading when only a few samples were classified to drought; for
example, the 50% user’s Drought Accuracy of SPEI3 with 6-month
lead time was obtained because only two samples were classified
111
to drought and one of them was correctly classified (Fig. 4; sample
number data not shown).
4.2. Accuracy for ungauged areas
Machine learning models can be either classification models or
regression models. If regression models are developed, drought
classes can be derived based on regression results and drought
classes (Table 1). Machine learning DT, RF, and ERT models were
developed as both classification and regression models in this
study, and the regression models mostly performed better. Only
Drought Accuracy values derived from the regression models were
further examined.
The Drought Accuracy of C-I, F-I, C-ML (DT, RF, ERT models), and
F-ML (DT, RF, ERT models) methods were compared to SPI and SPEI
with 3-, 6-, 9-, and 12-month time scales and lead times of 1–6
months (Figs. 5 and 6). For producer’s Drought Accuracy based on
C-I and F-I methods, producer’s Drought Accuracy values for both
SPI and SPEI change with lead time as they do at gauge locations
(Figs. 3 and 5). The only difference is that the accuracy values are
smaller due to spatial interpolation. Other machine learning-based
methods show quite different changes with lead time. Although the
accuracy of the C-I and F-I methods decreases with increased lead
time, many of the methods still show somewhat high producer’s
Drought Accuracy with longer lead times for SPI and SPEI with 9-
and 12-month time scales and also for SPEI3 and SPEI6 (Fig. 5).
Methods based on machine learning mostly performed bet-
ter than the methods based on spatial interpolation. While there
appeared not much difference between methods using climatologi-
cal data and long-range forecast data, methods using climatological
data performed slightly better in many cases (Fig. 5). As a result, the
C-ML (ERT model) method and the F-ML (ERT model) method per-
formed the best except for the 6-month lead SPI6 forecast. In that
case, C-ML (RF model) performed the best (Fig. 5). The uncertainty
ranges of producer’s Drought Accuracy are provided for C-ML (ERT
model) method using mean ± one standard deviation, showing that
the method performs significantly better than many other methods
(Fig. 5).
For user’s Drought Accuracy, methods based on machine learn-
ing show consistently low accuracy (Fig. 6). Methods based
on spatial interpolation perform better than methods based on
machine learning for drought indices with longer time scales (SPI9,
112 J. Rhee, J. Im / Agricultural and Forest Meteorology 237–238 (2017) 105–122

Fig. 3. Producer’s Drought Accuracy of SPI and SPEI forecasts with 3-, 6-, 9-, and 12-month time scales and lead times of 1–6 months based on C-method and F-method.

SPEI9, SPI12, and SPEI12) as well as for drought indices with shorter
time scales (SPI3, SPEI3, SPI6, and SPEI6) with shorter lead times
(Fig. 6). The uncertainty ranges of user’s Drought Accuracy are pro-
vided for F-I method using mean ± one standard deviation, showing
that the method performs significantly better than many other
methods (Fig. 6). Low user’s Drought Accuracy by machine learn-
ing implies that non-drought samples were not effectively learned
by machine learning compared to the drought training samples,
resulting in overestimation of drought conditions.
Weather stations’ density may affect the results because fewer
stations provide less information for spatial interpolation with
increased distances between stations, and also a smaller dataset for
training of machine learning models. Instead of using 61 weather
stations, 41 and 21 stations randomly selected were tested; the
results were similar though.
4.3. Improved ranges for ungauged areas
What if the skill of the long-range forecast improves? Different
levels of decrease in forecast error were assumed using observation
data, and their performance were examined (Figs. 7 and 8). Three
simulation scenarios were designed:
 J. Rhee, J. Im / Agricultural and Forest Meteorology 237–238 (2017) 105–122 113

Fig. 4. User’s Drought Accuracy of SPI and SPEI forecasts with 3-, 6-, 9-, and 12-month time scales and lead times of 1–6 months based on C-method and F-method.

