Biosci. Biotech. Res. Comm. Special Issue No 1:181-194 (2017)

Analysis of post-processing method for dynamic models

output using network data for the drought in North

West of Iran

Behrooz Sari Sarraf

, Ali Akbar Rasouli

, Majid Habibi Nokhandan

, Sina Samadi Naghab

and Sharareh Malboosi

Professor, Faculty of Geography and Planning, Tabriz, Iran

Associate Professor, Meteorology Research Institute, Tehran, Iran

1,3

PhD student of Climatology, University of Tabriz, Tabriz, Iran,

Expert of Climatological Research Institute, Mashhad, Iran

ABSTRACT

Since long time ago, prediction of precipitation status and investigation of drought hazards in catchment areas of North West of

Iran, due to the critical importance of discharge rate of related catchments for Lake Uromia, has been one of the most important

challenging issues in ef cient management of water resources; management of vast capital of water resources and energy produc-

tion of the country is highly affected by the aforesaid factors. Therefore, application of dynamic methods may play signi cant

role in adjustment of such conditions concerning the frequencies of climate parameters and occurrence of imbalance behaviors

in precipitation pattern of the country. Regarding improper distribution of observed data, this research  rstly completes post-

processing operation using precipitation data of Aphrodite network, and Model Output Statistics(MOS) post-processing methods

on the output of dynamic prediction model MRI-CGCM3 in a 28-year period(1980-2007), the precipitation grid of post-processed

model and upon weighting output climate variables of dynamic model for each cell of data network and also, determining sta-

tistical model coef cients of multivariable correlation; output systematic error of the model highly reduced to be used in small

scale applications. Then, post-processed prediction data of dynamic model were applied for computing Standardized Precipitation

Index (SPI) provided in order to predict drought. Capabilities of selected post-processing method were assessed using evaluation

criteria. Findings showed that application of statistical post-processing on direct output of dynamic model results in developing

the monthly prediction of precipitation up to 29% in selected post-processing method. Accuracy of Standardized Precipitation

Index (SPI) predicting may increase up to 22.3% than no post-processing mode, in a way that this value reaches to 79.5% after

the implementation of post-processing operation.

KEY WORDS: POST-PROCESSING, DROUGHT, DYNAMIC MODELS, SEASONAL PREDICTION

181

ARTICLE INFORMATION:

*Corresponding Author: sisamadi@yahoo.com

Received 27

Nov, 2016

Accepted after revision 26

Dec, 2016

BBRC Print ISSN: 0974-6455

Online ISSN: 2321-4007

Thomson Reuters ISI ESC and Crossref Indexed Journal

NAAS Journal Score 2017: 4.31 Cosmos IF : 4.006

reserved.

Online Contents Available at: http//www.bbrc.in/

182 ANALYSIS OF POST-PROCESSING METHOD FOR DYNAMIC MODELS OUTPUT BIOSCIENCE BIOTECHNOLOGY RESEARCH COMMUNICATIONS

Behrooz Sari Sarraf etal.

INTRODUCTION

Drought has been regarded as a basic parameter in sus-

tainable development issues; it is considered as one of

the most prominent climatic hazards, both in the short

and long-term scale. Concerning that drought is a prom-

inent natural hazard in Iran, and that in the last few

years, various parts of the country have been affected

by this hazard; therefore, it is of particular importance to

perform its assessment, monitoring and prognosis. One

of the methods used for quantifying drought hazards

deals with using drought indices which may be applied

for determining intensity and extension of drought on a

periodical style. So far, these indices have been used for

monitoring drought; but, we may use output of seasonal

predictions to predict such pro les. Seasonal predic-

tions provide some information about long-term aver-

ages. Land surface properties, especially calm variabil-

ity of ocean surface temperature can affect the Earth’s

weather. These effects are not observable in diurnal

scale, but they are observable on a larger time scale of

months and the seasonal averages.

