You are on page 1of 35

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/335712628

Statistical Downscaling Seasonal Forecast for Precipitation and Temperature


in Sumatera, Indonesia based on APCC Multi-Model Ensemble (MME) System

Preprint · September 2018

CITATIONS READS

0 68

2 authors:

Donaldi S Permana Jaepil Cho


Meteorological Climatological and Geophysical Agency APEC Climate Center
37 PUBLICATIONS   47 CITATIONS    74 PUBLICATIONS   624 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Conservation Practice Effects in the SE Coastal Plain, USA View project

Seasonal Forecast using Statistical Downscaling View project

All content following this page was uploaded by Donaldi S Permana on 10 September 2019.

The user has requested enhancement of the downloaded file.


APCC Young Scientist Support Program 2018-01

Statistical Downscaling Seasonal


Forecast for Precipitation and
Temperature in Sumatera, Indonesia
based on APCC Multi-Model Ensemble
(MME) System

Name: Dr. Donaldi Sukma Permana

Affiliation: Indonesia Agency for Meteorology Climatology and Geophysics (BMKG)

APCC Counterpart: Dr. Jaepil Cho

1
ABSTRACT
A statistical downscaling model is applied to forecast seasonal precipitation and temperature in
Sumatera Island, Indonesia using multi-model ensemble (MME) output variables as predictors. The
predictors are multiple variables from the output of 11 operational coupled general circulation
models (GCMs) associating in the APEC Climate Center (APCC) MME prediction system, climate
indices, and reanalysis data. The downscaling methods used in this study consists of four different
methods namely Simple Bias Correction (SBC), Climate Index Regression (CIR), Moving Window
Regression (MWR) and Observed Moving Window Regression (MWR-Obs).

Ten observed station data in Sumatera, Indonesia were used as the input to construct the forecast
model over the hindcast period of 1983-2012. The quality control of observed station data was
conducted by filling the missing data with the nearest station data and the closest MERRA-2 gridded
data that has a similar daily pattern. Data clustering of the seasonal precipitation data divides
Sumatera into two climate zones called Climate Zone North (CZN) and Climate Zone South (CZS).

In CZN, for precipitation, SBC is the dominant method that has good forecasting skill for all months,
except April and November. MWR has good skill forecast from May to October. No good predictor is
for November. APCC_SCOPS, APCC, NASA, and PNU are the GCMs served as best predictors for
precipitation. For temperature, most of the downscaling methods and GCMs performed good
forecasting skill for every month. The MME prediction system has the highest skill scores in August
to October and the lowest in April and December for precipitation. For temperature, the skill scores
are high for every month, even with 6-month lead prediction.

Likewise CZN, SBC has also good forecasting skill for all months for precipitation in CZS, except
February and April. There is no good predictor in February and April. In contrast to CZN, MWR is
more dominant than SBC for temperature. The highest skill scores of the MME system is in July to
September and the lowest is in January and November for precipitation. For temperature, the skill
scores are high for every month at least 4-month lead prediction.

Keywords: statistical downscaling, seasonal forecast, Sumatera, MME, SBC, CIR, MWR, MWR-Obs

1. Introduction
Indonesia is one of the most vulnerable countries for natural disasters with more than 80% of the disasters
are hydrometeorological disasters including floods, landslides, wind gusts, drought, and forest fires,
according to 2010-2017 disaster data from the Indonesian National Agency for Disaster Management
(BNPB). The climate of Indonesia is influenced by the following climate features; Madden-Julian Oscillation
(MJO) on intraseasonal timescale, Asian-Australian Monsoon and Inter-Tropical Convergence Zone (ITCZ) on
seasonal timescales, and El Niño Southern Oscillation (ENSO) and Indian Ocean Dipole (IOD) phenomena on
interannual timescales. In addition, some regions in Indonesia are also influenced by the tropical cyclones.

Seasonal forecasting has played an important role in agriculture, forestry, health, infrastructure, water
management, disaster risk management and other related application sectors of the country. The users
from each sector would like to use reliable seasonal forecast information to make important decisions and
design planning for the future activities. In Indonesia, Agency for Meteorology Climatology and Geophysics
(BMKG) has the main task to provide seasonal climate forecasting and information to those
aforementioned sectors.

2
BMKG has used both statistical and dynamical models to provide the seasonal forecast information. Since
2011 BMKG has used In-House statistical software application called Hybrid BMG (HyBMG) based on time
series analysis. It consists of several methods such as Autoregressive Integrated Moving Average (ARIMA),
Adaptive Neuro-Fuzzy Inference Systems (ANFIS) and Wavelet Transformation. This application produces a
deterministic seasonal rainfall forecast of up to 1 year period (Kadarsah, 2010; Komalasari et al., 2016;
Kuswanto and Prayoga, 2015). Since 2014 BMKG has used the dynamical model from ECMWF (Molteni et
al., 2011) with a spatial resolution of 0.75° and 51 ensemble members to produce deterministic and
probabilistic climate forecasts of up to 7 months. BMKG has also used additional NCEP-NOAA CFS v2 (Saha
et al., 2014) with a spatial resolution of 1° since 2016 to support monthly and seasonal forecasts (up to 9
months). Based on those models, BMKG produces seasonal forecast information which includes the
prediction of the onset of the dry and rainy season and rainfall characteristics (below-normal, normal and
above-normal).