• F1: 30% decrease of forecast error in precipitation
• F2: 30% decrease of forecast error in mean temperature
• F3: 30% decrease of forecast error in precipitation and mean tem-
perature
Producer’s Drought Accuracy was improved by F1-ML and F3-
ML methods compared to F-ML methods, while F2-ML methods
produced similar results to F-ML methods (Fig. 7). The results indi-
cate that the decrease of forecast error in precipitation contributed
to the improvement of drought forecast. For methods based on spa-
tial interpolation, F1-I, F2-I, and F3-I methods showed even lower
accuracy values than F-I method in many cases (Fig. 7). The shorter
the time scale of SPI or SPEI, the larger the improved range of the
accuracy (Table 4). User’s Drought Accuracy was also improved by
F1-ML and F3-ML methods compared to F-ML methods in most
cases, and F1-I, F2-I, and F3-I methods showed even lower accuracy
values than F-I method (Fig. 8; Table 5).
4.4. Relative importance of input variables
The machine learning models used in this study provides rela-
tive importance information about the input variables. The relative
importance of input variables of the ERT model was examined for
SPI and SPEI with 3-, 6-, 9-, and 12-month time scale and 1–6-month
114 J. Rhee, J. Im / Agricultural and Forest Meteorology 237–238 (2017) 105–122

Fig. 5. Producer’s Drought Accuracy of SPI and SPEI forecasts with 3-, 6-, 9-, and 12-month time scales and lead times of 1–6 months based on C-I, F-I, C-ML (DT, RF, ERT
model), and F-ML (DT, RF, ERT model) methods. The blue background indicates the uncertainty ranges of the accuracy using mean ± one standard deviation for the C-ML
(ERT model) method. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
 J. Rhee, J. Im / Agricultural and Forest Meteorology 237–238 (2017) 105–122 115

Fig. 6. User’s Drought Accuracy of SPI and SPEI forecasts with 3-, 6-, 9-, and 12-month time scales and lead times of 1–6 months based on C-I, F-I, C-ML (DT, RF, ERT model),
and F-ML (DT, RF, ERT model) methods. The red background indicates the uncertainty ranges of the accuracy using mean ± one standard deviation for the F-I method. (For
interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
116 J. Rhee, J. Im / Agricultural and Forest Meteorology 237–238 (2017) 105–122

Fig. 7. Producer’s Drought Accuracy of SPI and SPEI forecasts with 3-, 6-, 9-, and 12-month time scales and lead times of 1–6 months based on C-I, F-I, F1-I, F2-I, F3-I, C-ML
(ERT model), F-ML (ERT model), F1-ML (ERT model), F2-ML (ERT model), and F3-ML (ERT model) methods.

lead times (Fig. 9). The relative importance values are given using
the importance of the most important variable as 100.
The most important variable in all case is PRCP. It is not surpris-
ing that PET also ranked high for SPEI forecast, especially for SPEI9
and SPEI12 (PET is also on the right-axis for them in Fig. 9f and h).
This is because SPEI is originally calculated using precipitation and
potential evapotranspiration.
The relative importance of other input variables varies with
drought index, time scale, and lead time of the forecast. If only input
variables with greater than 10% relative importance for at least one
 J. Rhee, J. Im / Agricultural and Forest Meteorology 237–238 (2017) 105–122 117

Fig. 8. User’s Drought Accuracy of SPI and SPEI forecasts with 3-, 6-, 9-, and 12-month time scales and lead times of 1–6 months based on C-I, F-I, F1-I, F2-I, F3-I, C-ML (DT
model), F-ML (DT model), F1-ML (DT model), F2-ML (DT model), and F3-ML (DT model) methods.

lead time are considered, MONTH, NDVI, MEI, and LST DAY were
chosen as important variables in addition to PET and PRCP. Tempo-
ral information MONTH appears more important for SPI3 and SPEI3
forecasts with longer lead times while NDVI appears more impor-
tant for SPI3 and SPEI3 forecasts with 3-month lead time (Fig. 9a
and b). LST DAY has more than 10% relative importance for SPI6
and SPEI9 forecasts with 2-month lead time (Fig. 9c and f). The
large-scale climate index MEI also appears relatively important for
drought index forecast with shorter time scales, which are SPI3,
118 J. Rhee, J. Im / Agricultural and Forest Meteorology 237–238 (2017) 105–122

Fig. 9. Relative importance of input variables for SPI and SPEI forecasts with 3-, 6-, 9-, and 12-month time scales and lead times of 1–6 months based on ERT model.
 J. Rhee, J. Im / Agricultural and Forest Meteorology 237–238 (2017) 105–122 119

Table 4
Improved ranges of producer’s Drought Accuracy (%) for ungauged areas for F1-ML (ERT model) method.