A wide range of studies have been conducted within

and outside the country; but, most studies applied

hydro-climate observed data or climatic indices such

as El Niño–Southern Oscillation (ENSO) and the North

Atlantic Oscillation (NAO) and establishing interrelation

of rainfall and large-scale climate signals and they rarely

have used dynamic methods in the prediction of precipi-

tation. Concerning climatic variations of the recent years

and the occurrence of unusual behaviors in precipitation

pattern of the most parts of the country, application of

dynamic methods bear advantages than the statistical

methods which are solely based on the behaviors of sta-

tistical periods. Nowadays, the most common method

used in international centers for predicting rainfall on

hourly to seasonal time scales is the application of the

numerical dynamic models. The output of the aforesaid

models can be used as input for other applied models.

The main objective for predicting climatic conditions in

dynamic method is predicting the future of the status

of climatic variables according to their current condi-

tions and information and the application of numerical

approximations for dynamic equations. In the dynamic

method, we  rstly provide prediction using a general

circulation model; then, dynamic downscaling will be

done on the desired area using a regional model. General

Circulation Models (GCM) simulate the climate system

with more complexities. Dynamic part includes numeri-

cal schemes which compute large scale atmospheric

transmissions. These transmissions are calculated in a

physical space or a spectrum space.

Nowadays, the outputs of these models are pre-

sented to users by international seasonal predicting

centers. Concerning the relatively large scale output of

these models, ranged from about 0.1 * 0.1 geographical

degrees to 2.5 * 2.5 geographical degrees which results

in lower resolution and more errors in the direct use

of the output of such models. Therefore, their output

has errors especially for the near ground surface vari-

ables including precipitation which requires correction

and post-processing analysis. There are a wide range of

methods used for post-processing analysis of the output

of numerical Prediction models of which we may name

Model Output Statistics (MOS). Application of tech-

niques and special conditions is required for determin-

ing correlation coef cient and effect of each parameter

of output model with the regional climate conditions.

Another problem in using post-processing methods is

lack of spatial and temporal distribution of observed

data to be used in the post-processing analysis of the

output of dynamic model, (Azadi etal., 2011).

In this regard, application of Aphrodite data can

greatly reduce calculation errors and considers proper

distribution of time and place in the category of post-

processing framework (Yatagai , et al., 2012). A wide

range of methods have been used in the post-processing

analysis of the output of dynamic methods; for example.

Babaeian et al. applied linear multivariable regression

method for post-processing the precipitation output of

MRI-CGCM3 model, (Babaeian etal., 2013).

In another study, (Kim et al 2012) performed sea-

sonal prediction of winter in the Northern Hemisphere

using seasonal predicting systems which were recently

updated using ECMWF and NCEP; through the revision

of predicting period (1982-2010), the paper evaluated

coupled seasonal climate prediction systems of ocean-

atmosphere and using ECMWF System 4 (Sys4) and the

National Center for Environmental Prediction (NCEP)

for model (CFSv2); they evaluated analysis with the use

of both data sets (Kim et al., 2012a). Also in another

research, Wilkes (2008) presented seasonal prediction

of temperature parameter in network form on North

America using developed statistical methods based on

the surface temperature data of the water bodies of the

North Paci c. Two time-series including long-term data

(from 1880 to 2007) and short term data (from 1950

to 2007) were tested the surface temperature of water

bodies in terms of Extended Reconstructed Sea Surface

Temperature version 2 (ERSST v.2) grid data with a hori-

zontal resolution of 2 × 2 degrees in statistical models

of Canonical Correlation Analysis (CCA) and Maximum

Covariance Analysis (MCA) and it showed that applica-

tion of long-term data, despite low accuracy of some of

them, may result in major promotion in accuracy of sea-

sonal predictions for winter temperature (Wilks 2008).

Also, Kim et al (2012) applied multivariable linear

regression method to provide seasonal predicts in South

BIOSCIENCE BIOTECHNOLOGY RESEARCH COMMUNICATIONS ANALYSIS OF POST-PROCESSING METHOD FOR DYNAMIC MODELS OUTPUT 183

Behrooz Sari Sarraf etal.