In comparison to Indonesia, the Bureau of Meteorology (BoM) of Australia issued seasonal climate outlooks
based on the statistical relationships between sea surface temperature (SST) patterns and Australian
rainfall before 2013. Afterward, BoM switched to the dynamical model-based system, the Predictive Ocean
Atmosphere Model for Australia (POAMA) (Charles et al., 2015) because it has higher forecast skill than the
previous operational model. Another reason is that dynamical models are less susceptible to changes in
statistical relationships, whether these occur naturally or through climate change. To improve the forecast
skill and statistical reliability of POAMA forecasts, statistical post-processing methods such as calibration
and bridging (Schepen et al., 2014) have been applied to increase the skill of Australian seasonal rainfall
forecasts. Meanwhile, the North American countries have developed the North American Multi-Model
Ensemble (NMME) with 9 ensemble members of general circulation models (GCMs) and seasonal
prediction up to 7 months (Kirtman et al., 2014).

A multi-model ensemble (MME) technique is widely used in operational centers for long-range forecasting.
An advantage of the multi-model approach over a single model is based on diversity of the model errors
and their mutual offset which has been documented in many previous studies (e.g., Min et al., 2014;
DelSole et al., 2014; Yang et al., 2016). Nowadays, it is common practice to present climate forecasts in
terms of probability distribution functions (PDFs) because, in contrast to the deterministic forecast, it
provides additional information needed to quantify climate forecast uncertainty. Also, the MME is proved
more effective in a probabilistic framework (Yang et al., 2016; Min et al., 2017). The Asia-Pacific Economic
Cooperation Climate Center (APCC) has been issuing real-time global MME forecasts of temperature and
precipitation since November 2007 (Min et al., 2014). These forecasts contain monthly, 3-month and 6-
month mean values (since September 2013) which are issued monthly with one-month lead.

The skill of real-time operational forecasts with the APCC MME prediction system strongly depends on
seasons and regions that are higher for the tropics and boreal winter season than for extratropics and
boreal summer season (Juneng et al., 2010). The use of seasonal forecasts that is produced by APCC MME
would be suitable and useful for operational climate services in Indonesia. The global APCC MME seasonal
forecasts have coarse spatial resolution with 2.5° x 2.5° grid. In order to obtain seasonal forecasts with
higher spatial resolution or at specific station location, downscaling methods were applied to the APCC
MME prediction system. One way of downscaling techniques with low computing resources in term of
computing time and storage capacity is a statistical downscaling. This method has been widely used and
implemented in Southeast Asian countries (e.g. Kang et al., 2007; Juneng et al., 2010). Indonesia is the
3
largest country in tropical Southeast Asia would need to implement this technique to get a benefit from the
real-time seasonal forecasting based on the APCC MME system. This study aims to implement statistical
downscaling methods to the real-time seasonal forecast (lead time up to 6 months) of APCC MME in
Indonesia. The best skill predictors were selected by finding the variables with the highest temporal
correlation coefficients. Due to limited data availability, the study area is focused on Sumatera Island which
has two types of seasonal rainfall pattern according to a previous study (Aldrian and Susanto, 2003). In
addition, Sumatera Island is also vulnerable to the forest fires and haze dispersion hazards. Therefore,
reliable seasonal forecast information is very important for decision makers in planning for disaster risk
management. A similar project has been conducted in Borneo Island to specifically develop the fire and
haze early warning system for Southeast Asia (Yoo et al., 2016).

2. Data and Methods


2.1. Data

In this study, the 10 meteorological stations in Sumatera Island, Indonesia were used. The available
meteorological parameters include daily precipitation (millimeter) and maximum and minimum
temperature (°C) from 1982 to 2014. The map of station location is depicted in Fig.1 and its data is
described in Table. 1.

Figure 1. The meteorological station locations used in this study


4
Table 1. The description of observed station data used in this study

Elevation Start End


No Station Name Station ID Longitude Latitude
(masl) Year Year
Blang Bintang Banda
1 ID96011 95.42 5.52 21 1982 2014
Aceh
2 Polonia Medan ID96035 98.68 3.57 25 1982 2014
3 Kijang Tanjung Pinang ID96091 104.53 0.92 18 1982 2014
Simpang Tiga
4 ID96109 101.45 0.46 31 1982 2014
Pekanbaru
5 Tabing Padang ID96163 100.35 -0.88 3 1982 2014
6 Sultan Thaha Jambi ID96195 103.65 -1.63 25 1982 2014
7 SMB II Palembang ID96221 104.7 -2.9 10 1982 2014
8 Pangkal Pinang ID96237 106.14 -2.16 33 1982 2014
9 Fatmawati Bengkulu ID96253 102.34 -3.86 16 1982 2014