Lead Time (month) SPI3 SPI6 SPI9 SPI12 SPEI3 SPEI6 SPEI9 SPEI12

1 5.51 2.74 −0.18 1.01 4.09 1.6 −0.36 −0.2

2 8.05 3.97 −0.67 2.55 5.88 2.93 0.9 1.41
3 11.54 3.11 1.88 5.19 10.15 3.72 2.75 1.6
4 7.78 9 −0.1 2.71 12.6 3.83 0.49 1.8
5 5.64 3.94 2.01 5.44 15.53 4.94 0.65 1.07
6 10.09 −5.83 0.21 7.5 15.39 3.67 1.54 2.98

Table 5
Improved ranges of user’s Drought Accuracy (%) for ungauged areas for F1-ML (DT model) method.

Lead Time (month) SPI3 SPI6 SPI9 SPI12 SPEI3 SPEI6 SPEI9 SPEI12

1 2.6 −0.03 0.37 0.45 2.99 2 0.19 −0.07

2 5.46 0.91 0.45 0.46 2.25 1.96 0.19 1
3 −11.01 1.11 0.34 1.3 11.15 2.1 0.93 1.31
4 15.54 1.12 −0.11 1.19 12.17 3.25 0.88 0.62
5 3.92 1.95 0.98 0.41 4.88 3.41 1 0.02
6 5.56 1.4 1.32 0.33 3.48 1.8 1.74 0.25

SPEI3, and SPEI6, especially for the SPEI3 forecast with all lead times
(Fig. 9b).
4.5. Drought forecast maps
Although the machine learning models performed better than
spatial interpolation for drought index forecast in terms of
producer’s Drought Accuracy, there is still some room for improve-
ment. The resultant drought forecast maps for August 2015 with
1–3-month lead times were created for SPI3, SPEI3, SPI6, and SPEI6
(Figs. 10 and 11). In August 2015, the study area was in Mod-
erate Drought when drought index values for 61 ASOS weather
stations were averaged, while the northwest part of the study area
including Seoul, Incheon, and Gyeonggi was in Extreme Drought
when drought index values for 6 weather stations were averaged
(Table 6).
Drought conditions in weather stations (circles on maps) are
actual observation values of August 2015, while spatially dis-
tributed values are forecasts (Figs. 10 and 11). Drought conditions in
the 1-month lead SPI3 forecast look close to the observation data
except in the southeast part of the study area where no drought
conditions could not be forecasted correctly (Fig. 10a). For the 1-
month lead SPEI3 forecast, on the other hand, the Extreme Drought
in the northwest seems not adequately forecasted (Fig. 10d). Similar
conditions were observed for SPI6 and SPEI6 too (Fig. 11). Spatial
interpolation generally predicted less serious drought conditions
overall (data not shown).
4.6. Comparisons of the methods
4.6.1. Climatology and long-range climate forecast
Although not in all cases, the methods using climatological data
(C, C-I, and C-ML) performed better than the methods using long-
range forecast data (F, F-I, and F-ML) for SPI and SPEI forecast in
many cases in terms of producer’s Drought Accuracy. This may be
because the simple bias-correction method of long-range forecast
data used here could not reflect the spatial variability within the
study area. More sophisticated downscaling methods need to be
tested in further study. The improvement of drought forecast in
ungauged areas, however, can be expected with the enhancement
of the forecast skill of precipitation as was proven with assumed
forecast data scenarios.
4.6.2. Spatial interpolation and machine learning models
Three machine learning models, DT, RF, and ERT, were tested in
this study and the ERT model generally performed the best in terms
of producer’s Drought Accuracy. Both the C-ML (ERT model) and F-
ML (ERT model) methods produced higher Drought Accuracy than
the C-I and F-I methods. This means machine learning models could
provide more promising drought forecast data compared to spatial
interpolation techniques, such as the Kriging technique used in this
study.
In many cases, machine learning models (C-ML and F-ML meth-
ods based on the ERT model) even performed better than F1-I, F2-I,
and F3-I using spatial interpolation with assumed enhanced fore-
cast data. It was shown that the use of machine learning models
supplements the current shortcomings of long-range forecast data
in this study area.
4.6.3. Limitations
This study still has some limitations. It may not be appropriate to
use climatological data due to the non-stationary characteristics of
changing future climate, although in many cases the methods using
climatological data perform better than the methods using long-
range climate forecast data for drought indices such as SPI3 and
SPEI3. Fortunately, the methods using long-range climate forecast
data produced comparable Drought Accuracy values to the meth-
ods using climatological data, especially those based on machine
learning. This implies that machine learning-based drought pre-
diction would be more promising with improved climate forecast.
Soil moisture data could not be included due to very large dis-
crepancies between satellite-based soil moisture estimates and flux
observation data. The use of a simple bias-correction method for
long-range forecast data can also be a weakness of the study. These
limitations will be addressed in further studies. Additional input
variables for initial drought conditions and geographical informa-
tion will be explored.
5. Conclusion
In this study, drought forecasting using long-range forecast data
was tested and compared to the method using climatological data.
In order to obtain drought forecast information in ungauged areas,
machine learning of long-range forecast data or climatological data
combined with remote sensing data was performed and the perfor-
mance was compared to the spatial interpolation of Kriging. One of
the machine learning models, the ERT model, performed the best in
most cases in terms of producer’s Drought Accuracy. Although the
contribution of long-range forecast data was not significant under
the conditions used in this study, further improvement is expected.
The use of machine learning models supplements the limitations of
120 J. Rhee, J. Im / Agricultural and Forest Meteorology 237–238 (2017) 105–122