Korea using teleconnection indices. The present study

applied a maximum of  ve predictive variables used for

the multivariable correlation models. The results of Kim

et al showed that monthly correlation coef cients of

temperature varied from 0.42 to 0.65 and for precipita-

tion changes from 0.37 to 0.63. Correction coef cients

for temperature vary from 18% to 42% and for precipi-

tation changes from 14% to 39% (Kim etal., 2012b). In

a research, Lim etal (2009) performed downscaling of

predicted seasonal precipitation (NCEP/CFS) with a reso-

lution of 2.5° to spatial scale of 20km on the South-East

United States, including Florida, Georgia and Alabama

using CSEOF model which is based on statistical down-

scaling (Lim etal., 2009). Jeffery and colleagues (2005)

demonstrated that the use of MOS technique (applica-

tion of statistical methods on the output of dynamic

models) results in development of two-week predictions.

They applied this method on global models of NCEP and

ECMWF. Application of MOS technique on two mod-

els results in development of prediction results higher

than application of MOS technique on one model (Jef-

frey, ,et al., 2005). Krishnamurti and colleagues (2000)

performed a research in connection with the seasonal

and climatic predictions with the use of research based

corrective multi-model predicts.

Finally, statistical weight of each model was deter-

mined using linear multivariable regression. They con-

cluded that the predictions of multiple models have

better performance than single models. The  ndings

showed that the use of statistical methods in post-

processing multi-model predicts can improve multiple

predict system of distinct models (Krishnamurti, T. K.,

et al., 2000). Some of the statistical post-processing

methods do not require long-term data of the model

including neural network method (Fathi et al. 2010,

Hasanzadeh etal. 2012) and Genetic Algorithm method

(Kishtawal 2003) or Kalman  ltering method (Rastgu

et al. 2010,) and the moving average method (Azadi,

et al. 2011, McCollor 2008, Johnson and Swinbank.

2010).

Gene and Renwick (2003) performed Seasonal Pre-

dicting of New Zealand temperature by using linear

multivariable correlation method and the parameters

of temperature, rainfall and water bodies’ surface tem-

perature for Paci c Ocean. (Zheng, Renwick 2003).Quito

etal. (2011) conducted a research on precipitation cli-

mate data of the Middle East using Aphrodite data and

comparing them with the output of MRI-CGCM3 Model.

They found that application of Aphrodite data may

highly increase spatial accuracy of the research; espe-

cially, application of network data instead of station-

based observed data in areas with mountainous con-

ditions may increase ef ciency of the downscaling of

climate models (Kitoh etal., 2011).

In another study conducted by Kasanicky and Kob-

ayashi (2003) evaluated the ef ciency of prediction prob-

abilities and the seasonal predictability of atmosphere

using Atmospheric General Circulation Model (AGCM) at

Japan Meteorological Agency (JMA), which is a global

spectral model with T63 resolution. The results showed

that the probable prediction contrary to de nite predic-

tion related to some seasonal and regional similarities,

such as higher relative ability in winters of the North-

ern hemisphere, East Asia and North America (Kusunoki

Kobayashi, 2003).

Meanwhile in another research, Rasa et al (2012)

developed Aphrodite data by the Research Institute for

Humanity and Nature (RIHN) and the Meteorological

Research Institute |Japan Meteorological Agency (MRI/

JMA) for wet areas and wet adjacent areas of Pakistan

with a Resolution of 0.05 degrees in decade form (Rasu

etal., 2012). Yasutomi etal. (2011) studied the develop-

ment of long-term networked temperatures data series

and its application in the separation of rain / snow at

daily precipitations (Yasutomi etal., 2011)

Therefore, application of post-processing technique

may develop outputs of dynamic models to be used

in subscales and the outputs of these models may be

implemented in macro-environment management with

a more comprehensive approach. Main objective of the

present research consisted of developing the accuracy

of seasonal predictions of precipitation of North-west of

the country using dynamic model output post-process-

ing method used toward managing the drought hazard.