10 Raden Inten Lampung ID96295 105.18 -5.24 96 1982 2014

In conducting statistical downscaling of real-time seasonal forecast provided by APCC-MME, the following
datasets are used:

 The individual models for real-time forecast provided by APCC-MME prediction system
(http://adss.apcc21.org/)
 Monthly climate indices from NOAA (16 indices)
(http://www.esrl.noaa.gov/psd/data/climateindices/list/)
 Monthly climate indices from NCEP Reanalysis I or APCC (9 indices)
(http://www.apcc21.org/ser/indic.do?lang=en)
 NCEP/NCAR Reanalysis I monthly data
(www.esrl.noaa.gov/psd/data/gridded/data.ncep.reanalysis.derived.pressure.html)

2.2. Methodology

The overall procedures for generating statistical downscaling seasonal forecasting in this study include (1)
conduct quality control for station data with nearest station data and or satellite-derived and reanalysis
data, (2) conduct data clustering for defining climate zone, (3) construct statistical downscaling model for
seasonal forecasting, (4) combine the best predictors and methods into MME system and calculate monthly
forecasting skills, and (5) run the real-time statistical downscaling for seasonal forecast.

The observed station data are sometimes incomplete due to various reasons such as instrument
malfunction and human error. In order to conduct statistical downscaling, the completeness of the
observed data is needed. Therefore, the data filling in the observation station were conducted by filling the
gap with the nearest station data, satellite-derived and reanalysis data.

In addition, the statistical downscaling methods used in this study could only be applied to the observed
station data that has a similar seasonality pattern. Therefore, before conducting the statistical downscaling,
5
data clustering was conducted on the observed station data to obtain distinct climate zones. Although, a
previous study suggests that there are two seasonal rainfall types in Sumatera (Aldrian and Susanto, 2003),
but the study was conducted with old data from a period of 1961-1993. Therefore, data clustering were
conducted in this study with newer data using the kmlShape R-package (Genolini et al., 2016) and using
Voronoi polygons in QGIS 3.2 (QGIS development team, 2009) to generate lines between station data
points.

The products of real-time seasonal forecast generated from individual and multi GCMs in the APCC-MME
can be statistically downscaled based on the coupled pattern selection and projection (Yoo et al., 2016).
Coupled patterns represent the relationship between local precipitation and the variation of a large-scale
pattern. The pattern projection is thus based on the premise that the large-scale pattern can be well
simulated by dynamical models and that local precipitation forecasts may be retrieved from the
information in the coupled pattern as a proper transfer function. The individual GCMs used in this study
consists of models APCC, CWB, GLOSEA5, HMC, JMA, MGO, and MSC with 3-month lead time, and
APCC_SCOPS, NASA, PNU, and CMCC with 6-month lead time.

In this study, four different downscaling methods were used in accordance with the degree of utilizing the
seasonal climate forecast information. These methods are:

1. Simple Bias Correction (SBC).


SBC is a forecast-based direct downscaling method, which uses GCM’s prediction data after adjusting
the monthly mean of predicted data. For example, if the precipitation prediction data on a specific
region is needed, SBC directly uses the grid values of precipitation variables, which are produced from
GCMs over the given area. The systematic bias is adjusted for making the monthly average of
prediction the same as the average of observation for the same period.
2. Moving Window Regression (MWR)
MWR is a forecast-based indirect statistical downscaling method, which uses the simultaneous proxy
variables produced by GCMs as predictors of regression model when high correlation exists between
proxy variables and regional target variables. The MWR method uses the oceanic and atmospheric
circulation variables as predictors to improve the seasonal prediction predictability in the target area
(e.g. Kang et al., 2009; 2014). The MWR method selects the coupled pattern by scanning a moving
window over the globe (for each individual GCM). The optimal window of the coupled pattern is
defined by the area in which the correlation value between predictand and predictor is the highest.
3. Climate Index Regression (CIR)
CIR is an observation-based indirect statistical downscaling method that can be used when there is a
high correlation between global climate indices and regional target variables with lag time (e.g. Kim et
al., 2007; Kim & Kim, 2010).
4. Observed Moving Window Regression (MWR-Obs)
MWR-Obs is an observation-based direct statistical downscaling method that can be used when there
is a high correlation between NCEP/NCAR Reanalysis I and regional target variables with lag time.

The concepts of both cross-validation and split-validation were applied in order to prevent overfitting
problems. The Leave-One-Out Cross-Validation (LOOCV) technique was applied to the training period of
1983 – 2012 (30 years) during hindcast model reconstruction.

6
The procedure of statistical downscaling of seasonal climate forecast was conducted using the rSForecast R-
package version 0.6.2 developed by Dr. Jaepil Cho (APCC). The rSForecast package uses observed monthly
area average precipitation and temperature as predictands. This is the reason why the climate zone is
important. The model running was conducted in APCC. The SBC and CIR methods were run using a regular
PC within a few hours, while the MWR and MWR-Obs methods were run using the High-Performance
Computing (HPC) machine in APCC facilities within a few days. The flowchart of statistical downscaling of
real-time seasonal forecast using the APCC-MME system is given in Fig. 2.