Fig. 10. 1- to 3-month lead SPI3 and SPEI3 forecast in the study area for August 2015.

Fig. 11. 1- to 3-month lead SPI6 and SPEI6 forecast in the study area for August 2015.

Table 6
Drought index values in August 2015.

Drought Index Average of 61 stations (Entire study area) Average of 6 stations (Northwest part)

SPI3 −1.35 −2.12

SPEI3 −1.38 −2.69
SPI6 −1.38 −2.64
SPEI6 −1.47 −3.09
 J. Rhee, J. Im / Agricultural and Forest Meteorology 237–238 (2017) 105–122
the simple bias-correction method used in this study for drought
forecast in ungauged areas.
For high-resolution drought forecasting in ungagged areas, it
is recommended to use the machine learning ERT model with cli-
matological data or long-range climate forecast data for drought
indices with longer time scales. This model produces drought
forecast data spatially distributed in 0.05 × 0.05◦ resolution. The
producer’s Drought Accuracy of the machine learning-based meth-
ods with long-range climate forecast data reaches up to 56% for SPI
forecast and 64% for SPEI forecast.
Drought forecast skill using long-range forecast data can be
improved if the forecast skill of long-range forecast data is
improved. The assumed enhanced forecast F1-ML and F3-ML
showed improved Drought Accuracy in many cases. The drought
forecast skill based on long-range forecast data may also be
improved if more sophisticated bias-correction and downscaling
methods are used. Further studies deserve to examine this.
The level of Drought Accuracy derived in this study may be used
as a baseline to compare in future studies trying to produce high
resolution drought forecast data. Although there is still some room
for improvement, the developed model may be used for drought-
related decision making in ungauged areas.
Acknowledgment
This research was supported by the APEC Climate Center and the
National Disaster Management Research Institute of Korea and by
National Space Lab Program through the National Research Founda-
tion of Korea (NRF) funded by the Ministry of Science, ICT, & Future
Planning (Grant: NRF-2013M1A3A3A02042391).
References
AghaKouchak, A., Farahmand, A., Melton, F.S., Teixeira, J., Anderson, M.C.,
Wardlow, B.D., Hain, C.R., 2015. Remote sensing of drought: Progress,
challenges, and opportunities. Rev. Geophys. 53, 452–480, http://dx.doi.org/10.
1002/2014RG000456.
Alves, O., Wang, G., Zhong, A., Smith, N., Tseitkin, F., Warren, G., Schiller, A.,
Godfrey, S., Meyers, G., 2003. POAMA: Bureau of Meteorology operational
coupled model seasonal forecast system. In: Stone, R., Partridge, I. (Eds.),
Science for Drought: Proceedings of the National Drought Forum. Brisbane, Apr
2003, Department of Primary Industries, pp. 49–56.
Anderson, L.O., Malhi, Y., Aragao, L.E.O.C., Ladle, R., Arai, E., Barbier, N., Philips, O.,
2010. Remote sensing detection of droughts in Amazonian forest canopies.
New Phytol. 187, 733–750.
Breiman, L., 2001. Random forests. Mach. Learn. 45, 5–32.
CPC (Climate Prediction Center), 2016. Seasonal Drought Outlook. http://www.cpc.
ncep.noaa.gov/products/expert assessment/sdo summary.html.