MATERIAL AND METHODS

In this study, we used three data series; the  rst series

is observed data of monthly precipitation obtained

from Meteorological Stations in the North-West of Iran,

including West Azerbaijan and East Azerbaijan and

Ardabil provinces (Figure 1). Selection of stations with

regard to the availability of long-term observed data of

precipitation (1980 - 2007) was according to predict data

and Aphrodite data.

Table 1 shows existing observation stations of the

studied region along with precipitation data of the obser-

vation stations, respectively presented for all seasons of

the year Second Series of Data are Aphrodite Data.

Aphrodite Project was developed in 2006 with the

aim of creating diurnal precipitation data in high res-

olution networks across Asia (Yatagai, et al. 2012). In

the same year, a project named APHRODITE was devel-

oped by the Research Institute for Humanity and Nature

(RIHN) and the Meteorological Research Institute |Japan

Meteorological Agency (MRI/JMA)in order to establish

networked diurnal precipitation databases across Asia

with high spatial resolution and as per the observations

Behrooz Sari Sarraf etal.

184 ANALYSIS OF POST-PROCESSING METHOD FOR DYNAMIC MODELS OUTPUT BIOSCIENCE BIOTECHNOLOGY RESEARCH COMMUNICATIONS

FIGURE 1. The studied region (North West of Iran)

Table 1. Climate statistics of observation stations in the studied region

Province

Station

Precipitation (mm) Total Precipitation (mm)

Precipitation Percentage

(%)

Variation

Range

Variance

Average

Max

Min

Spring

Summer

Autumn

Winter

Spring

Summer

Autumn

Winter

West Azerbaijan

Uremia

312.3 51 2.2 114.2 14.1 91.7 92.3 %37 %5 %29 %30 48.8 307.23

Piranshahr

662.9 106.5 1.6 170.7 7.1 196.8 288.3 %26 %1 %30 %43 104.9 1927

Takab

337.5 60 3.1 125.3 13.3 93 105.9 %37 %4 %28 %31 56.9 395

Khoy 263.2 52.6 6.2 114.9 29.2 64.7 54.4 %44 %11 %25 %21 46.4 163.55

Sardasht

833.5 135.9 1.3 214.4 4.9 253 361.2 %26 %1 %30 %43 134.6 3248.13

Mako

305.1 60.1 12.6 142.2 48.9 62.4 51.6 %47 %16 %20 %17 47.5 214.78

Mahabad 397 63.1 1.2 115.9 6.6 124 150.5 %29 %2 %31 %38 61.9 598.42

East zerbaijan

Ahar

287.6 55.4 6.4 123.8 25.5 75.5 62.8 %43 %9 %26 %22 49 190.77

Tabriz

247.2 46.8 3.4 102.3 13.7 65.6 65.6 %41 %6 %27 %27 33.4 180.56

Sarab

244.4 51 7 107.7 29.1 59.6 48 %44 %12 %24 %20 44 158.97

Mianeh 278.4 48.4 3.1 101.6 14.3 75.7 86.8 %36 %5 %27 %31 45.3 218.64

Ardebil

275.3 34.6 5.1 99.8 22.7 80.5 72.3 %36 %8 %29 %26 38.7 137.74

Pars Abad

274.2 62.2 6.2 91.1 32.6 92.7 57.8 %33 %12 %34 %21 28.4 94.42

Khalkhal 372.8 6.1 144.2 25.7 101.8 101.1 %39 %7 %27 %27 56.1 336.23

made by rain gauge. APHRODITE consisted of an inter-

national cooperation plan for collection and analysis of

rain gauging observed data collected from thousands of

stations across Asia plus the reports provided by World

Meteorological Organization (WMO) which resulted in

providing diurnal precipitation data for 57 years. This

database was provided by ADW interpolation method.

Aphrodite Data consists of integrated observed data of

precipitation in Asia with high-resolution which are

used for evaluating water resources in form of three

separate collections consist of the monsoon regions of

Asia, the Middle East and Russia with the spatial reso-

lution of 0.25 × 0.25 and 0.5 × 0.5 and with diurnal

timescale.