Figure 2. Flowchart of statistical downscaling of the real-time seasonal forecast of APCC-MME (Cho, 2017)

7
3. Results and Analysis
3.1 Station Data Quality Control

The daily observation data often have missing values due to many reasons. In order to fill the missing
values, there are several ways to conduct. First, by filling the missing values with nearest station data that
has a similar daily pattern and more complete data. In this study, after considering the linear regression
between the observed station and its nearest station, only station ID96253 (Meteorology Station
Fatmawati Bengkulu) that has a similar daily pattern with its nearest station ID96255 (Climatology Station P.
Baai Bengkulu) (Fig. 3). The distance between station ID96253 and ID96255 is about 2 km (not shown).
Therefore, some of missing values in ID96253 were filled with data from ID96255

Even after filling the missing values with nearest station data, the completeness of daily observed station
data was still not good (Table 2). The best data availability was found in station ID96035 and ID96253 with 0
- 2 % missing values of monthly average and the worst was found in station ID96237 with 20 - 60 % missing
values of monthly average.

Secondly, the missing values in observed station data would be filled with gridded observed satellite-
derived and reanalysis data. In this study, two satellite-derived i.e. CHIRPS and PERSIANN and two
reanalysis datasets i.e. ERA-Interim and MERRA-2 (Gelaro et al., 2017) were compared with observed
station data to obtain the best gridded data to fill the missing values in observed station data. All of four
datasets were compared for precipitation, but only ERA-Interim and MERRA-2 were compared for
temperature. The comparison results for precipitation, maximum and minimum temperature were shown
in Fig 4, Fig. 5 and Fig. 6, respectively. The results show that MERRA-2 has the best correlation with
observed station data in Sumatra Island for all parameters. MERRA-2 has daily correlation coefficients of
0.6, 0.55, and 0.4 for precipitation, maximum and minimum temperature, respectively.

Furthermore, the missing values of observed station data were filled with the closest MERRA-2 grid data.
MERRA-2 has a spatial resolution of 0.67° longitude and 0.5° latitude. For instances, the linear regressions
of precipitation from some stations and precipitation from MERRA-2 are shown in Fig. 7. By using these
linear regression equations, the missing values at observed station data were filled. After data gap filling
with MERRA-2 data, all observed station data are complete with 0 % missing values.

8
Figure 3. Comparison of daily precipitation, maximum and minimum temperature between station ID96253
and ID96255 in the period of 1982-2014

Table 2. The completeness of observed station data after filling the gap with nearest station data

ID96011 ID96035 ID96091 ID96109 ID96163 ID96195 ID96221 ID96237 ID96253 ID96295
Mon (NA%) (NA%) (NA%) (NA%) (NA%) (NA%) (NA%) (NA%) (NA%) (NA%)
1 7.5 2.4 5.8 13.0 2.6 11.4 8.9 25.9 2.2 1.2
2 9.7 0.8 9.1 9.4 2.1 11.3 8.0 43.1 2.7 2.0
3 10.9 1.0 4.0 9.9 2.3 13.2 7.6 34.3 2.1 1.4
4 15.1 1.1 5.5 12.9 2.1 12.3 7.5 34.8 2.0 1.0
5 14.0 1.1 4.7 6.6 1.4 11.1 6.8 38.4 1.8 1.1
6 10.2 1.0 4.6 9.9 1.9 11.5 7.7 50.0 1.6 0.5
7 11.5 1.6 2.8 6.2 2.2 9.9 7.5 59.7 1.7 1.0
8 11.2 1.0 5.7 8.1 3.3 9.7 7.2 64.0 0.9 0.7
9 17.5 0.9 7.0 13.6 3.0 9.6 7.0 61.4 1.8 0.4
10 11.4 0.4 4.5 11.7 3.2 9.9 5.6 45.7 1.9 0.6
11 11.4 0.7 5.6 10.5 2.7 11.0 7.9 30.2 1.4 1.8
12 10.5 0.4 5.7 15.0 2.1 11.4 9.4 24.3 2.3 5.1

9
Figure 4. Correlation coefficients between observed station data and satellite-derived and reanalysis data
for precipitation

Figure 5. Correlation coefficients between observed station data and reanalysis data for maximum
temperature

10
Figure 6. Correlation coefficients between observed station data and reanalysis data for minimum
temperature

Figure 7. Linear regressions of precipitation at station ID96011, ID96221 and ID96237 and precipitation at
the closest MERRA-2 grid data

11
3.2 Station Data Climatology

After filling the missing values with MERRA-2 grid data, the seasonality of precipitation and temperature at
each station could be obtained. In addition, since MERRA-2 data could represent the precipitation and
temperature at stations in Sumatera Islands, it is assumed that MERRA-2 can also spatially represent the
climatology of precipitation and temperature in Sumatera Island. The climatology of precipitation and
temperature at each station and Sumatera Island are depicted in Fig 8 - 10. In general, there are two
seasonal rainfall types in Sumatera, which would further be tested using data clustering method in Section
3.3. The first seasonal rainfall type has two peaks of rainfall on March-May and October-November, while,
the second type has only one rainfall peak on December-February. In addition, Figure 10 shows the
peculiarity of maximum temperature in MERRA-2 data at station ID96295 or Lampung Meteorological
Station (red circles) where it shows a relatively high monthly maximum temperature from September to
November in comparison to other months. This peculiarity is caused by the relatively high maximum
temperature values on September to November in more than 15 years from 1983 to 2014 (Fig. 11). This
peculiarity could be considered as the weaknesses of MERRA-2 data.