Cancelliere, A., Mauro, G.D., Bonaccorso, B., Rossi, G., 2007. Drought forecasting
using the Standardized Precipitation Index. Water Resour. Manag. 21, 801–819.
Cha, E.-J., 2012. Summer 2012. Short Changma, drought, heatwave, tropical night,
heavy rain, and typhoon. J. Korean Soc. Hazard Mitig. 12 (3), 4–9 (In Korean).
Chakraborty, A., Shehgal, V.K., 2010. Assessment of agricultural drought using
MODIS derived normalized difference water index. J. Agric. Phys. 10, 28–36.
Cressie, N.A.C., 1993. Statistics for Spatial Data, Revised ed. John Wiley & Sons Inc.,
pp. 928.
Gao, B.-C., 1996. NDWI – a normalized difference water index for remote sensing
of vegetation liquid water from space. Remote Sens. Environ. 58, 257–266.
Gu, Y., Brown, J.F., Verdin, J.P., Wardlow, B., 2007. A five-year analysis of MODIS
NDVI and NDWI for grassland drought assessment over the central Great
Plains of the United States. Geophys. Res. Lett. 34, L06407, http://dx.doi.org/10.
1029/2006GL029127.
Guttman, N.B., 1999. Accepting the Standardized Precipitation Index: a calculation
algorithm. J. Am. Water Resour. Assoc. 35, 311–322.
Han, P., Wang, P.X., Zhang, S.Y., Zhu, D.H., 2010. Drought forecasting based on the
remote sensing data using ARIMA models. Math. Comput. Model. 51,
1398–1403.
Han, H., Lee, S., Im, J., Kim, M., Lee, M.-I., Ahn, M.H., Chung, S.-R., 2015. Detection of
convective initiation using Meteorological Imager onboard Communication,
Ocean, and Meteorological Satellite based on machine learning approaches.
Remote Sens. 7, 9184–9204.
Harou, J., Medellin-Azuara, J., Zhu, T., Tanaka, S., Lund, J., Stine, S., Olivares, M.,
Jenkins, M., 2010. Economic consequences of optimized water management
for a prolonged, severe drought in California. Water Resour. Res. 46, W05522,
http://dx.doi.org/10.1029/2008WR007681.
121
Hayes, M., Svoboda, M., Wall, N., Widhalm, M., 2011. The lincoln declaration on
drought indices: universal meteorological drought index recommended. Bull.
Am. Meteorol. Soc. 92 (4), 485–488.
Higgins, R.W., Leetmaa, A., Kousky, V.E., 2002. Relationships between climate
variability and winter temperature extremes in the United States. J. Clim. 15,
1555–1572.
Huete, A., Justice, C., van Leeuwen, W., 1999. MODIS Vegetation Index (MOD 13)
Algorithm Theoretical Basis Document, Version 3. Goddard Space Flight
Center, National Aeronautics and Space Administration, 30 April 1999. 120 p.
Huffman, G.J., Adler, R.F., Bolvin, D.T., Gu, G., Nelkin, E.J., Bowman, K.P., Stocker, E.F.,
Wolff, D.B., 2007. The TRMM multi-satellite precipitation analysis:
quasi-global, multi-year, combined-sensor precipitation estimates at fine
scale. J. Hydrometeorol. 8, 33–55.
Huffman, G.J., Bolvin, D.T., Braithwaite, D., Hsu, K., Joyce, R., Kidd, C., Nelkin, E.J.,
Xie, P., 2015. Algorithm Theoretical Basis Document (ATBD) Version 4.5. NASA
Global Precipitation Measurement (GPM) Integrated Multi-satellitE for GPM
(IMERG) National Aeronautics and Space Administration, 16 November 2015.
30 p.
IPCC SREX, 2012. Managing the risks of extreme events and disasters to advance
climate change adaptation. In: Field, C.B., Barros, V., Stocker, T.F., Qin, D.,