The initial step in using such data is to perform their

veri cation and contrast with the observed data obtained

from meteorological stations which are equipped with

rain gauge stations. Toward this, observed data in the

Behrooz Sari Sarraf etal.

BIOSCIENCE BIOTECHNOLOGY RESEARCH COMMUNICATIONS ANALYSIS OF POST-PROCESSING METHOD FOR DYNAMIC MODELS OUTPUT 185

same period were compared with Aphrodite data on the

synoptic station and rain gauge stations of North West

region of Iran including Ardabil and East Azerbaijan

and West Azerbaijan provinces. Figure 2 shows con gu-

ration of considered region in terms of using Aphrodite

data. The target area has 59 grids with 0.5x0.5 degrees;

where, corresponding Aphrodite data were extracted

through programming and changing format.

The third series of data is retrospective predicts of

rainfall and some meteorological variables affecting it,

such as geo-potential height, thickness of different lay-

ers, ground level pressure and other meteorological vari-

ables of model outputs. Generally, each seasonal pre-

diction model should be performed during each model

development for a 30-year period in order to compare

its results with the observed data. Comparing the predic-

tions of the last 30-year period with observed values,

accuracy of seasonal prediction model will be evaluated

and in terms of verifying the aforesaid model and it will

be used for the issuance of seasonal prediction.

In this research, we applied re-prediction data of MRI-

CGCM model output including 14 variables and general

index of the model output and 6 retrospective predictive

variables related to the network of the studied stations.

MRI-CGCM3 Model (Yukimoto etal., 2012) consists of

two components including atmospheric general circula-

tion model (MRI-AGCM3) and oceanic general circula-

tion model (MRI.COM) where its atmospheric component

is coupled with aerosol model of MASINGAR km

-2

. Cou-

pling intervals or data exchange between atmospheric

and oceanic models is one hour and the same interval

for Aerosol is 0.5 hour. In the atmospheric model of

MRI-AGCM3, atmospheric component of the model is

in spectral form in which the hydrostatic equations are

used as predictors.

The horizontal resolution of Tl159 model (about 120

km) is with 48 vertical layers in ETA coordinate system.

The structure of this model consists of three major com-

ponents, namely: (a) the initial  eld input data of the

model, obtained through the analysis of meteorological

variables, ocean and land surface variables, (b) an inte-

grated prediction system of the atmosphere, oceans and

land and (c) the map products and error analysis and

assessment system. Applied variables and indices with

their explanations are given in Table 2.

Total of the  rst 14 parameters of the output of MRI-

CGCM3 model have been used on a monthly basis at

Tokyo Climate Centre (TCC) for post-processing. They

have been chosen in a way to be appropriate for the

climate of South-East Asia; but, a signi cant number of

them are suitable for our climate, too. In addition to the

aforesaid 14 indices, 6 other variables including H500,

SLP, T2M, T850 and Model-Pr will be extracted from the

model output  les.

Table 2. Output Parameters of MRI-CGCM3 Model

Parameter Variable Parameter Variable Parameter Variable

Z2030

Geopotential height of 500

millibar

NINOSST

Water surface

temperature

H500

Geopotential height of 500

millibar

Z3040

Average Geopotential height of

500 millibar

NINOWESTSST

Water surface

temperature

SLP Mean sea level pressure

Z4050

Geopotential height of 500

millibar

DLRAIN Precipitation SST Water surface temperature

Z5060

Geopotential height of 500

millibar

WIORAIN Precipitation T2M Temperature in 2m

THMD

Thickness between 300 and 850

millibar

SAMOIRAIN Precipitation T850 Temperature in 850milibar

THEX

Thickness between 300 and 850

millibar

WNPRAIN Precipitation Model-Pr Precipitation in Network

WIOSST Water surface temperature MCRAIN Precipitation

FIGURE 2. Regular grids of Aphrodite data with

0.5x0.5 degrees on the studies region

186 ANALYSIS OF POST-PROCESSING METHOD FOR DYNAMIC MODELS OUTPUT BIOSCIENCE BIOTECHNOLOGY RESEARCH COMMUNICATIONS

Behrooz Sari Sarraf etal.