12
Figure 8. (Top panel) Seasonality of precipitation in each station. (Bottom panel) The spatial monthly
average of precipitation from MERRA-2 (1980-2017)

13
Figure 9. (Top panel) Seasonality of minimum temperature in each station. (Bottom panel) The spatial
monthly average of minimum temperature from MERRA-2 (1980-2017)

14
Figure 10. (Top panel) Seasonality of maximum temperature in each station. (Bottom panel) The spatial
monthly average of maximum temperature from MERRA-2 (1980-2017)

15
Figure 11. Daily and monthly maximum temperature at station ID96295 and the closest MERRA-2 grid data
from 1983 to 2014

16
3.3 Station Data Clustering

A previous study suggests that there are two seasonal rainfall types in Sumatera (Aldrian and Susanto,
2003). However, the study was conducted with observation data from the period 1961 to 1993. In this
study, data clustering was conducted with newer data (1982 - 2014) using the kmlShape R-package
(Genolini et al., 2016) and using Voronoi polygons in QGIS 3.2 (QGIS development team, 2009) to generate
lines between station data points. The advantage of kmlShape method is because this algorithm works
based on the shapes of the trajectories rather than on classical distance (Genolini et al., 2016). In climate
data, the shape of the trajectory is associated with the time-series (seasonal) pattern while the distance is
associated with the data values.

Data clustering for precipitation is conducted with the input of monthly-average precipitation from 10
stations (1983-2014) and its result is shown in Fig. 12. Two climate zones are obtained: Climate Zone 1
which is characterized by two precipitation peaks in April and November, and Climate Zone 2 which is
characterized by one precipitation peak in December.

Because of the similarity of seasonal pattern between observed station data and MERRA-2 over Sumatera
(top panel in Fig. 8), for the sake of comparison, data clustering is also performed with the input of
monthly-average precipitation from 126 gridded data of MERRA-2 over Sumatera region (Fig. 13). Figure 13
shows that there are 3 climate zones over Sumatera derived from MERRA-2 data. However, Climate Zone 1
and Climate Zone 3 in Fig.13 have a similar seasonal pattern which is having two precipitation peaks on
March-April and November but the only difference is in precipitation intensity. The precipitation intensity in
Climate Zone 3 is higher than in Climate Zone 1 because Climate Zone 3 covers the mountainous area with
high elevation (see Fig. 1). For this reason, then Climate Zone 1 and 3 in Fig. 13 are combined into one
climate zone.

The combination of climate zones shown in Fig. 12 and Fig. 13 produce the final climate zones for observed
station data which is given in Fig. 14. The results agree with a previous study by Aldrian and Susanto (2003),
only a bit different in the middle of Sumatera Island. Later, Climate Zone 1 in Fig. 14 is referred to Climate
Zone North (CZN) or Climate Zone of Northern Sumatera and Climate Zone 2 in Fig. 14 is referred to Climate
Zone South (CZS) or Climate Zone of Southern Sumatera.

Similar procedures for data clustering are also performed for maximum and minimum temperature over
Sumatera Island. The results are shown in Fig. 15 for maximum temperature and in Fig. 16 for minimum
temperature. Although, there are two climate zones for maximum and minimum temperature they have a
similar seasonal pattern. The only difference is the magnitude of temperature. For maximum temperature,
the high values are observed twice a year in April and October and the low values are observed in July and
December. For minimum temperature, the high and low values are observed once a year on April-May and
July-August, respectively. Therefore, there is only one climate zone for maximum and minimum
temperature over Sumatera.

17
450
400

Precipitation (mm)
350
300
250
200
150
100
50

OCT
JAN

AUG
FEB

SEP

DEC
JUN
JUL

NOV
APR
MAR

MAY
Climate Zone 1 Climate Zone 2

Figure 12. Data clustering for precipitation from 10 stations over Sumatera

450
400
Precipitatation (mm)

350
300
250
200
150
100
50
AUG

OCT
JAN

JUN
JUL

SEP

NOV
FEB

APR

DEC
MAR

MAY

Climate Zone 1 Climate Zone 2


Climate Zone 3

Figure 13. Data clustering for precipitation from 126 gridded data of MERRA-2 over Sumatera

18
MERRA2 Final Climate Zone
(Combined MERRA2 and Station Data)

2
QC station data

Figure 14. Combination of climate zones for precipitation from observed station data and MERRA-2 gridded
data over Sumatera