Dokken, D.J., Ebi, K.L., Mastrandrea, M.D., Mach, K.J., Plattner, G.-K., Allen, S.K.,
Tignor, M., Midgley, P.M. (Eds.), A Special Report of Working Groups I and II of
the Intergovernmental Panel on Climate Change. Cambridge University Press,
Cambridge, UK, and New York, NY, USA, 582 p.
Im, J., Jensen, J., Jensen, R., Gladden, J., Waugh, J., Serrato, M., 2012. Vegetation
cover analysis of hazardous waste sites in Utah and Arizona using
hyperspectral remote sensing. Remote Sens. 4, 327–353.
Jensen, J.R., Im, J., 2007. Remote sensing change detection in urban environments,
In Geo-Spatial Technologies in Urban Environments. In: Jensen, R.R., Gatrell,
J.D., McLean, D. (Eds.), Policy, Practice, and Pixels. , 2nd edition. Springer Berlin
Heidelberg, Berlin, Heidelberg, pp. 7–31.
Jensen, J.R., 2005. Introductory Digital Image Processing: A Remote Sensing
Perspective, 3rd ed. Prentice Hall, Upper Saddle River, New Jersey, USA, pp. 544.
Ji, L., Peters, A.J., 2003. Assessing vegetation response to drought in the northern
Great Plains using vegetation and drought indices. Remote Sens. Environ. 87,
85–98.
KOSIS (Korean Statistical Information System), 2016a. Land Use Statistics,
Available online at http://kosis.kr.
KOSIS (Korean Statistical Information System), 2016b. Climatology, Available
online at http://kosis.kr.
Kang, I.S., 1998. Relationship between El Nino and climate variability in Korean
Peninsula. Atmos. J. Korean Meteorol. Soc. 34 (3), 390–396 (In Korean).
Ke, Y., Im, J., Lee, J., Gong, H., Ryu, Y., 2015. Characteristics of Landsat 8 OLI-derived
NDVI by comparison with multiple satellite sensors and in-situ observations.
Remote Sens. Environ. 164, 314–329.
Kim, Y.H., Im, J., Ha, H.K., Choi, J.-K., Ha, S., 2014. Machine learning approaches to
coastal water quality monitoring using GOCI data, GISci. Remote Ses. 51,
158–174.
Kim, M., Im, J., Han, H., Kim, J., Lee, S., Shin, M., Kim, H.-C., 2015. Landfast sea ice
monitoring using multisensory fusion in the Antarctic. GISci. Remote Sens. 52,
239–256.
Kim, S., 2004. The Pattern Change of Atmospheric Circulation in the North
Hemisphere Associated with Spring Precipitation in Korea Master’s Thesis.
Gongju University. Department of Atmospheric Science, pp. 61.
Kogan, F.N., 1995. Application of vegetation index and brightness temperature for
drought detection. Adv. Space Res. 15, 91–100.
Lee, J.H., Kim, C.J., 2012. A multimodel assessment of the climate change effect on
the drought severity-duration-frequency relationship. Hydrol. Process. 27,
2800–2813, http://dx.doi.org/10.1002/hyp.9390.
Leilah, A.A., Al-Khateeb, S.A., 2005. Statistical analysis of wheat yield under
drought conditions. J. Arid Environ. 61, 483–496.
Li, M., Im, J., Beier, C., 2013. Machine learning approaches for forest classification
and change analysis using multi-temporal Landsat TM images over Huntington
Wildlife Forest. GISci. Remote Sens. 50, 361–384.
Lu, Z., Im, J., Rhee, J., Hodgson, M., 2014. Building type classification using spatial
and landscape attributes derived from LiDAR remote sensing data. Landsc.
Urban Plan. 130, 134–148.