POST-PROCESSING OF PRECIPITATION DATA

AND VERIFYING RESULTS

The method applied in statistical post-processing is lin-

ear multivariable regression method on precipitation

data. Multivariable regression methods may modify both

types of random and systematic errors in model outputs.

The predictability of random error is much more dif-

 cult than systematic errors. In this method, prediction

and observed data are divided into two courses “statis-

tical post-processing model” and “examination”. Multi-

variable regression method is a method of making model

equation from past data series (Shimizukawa etal. 2009).

This method is one of the most powerful ways to

explain the inter-relationship of observed and modeled

variables. The general form of multivariable regression

equation is as follows:

(1)

Where, Y

is dependent variable or predictant and X

is independent variable or predictor. Since, total of 20

seasonal prediction variables are applied for the devel-

opment of multivariable correlation model and some of

which have no signi cant relation with observed precip-

itation of the region; therefore, all predictors (independ-

ent variables) were inserted in model toward omission

of non-effective variables; then, the variable with least

correlation became omitted.

In selection of the  nal variables for adjustment coef-

 cient of R

, balanced adjustment coef cient of R

-Adjust

and F and t statistics are also effective. R

presents the

percentage of variable changes of the predictant using

predictors. R

-Adjust or balanced R

will be used when the

number of independent variables increased. Negative val-

ues of balanced adjustment coef cient are not accepted.

Advantage of linear multivariable correlation method

is that despite the non-intervention of atmosphere phys-

ical processes, a signi cant relationship may be estab-

lished between predictants of the region and predictors

of large-scale atmospheric circulation model output and

then applied its results for downscaling local parameters

including precipitation (Lee, J., Y., 2003).

In this study, the 28-year period of seasonal Predic-

tion model can be divided into two periods of 22-years

and 6-years. Data of the 22years period are used for the

extraction of the precipitation behavior of MRI-CGCM3

Model on the studied network points. This has been per-

formed through determining the variable of prediction

indices with highest correlation with point precipitation

of the network and determining statistical model coef-

 cients of multivariable correlation. Then, the statistical

model obtained from 22 years output of MRI-CGCM3

model and precipitation network data were applied for a

6 years period to predict monthly rainfall. Jump (JMP4)

software was used in this research for the determination

of partial correlation between 20 variable output indices

of MRI-CGCM3 Model with observed precipitation of the

station. Investigations showed that if number of input

variables in multivariable model exceeds from 3, post-

processing errors increase the same and prediction of the

precipitation points of the network increases with the

same trend; therefore, multivariable model was designed

based on 3 input variables.

Also, four evaluation indices of Mean Square Skill

Score (MSSS), Relative Operating Characteristics (ROC),

Mean Bias Error (MBE) and Relative Error (RE) were

applied for investigating capabilities of selected post-

processing method in predicting point precipitation of

studied regional stations network.

Mean Square Skill Score (MSSS) index predicts the

relative accuracy of post-processed model compared

with the actual values of observed data; whereas:

It is necessary to calculate the Mean Square Error (MSE)

of observed data (MSE

) and prediction (MSE

Where  and xi respectively are the ith predicted value

and ith observed value of n data. RMSE

and RMSE

values are obtained respectively for square root of the

mean square error prediction and observed values. In an

accurate prediction, square root of the mean square error

prediction equals 1 (MSSS=1) and in a full incorrect pre-

diction, it equals 0 (MSSS=0). It shows that application

of post-processed model output is more successful in

comparison with climate means (Gheti, 2007).

In addition to the two aforementioned indices, mean bias

error (MBE) and the mean relative error were also applied

in examining the capabilities of post-processing method

which are calculated according to the following formula.

Where Mi and Oi are respectively predicted and observed

values. Relative errors (RE) of predictions are calculated

as follows:

(2)

(3)

(4)

(5)

(6)

Behrooz Sari Sarraf etal.