Station
data
MERRA2

34
Temperature (C)
Temperature (c)

29
32
30 27

28 25
AUG

OCT
JAN
FEB

JUL

SEP

DEC
JUN

NOV
APR
MAY
MAR

AUG

OCT
JAN
FEB

SEP
JUN
JUL

NOV
APR

DEC
MAR

MAY

CZ1 CZ2 CZ1 CZ2

Figure 15. Data clustering for maximum temperature from observed station data (left) and MERRA-2
gridded data (right) over Sumatera

19
Station
data MERRA2

25 25

Tempearature (C)
Temperature (C)

24 23
23
22 21
21 19
AUG

OCT
JAN

JUN
JUL

SEP

NOV
FEB

APR

DEC
MAY
MAR

AUG

OCT
JAN
FEB

SEP
JUN
JUL

NOV
APR

DEC
MAR

MAY
CZ1 CZ2 CZ1 CZ2

Figure 16. Data clustering for minimum temperature from observed station data (left) and MERRA-2
gridded data (right) over Sumatera

3.4 Statistical Downscaling Seasonal Forecast

The procedure of statistical downscaling of seasonal forecast was conducted using rSForecast R-packages
version 0.6.2 developed by Dr. Jaepil Cho (APCC) with four different statistical downscaling methods namely
(1) Simple Bias Correction – SBC, (2) Climate Index Regression – CIR, (3) Moving Window Regression – MWR,
and (4) Observed Moving Window Regression – MWR-Obs. The rSForecast package uses observed monthly
area average precipitation and temperature as predictands. Since there are two climate zones of
precipitation in Sumatera, the statistical downscaling seasonal forecast is performed for each climate zone.

For the Climate Zone of Northern Sumatera (CZN), the selected statistical downscaling methods and GCMs
for precipitation and temperature with lead time up to 3-month are given in Table 3. For precipitation, it is
clear that SBC is the dominant method that has good forecasting skill for all months, except April and
November. The second dominant method is MWR which has good skill forecast from May to October. The
CIR has only good skill forecast in August and September. In April, only MWR-Obs has a good skill forecast.
While no statistical downscaling method has a good prediction skill in November. It is interesting that both
April and November are the two peaks of precipitation in CZN. In terms of GCMs, APCC_SCOPS, APCC, NASA,
and PNU are the best predictors for precipitation up to 3-month lead. For temperature, most of the
downscaling methods performed good forecasting skill and most of GCMs served as good predictors for
every month.

The performance of the MME seasonal forecast up to 6-month lead for CZN is shown in Fig. 17. For
precipitation, the highest skill scores are obtained for prediction in August, September, and October with
TCC of 0.7 – 0.8, whereas the lowest skill scores are obtained for prediction in April and December with TCC
of 0.45 – 0.56. As expected, the TCC of prediction for each month decreases with increasing leading month.

20
For temperature, the predictability for CZN is high for every month, even with 6-month lead time (TCC of
0.6 – 0.85).

MWR and MWR-Obs search the best predictors, in terms of variables, location area and lead time, from
forecasted models and observation NCEP/NCAR Reanalysis I, respectively. For CZN, examples of selected
predictors in MWR and MWR-Obs are given in Fig. 18 and Fig. 19, respectively.

Table 3. The selected statistical downscaling methods and models for precipitation (top) and temperature
(bottom) in Northern Sumatera for 1 to 3-month lead

21
Figure 17. Monthly TCC and NRMSE for precipitation and temperature in Northern Sumatera

22
Figure 18. Examples of the selected predictors in MWR for 1-month lead precipitation in May, August, and
September in Northern Sumatera

23
Figure 19. Examples of the selected predictor in MWR-Obs for 1-month lead June precipitation in Northern
Sumatera

After selecting best predictors from GCMs, climate index and reanalysis data using four different statistical
methods over the hindcast period (1983-2012), those information are integrated and saved to run the real-
time seasonal forecast model. The monthly skill scores are updated when the new observed station data
are available. For instance, the 6-month prediction of precipitation and temperature issued on June 2018
for July to December 2018 in Northern Sumatera is depicted in Fig. 20.

24
Figure 20. The 6-month forecast of precipitation and temperature for June to December 2018 in Northern
Sumatera

25
For the Climate Zone of Southern Sumatera (CZS), the selected statistical downscaling methods and GCMs
for precipitation and temperature with lead time up to 3-month are given in Table 4. For precipitation, SBC
is also the dominant method that has good forecasting skill for all months, except February and April. The
second dominant method is MWR which has good skill forecast in March, May-June, and August-October.
The CIR has only good skill forecast in September and October. There is no statistical downscaling method
that has a good forecasting skill in February and April. MWR-Obs has a good skill forecast for June with 3-
month lead time. In terms of GCMs, APCC_SCOPS, APCC, JMA, NASA, and PNU are the best predictors for
precipitation up to 3-month lead. This is almost similar to the best predictor’s GCMs for CZN. For
temperature, most of GCMs served as good predictors for every month. However, it is interesting that
MWR is more dominant than SBC for almost every month. The number of selected predictor’s GCMs for
SBC is decreasing from February-April to September-November.