McKee, T.B., Doesken, N.J., Kleist, J., 1993. The relationship of drought frequency
and duration of time scales. In: Proc. of the 8th Conf. of Applied Climatology,
Anaheim CA, USA, Amer. Meteor. Soc, pp. 179–184.
Merryfield, W.J., Lee, W.-S., Boer, G.J., Kharin, V.V., Scinocca, J.F., Flato, G.M.,
Ajayamohan, R.S., Fyfe, J.C., 2013. The Canadian seasonal to interannual
prediction system. Part I: models and initialization. Mon. Weather Rev. 141,
2910–2945.
Mishra, A.K., Desai, V.R., 2006. Drought forecasting using feed-forward recursive
neural network. Ecol. Model. 198, 127–138.
Mishra, A.K., Singh, V.P., 2010. A review of drought concept. J. Hydrol. 391,
202–216.
Mishra, A.K., Singh, V.P., 2011. Drought modeling – a review. J. Hydrol. 403,
157–175.
Mishra, A.K., Desai, V.R., Singh, V.P., 2007. Drought forecasting using a hybrid
stochastic and neural network model. J. Hydrol. Eng. ASCE 12 (6), 626–638.
Mo, K.C., Lyon, B., 2015. Global meteorological drought prediction using the North
American multi-model ensemble. J. Hydrometeorol. 16, 1409–1424.
122 J. Rhee, J. Im / Agricultural and Forest Meteorology 237–238 (2017) 105–122
Morid, S., Smakhtin, V., Bagherzadeh, K., 2007. Drought forecasting using artificial
neural networks and time series of drought indices. Int. J. Climatol. 27,
2103–2111.
Muriithi, F., Yu, D., Robila, S., 2016. Vegetation response to intensive commercial
horticulture and environmental changes within watersheds in central
highlands, Kenya, using AVHRR NDVI data. GISci. Remote Sens. 53, 1–21.
Palmer, W.C., 1965. Meteorological drought. In: U.S. Weather Bureau Research
Paper 45., pp. 58.
Park, C.-K., Schubert, S.D., 1997. On the nature of the 1994 East Asian summer
drought. J. Clim. 10, 1056–1070.
Park, S., Im, J., Jang, E., Rhee, J., 2016. Drought assessment and monitoring through
blending of multi-sensor indices using machine learning approaches for
different climate regions. Agric. Forest Meteorol. 216, 157–169.
Quan, X.-W., Hoering, M.P., Lyon, B., Kumar, A., Bell, M.A., Tippett, M.K., Wang, H.,
2012. Prospects for dynamicla prediction of meteorological drought. J. Appl.
Meteorol. Climatol. 51, 1238–1252.
Rhee, J., Im, J., Carbone, C.J., Jensen, J.R., 2008. Delineation of climate regions using
in-situ and remotely-sensed data for the Carolinas. Remote Sens. Environ. 112,
3099–3111.
Rhee, J., Im, J., Carbone, G.J., 2010. Monitoring agricultural drought for arid and
humid regions using multi-sensor remote sensing data. Remote Sens. Environ.
114, 2875–2887.
Rienecker, M.M., Suarez, M.J., Todling, R., Bacmeister, J., Takacs, L., Liu, H.-C., Gu,
W., Sienkiewicz, M., Koster, R.D., Gelaro, R., Stajner, I., Nielsen, J.E., 2008. NASA,
118, NASA/TM-2008-104606.
Saha, S., Moorthi, S., Wu, X., Wang, J., Nadiga, S., Tripp, P., Behringer, D., Hou, Y.-T.,
Chuang, H., Iredell, M., Ek, M., Meng, J., Yang, R., Mendez, M.P., van den Dool, H.,
Zhang, Q., Wang, W., Chen, M., Becker, E., 2014. The NCEP Climate Forecast
System version 2. J. Clim. 27, 2185–2208.
Scinocca, J.F., Mcfarlane, N.A., Lazare, M., Li, J., Plummer, D., 2008. Technical note:
the CCCma third generation AGCM and its extension into the middle