BIOSCIENCE BIOTECHNOLOGY RESEARCH COMMUNICATIONS ANALYSIS OF POST-PROCESSING METHOD FOR DYNAMIC MODELS OUTPUT 187

On one hand, we may calculate ROC curve and also cor-

rect Hit Rate (HR) prediction indices and False Alarm

Rate (FAR) incorrect prediction indices for each class

of predictions in which the ROC sub-curve area shows

evaluation of the prediction; where, much more closing

to 1 shows higher capability of the model (WMO, 2006).

DROUGHT PREDICTION INDEX (SPI)

Concerning the importance of being aware of drought

status for the future months in planning for agricul-

ture, water resources and environments, we may use

predicted precipitations to compute Drought Prediction

Index (SPI) in monthly and seasonal (3 months) scales

for studied network cells. The reason for this index deals

with the monthly data of Aphrodite network and conse-

quently, in having monthly post-processing data. Since,

SPI index may be calculated only if having monthly

precipitation data; therefore, this index is suggested for

predicting drought. SPI index was presented by McKee

and colleagues to quantify the precipitation and drought

observation. Wide range of applications enables SPI

index to observe drought in short-term scales includ-

ing soil moisture and long-term scales including surface

waters and ground waters (Fattahi etal., 2007). Based on

SPI method, drought period occurs when the SPI is con-

tinuously negative and reaches a value of -1 or less; and

it ends when the SPI is positive, and the cumulative val-

ues of SPI show the magnitude and severity of drought

period and wet periods. The classi cations of SPI values

are shown in Table 3 (Moghaddam 2007).

RESULTS AND DISCUSSION

This research regarded the selected course of study from

1980 to 2007; 70% of which i.e. 1980-2001 was consid-

ered as the test course and providing monthly post-pro-

cessing regression equations and 30% of which i.e. 2002-

2007 was considered as veri cation period. Aphrodite

data networked in 59 network cells of 0.5x0.5 degrees

were extracted for the studied region and validated with

the observed data of regional stations; whereas, results

showed proper accuracy of Aphrodite data which  nally

resulted in their application instead of sparse data of

stations (Figures 3 - 7). It is to be noted that networked

Aphrodite data absolutely increased accuracy of the

study with regard to topographic conditions of the stud-

ied region and dispersion of stations.

In the  rst step and corresponding to the test period,

precipitation prediction was calibrated during the period

of 1980-2001; variables with highest correlation with

Table 3. Classi cation of SPI index as per McKee etal

Classi cation of SPI drought index as per McKee etal (Ensa moghaddam, 2007)

Status SPI Index Status SPI Index

Sever wet ≤2 Relatively Dry -1 to -1.49

Very wet 1.5 to 1.99 Very dry -1.5 to -1.99

Relatively wet 1 to 1.49 Sever dry ≥-2

Near normal -0.99 to +0.99

FIGURE 3. Zoning rainfall data of March; Right: Observed Data and Left: Aphrodite data

Behrooz Sari Sarraf etal.

188 ANALYSIS OF POST-PROCESSING METHOD FOR DYNAMIC MODELS OUTPUT BIOSCIENCE BIOTECHNOLOGY RESEARCH COMMUNICATIONS

FIGURE 6. Zoning rainfall data of December; Right: Observed Data and Left: Aphrodite data

FIGURE 4. Zoning rainfall data of June; Right: Observed Data and Left: Aphrodite data

FIGURE 5. Zoning rainfall data of September; Right: Observed Data and Left: Aphrodite data

Behrooz Sari Sarraf etal.

BIOSCIENCE BIOTECHNOLOGY RESEARCH COMMUNICATIONS ANALYSIS OF POST-PROCESSING METHOD FOR DYNAMIC MODELS OUTPUT 189

monthly precipitations were extracted for each net-

work cell and their monthly post-processed equations

were designed. In next step, gross output of MRICGCM3

Model was amended for each of the network cells using

monthly post-processed equations for each network cell

in the studied area; and prediction of cell precipitation

was extracted for the test period of 2002-2007. Finally

for the accuracy of the performance of results and equa-

tions in future predictions, results of validation period

were validated using actual data. Applying the results of

the precipitation predicted in SPI index, we may study

zoning of drought prediction in the studied region.