The performance of the MME seasonal forecast up to 6-month lead for CZS is shown in Fig. 21. For
precipitation, the highest skill scores are obtained for prediction in July, August, September, and October
with TCC of 0.7 – 0.8, whereas the lowest skill scores are obtained for prediction in January and November
with TCC of 0.4 – 0.54. For temperature, the predictability for CZS is high for every month with at least 4-
month lead time (TCC of 0.47 – 0.84).

Likewise in CZN, MWR and MWR-Obs search the best predictors, in terms of variables, location area and
lead time, from forecasted models and observation NCEP/NCAR Reanalysis I, respectively. For CZS,
examples of selected predictors in MWR and MWR-Obs are given in Fig. 22 and Fig. 23, respectively.
Likewise in CZN, after selecting best predictors from GCMs, climate index and reanalysis data using four
different statistical methods over the hindcast period (1983-2012), those information are integrated and
saved to run the real-time seasonal forecast model. The monthly skill scores are updated when the new
observed station data are available. For instance, the 6-month prediction of precipitation and temperature
issued on June 2018 for July to December 2018 in Southern Sumatera is depicted in Fig. 24.

26
Table 4. The selected statistical downscaling methods and models for precipitation (top) and temperature
(bottom) in Southern Sumatera for 1 to 3-month lead

27
Figure 21. Monthly TCC and NRMSE for precipitation and temperature in Southern Sumatera

28
Figure 22. Examples of the selected predictors in MWR for 1-month lead precipitation in May, August and
September in Southern Sumatera

29
Figure 23. Examples of the selected predictor in MWR-Obs for 1-month lead September precipitation in
Southern Sumatera

30
Figure 24. The 6-month forecast of precipitation and temperature for June to December 2018 in Southern
Sumatera

31
4. Summary
The statistical downscaling model of the seasonal forecast for precipitation and temperature has been
constructed for 10 observed station data from 1983-2012 in Sumatera, Indonesia using four different
statistical downscaling methods (SBC, CIR, MWR, and MWR-Obs) with individual forecast models output
from APCC MME prediction system, climate indices, and reanalysis data as predictors. The quality control of
station data was conducted by filling the missing values with the nearest station data and then, the closest
MERRA-2 gridded data. The clustering of the seasonal precipitation data indicates two climate zones over
Sumatera (Climate Zone North-CZN and Climate Zone South-CZS) which agree with a previous study.

In CZN, for precipitation, SBC has good forecasting skill for all months, except April and November. MWR
has good skill forecast from May to October. There is no good predictor for November. APCC_SCOPS, APCC,
NASA, and PNU are the best predictors for precipitation up to 3-month lead prediction. For temperature,
most of the downscaling methods and most of GCMs performed good forecasting skill for every month. The
MME prediction system has the highest skill scores in August to October (TCC of 0.7 – 0.8) and the lowest in
April and December (TCC of 0.45 – 0.56) for precipitation. For temperature, the skill scores are high for
every month, even with 6-month lead prediction (TCC of 0.6 – 0.85).

In CZS, for precipitation, SBC has also good forecasting skill for all months, except February and April. MWR
has good skill forecast in March, May-June, and August-October. There is no good predictor in February and
April. APCC_SCOPS, APCC, JMA, NASA, and PNU are the best predictors for precipitation up to 3-month lead.
Unlike CZN, MWR is more dominant than SBC for temperature. This is likely due to a decrease in the
number of selected predictor’s GCMs for SBC from February-April to September-November. The MME
prediction system has the highest skill scores in forecasting precipitation in July to September (TCC of 0.7 –
0.8) and the lowest in January and November (TCC of 0.4 – 0.54). For temperature, the skill scores are high
for every month at least 4-month lead prediction (TCC of 0.47 – 0.84).

For the future activities, for months with no predictors, the seasonal forecasting values will be filled by the
APCC MME deterministic model. In addition, a similar project will be conducted for larger area and the
forecasting skill scores of the APCC MME deterministic forecasts will be examined over Indonesia.