atmosphere. Atmos. Chem. Phys. 8, 7055–7074.
Sim, K.O., 2009. A Study on Extreme Drought. National Disaster Management
Institute, pp. 170 (In Korean).
Steinemann, A.C., 2006. Using climate forecasts for drought management. J. Appl.
Meteorol. Climatol. 45, 1353–1361.
Strahler, A.H., Lucht, W., Schaaf, C.B., Tsang, T., Gao, F., Li, X., Muller, J.-P., Lewis, P.,
Barnsley, M.J., 1999. MODIS BRDF/Albedo Product: Algorithm Theoretical Basis
Document, Version 5.0. Goddard Space Flight Center, National Aeronautics and
Space Administration, April 1999. 53 p.
Sun, J., Ahn, J.B., 2011. A GCM-based forecasting model for the landfall of tropical
cyclones in China. Adv. Atmos. Sci. 28, 1049.
Thornthwaite, C.W., 1948. An approach toward a rational classification of climate.
Geogr. Rev. 38 (1), 5594, http://dx.doi.org/10.2307/210739.
Torbick, N., Corbiere, M., 2015. Mapping urban sprawl and impervious surfaces in
the northeast United States for the past four decades. GISci. Remote Sens. 52,
746–764.
Vernieres, G., Rienecker, M.M., Kovach, R., Keppenne, C.L., 2012. ). In: Suarez, M.J.
(Ed.), The GEOS-iODAS: Description and Evaluation, Technical Report Series on
Global Modeling and Data Assimilation, vol. 30. National Aeronautics and
Space Administration, Greenbelt, Maryland, 61 p.
Verseghy, D.L., 2000. The Canadian Land Surface Scheme (CLASS): its history and
future. Atmos. Ocean 38, 1–13.
Vicente-Serrano, S.M., Begueria, S., Lopez-Moreno, J.I., 2010. A multiscalar drought
index sensitive to global warming: the Standardized Precipitation
Evapotranspiration Index. J. Clim. 23, 1696–1718.
Vicente-Serrano, S.M., Gouveia, C., Camarero, J.J., Begueria, S., Trigo, R.,
Lopez-Moreno, J.I., Azorin-Molina, C., Pasho, E., Lorenzo-Lacruz, J., Revuelto, J.,
Moran-Tejeda, E., Sanchez-Lorenzo, A., 2013. Response of vegetation to
drought time-scales across global land biomes. Proc. Natl. Acad. Sci. U. S. A.
110, 52–57.
Wan, Z., 1999. MODIS Land-Surface Temperature Algorithm Theoretical Basis
Document (LST ATBD) Version 3.3. Goddard Space Flight Center, National
Aeronautics and Space Administration, April 1999. 75 p.
Wang, G., Kleeman, R., Smith, N., Tseitkin, F., 2001. The BMRC coupled general
circulation model ENSO forecast system. Mon. Weather Rev. 130, 975–991.
Wang, B., Kang, I.-S., Lee, J.-Y., 2004. Ensemble simulations of Asian-Australian
monsoon variability by 11 AGCMs. J. Clim. 17, 803–818.
Ward, F., Hurd, B., Rahmani, T., Gollehon, N., 2006. Economic impacts of federal
policy responses to drought in the Rio Grande Basin. Water Resour. Res. 42,
W03420, http://dx.doi.org/10.1029/2005WR004427.
Wilhite, D., Glantz, M.R., 1985. Understanding the drought phenomenon: the role
of definitions. Water Int. 10, 111–120.
Wolter, K., Timlin, M.S., 2011. El Nino/southern oscillation behavior since 1871 as
diagnosed in an extended multivariate ENSO index (MEI.ext). Int. J. Climatol.
31, 1074–1087.
Yagci, A., Di, L., Deng, M., 2015. The effect of corn-soybean rotation on the
NDVI-based drought indicators: a case study in Iowa, USA, using vegetation
condition index. GISci. Remote Sens. 52, 290–314.