32
REFERENCES

Aldrian, E., and Dwi Susanto, R, 2003: Identification of three dominant rainfall regions within Indonesia and
their relationship to sea surface temperature. Int. J. Climatol, 23(12), 1435-1452.
Charles, A. N., Duell, R. E., Wang, X., and Watkins, A. B, 2015: Seasonal Forecasting for Australia using a
Dynamical Model: Improvements in Forecast Skill over the Operational Statistical
Model. Australian Meteorological and Oceanographic Journal, 65(3), 356-375.
Cho, J, 2017: Introduction to downscaling of seasonal forecast using rSForecast. “User-oriented Statistical
Downscaling of Climate Information in Agriculture and Water Resources”. APCC Training
Program, August 21-26, 2017 APEC Climate Center, Busan, Korea
DelSole, T., Nattala, J., and Tippett, M. K, 2014: Skill improvement from increased ensemble size and model
diversity. Geophys. Res. Lett, 41, 7331–7342, https://doi.org/10.1002/2014GL060133
Gelaro, R., and Coauthors, 2017: The Modern-Era Retrospective Analysis for Research and Applications,
Version 2 (MERRA-2). J. Climate, 30, 5419–5454, https://doi.org/10.1175/JCLI-D-16-0758.1
Genolini, C., Ecochard, R., Benghezal, M., Driss, T., Andrieu, S., and Subtil, F, 2016: kmlShape: An
efficient method to cluster longitudinal data (time-series) according to their shapes. PLoS ONE
11(6): e0150738. https://doi.org/10.1371/journal.pone.0150738
Juneng, L., Tangang, F. T., Kang, H., Lee, W. J., and Seng, Y. K, 2010: Statistical downscaling forecasts for
winter monsoon precipitation in Malaysia using multimodel output variables. J. Climate, 23(1), 17-
27, https://doi.org/10.1175/2009JCLI2873.1
Kadarsah, 2010: Aplikasi ROC untuk uji kehandalan model HyBMG. Jurnal Meteorologi dan
Geofisika, 11(1).
Kang, H., An, K-H., Park, C-K., Solis, A. L. S., and Stitthichivapak, K, 2007: Multimodel output statistical
downscaling prediction of precipitation in the Philippines and Thailand. Geophys. Res. Lett. 34,
L15710, https://doi.org/10.1029/2007GL030730
Kang, H., Park, C.-K., Hameed, S. N., and Ashok, K, 2009: Statistical downscaling of precipitation in Korea
using multimodel output variables as predictors. Mon. Weather. Rev, 137(6), 1928–1938.
http://doi.org/10.1175/2008MWR2706.1
Kang, S., Hur, J., and Ahn, J.-B, 2014: Statistical downscaling methods based on APCC multi-model
ensemble for seasonal prediction over South Korea. Int. J. Climatol, 34(14), 3801–3810.
http://doi.org/10.1002/joc.3952
Kim, M.-K., Kim, Y.-H., and Lee, W.-S, 2007: Seasonal prediction of Korean regional climate from
preceding large-scale climate indices. Int. J. Climatol, 27(7), 925–934.
http://doi.org/10.1002/joc.1448
Kim, M.-K., and Kim, Y.-H, 2010: Seasonal prediction of monthly precipitation in China using large-scale
climate indices. Adv. Atmos. Sci, 27(1), 47–59. http://doi.org/10.1007/s00376-009-8014-x
Kirtman, B.P., and Coauthors, 2014: The North American Multimodel Ensemble: phase-1 seasonal-to-
interannual prediction; phase-2 toward developing intraseasonal prediction. B. Am. Meteorol. Soc.
95(4):585-601, https://doi.org/10.1175/BAMS-D-12-00050.1
Komalasari, K. E., Fajariana, Y., Nuraini, T. A., and Anggraeni, R, 2016: Aplikasi metode ensemble mean
untuk meningkatkan reliabilitas prediksi HyBMG. Jurnal Meteorologi dan Geofisika, 17(1).
Kuswanto, H., and Prayoga, G. S, 2015: Calibrating the rainfall forecast of the HyBMG outputs using
Bayesian Model Averaging: a case study. International Journal of Applied Mathematics and
Statistics, 53(6), 122-129.
Min, Y. M., Kryjov, V. N., and Oh, S. M, 2014: Assessment of APCC multimodel ensemble prediction in
seasonal climate forecasting: retrospective (1983–2003) and real-time forecasts (2008–2013). J.
Geophys. Res 199:12132–12150, https://doi.org/10.1002/2014JD022230

33
Min, Y. M., Kryjov, V. N., Oh, S. M., and Lee, H. J, 2017: Skill of real-time operational forecasts with the
APCC multi-model ensemble prediction system during the period 2008–2015. Clim. Dynam, 49
(11-12), 4141-4156, https://doi.org/10.1007/s00382-017-3576-2
Molteni, F., and Coauthors, 2011: The new ECMWF seasonal forecast system (System 4) (p. 49). Reading, U.
K: European Centre for Medium-Range Weather Forecasts.
QGIS Development Team (2009) Quantum GIS Geographic Information System. Open Source Geospatial
Foundation Project. http://qgis.osgeo.org
Saha, S., S. and Coauthors, 2014: The NCEP Climate Forecast System Version 2. J. Climate, 27, 2185–2208,
https://doi.org/10.1175/JCLI-D-12-00823.1
Schepen, A., Wang, Q. J., and Robertson, D. E, 2014: Seasonal forecasts of Australian rainfall through
calibration and bridging of coupled GCM outputs. Mon. Weather. Rev, 142(5), 1758-1770,
https://doi.org/10.1175/MWR-D-13-00248.1
Yang, D., Yang, X. O., Xie, Q., Zhang, Y., Ren, X., and Tang, Y, 2016: Probabilistic versus deterministic
skill in predicting the western North Pacific-East Asian summer monsoon variability with
multimodel ensembles. J. Geophys. Res 121:1079–1103, https://doi.org/10.1002/2015JD023781
Yoo, J.H., Cho, J., Hameed, S., Field, R., Kwan, K.F., and Albar, I, 2016: Toward a fire and haze early
warning system for Southeast Asia. APN Science Bulletin, 13

34

View publication stats

You might also like