Due to the high volume of maps and charts, which

were carried out separately for each cell, 505 cell analy-

sis with 45.25 degrees in longitude and 37.75 degrees

in latitude were presented as samples in this research

which can be generalized to other cells, as well.

Table 4 shows the input parameters for different

months of the year in regarded network cells which were

obtained through multivariable regression equations

and they were determined for the post-processed model

and they were given toward post-processed precipita-

tion model; they are mentioned along with the accu-

racy of the classi ed prediction of monthly precipitation

FIGURE 7. Zoning annual rainfall data; Right: Observed Data and Left: Aphrodite data

Table 4. Prediction model of the parameters of MRI-CGCM3 model toward post-processing network cell rainfall No. 505

during the statistical period of 1980-2007

MonthModel IndexMSSS

Accuracy of

classi ed

prediction %

MBEMRE

Before Post

Process

After Post

Process

Before Post

Process

After Post

Process

JanSst, h2,IOBW Rain

0.7472.2226.45-6.531.150.28

FebMC Rain, IOBW SST, Z4050

0.8888.8918.415.330.59-0.17

MarSST, h2, THEX

0.6150.0014.942.490.390.06

AprWIO SST, Z3040, DL Rain

0.7838.8916.34-21.520.25-0.33

MayH2, SST NNOWEST SST, NNO3

0.5738.8958.6314.171.400.34

JunZ2030, h2, Z4050

0.1683.3373.45-3.5711.510.56

JulSST, Z2030, THTR

0.1366.6789.78-6.4411.94-0.86

AugDL Rain, IOBW Rain, MC,Rain

0.1983.3378.650.9968.39-0.86

SepP850, WNP Rain, TPR

-12.22

66.67108.130.4481.983.30

OctMC Rain, WIO SST, DL Rain

0.40

75.2250.626.841.940.26

NovZ3040, Z4050, NNO WEST

0.6872.2252.29-12.791.350.33

DecELO SST, WIO Rain, ELO Rain

0.30100.0081.694.7712.630.74

Average

67.5955.786.67

76.770.11

Behrooz Sari Sarraf etal.

190 ANALYSIS OF POST-PROCESSING METHOD FOR DYNAMIC MODELS OUTPUT BIOSCIENCE BIOTECHNOLOGY RESEARCH COMMUNICATIONS

FIGURE 8. The post-processing results of MRI-CGCM3 model output during the training and prediction peri-

ods on the Aphrodite network of Grid 505(1980-2007)

FIGURE 9. ROC curve of post-processed precipitation data using multivariable regression method for the

training period (right) and prediction period (left) in an annual scale

and bias before and after post-processing procedure.

According to the above table, the least square error

(MSSS) index, the best capability of post-processing

model was observed in February with a value of 0.88

and the least value was observed in July as 0.13. There-

fore, monthly prediction accuracy average over the year

is 67.59%. Applying statistical bias post-processing and

relative error respectively reduced from 55.78 to 6.76mm

and 76.87 to -0.11.

Figure 8 shows rainfall graphs predicted by the MRI-

CGCM3 model for February (a), May (b), August (c) and

November (d) which include data of the model before

Behrooz Sari Sarraf etal.

BIOSCIENCE BIOTECHNOLOGY RESEARCH COMMUNICATIONS ANALYSIS OF POST-PROCESSING METHOD FOR DYNAMIC MODELS OUTPUT 191

FIGURE 10. Comparison of Aphrodite network rainfall,

raw model output and post-processed output obtained in

multivariable regression method for the network of 505

FIGURE 11. Zoning SPI index for March (winter), right: Aphrodite data and left: post-processed model data

FIGURE 12. Zoning SPI index for June (spring), right: Aphrodite data and left: post-processed model data

post-processing (MRICGCM) and after post-processing

(Train), Aphrodite network rainfall (Observation) and

modeling test period precipitation (Post-processing)

brought for 505 cells as a sample.

Figure 9 shows post-processing precipitation data of

ROC curve obtained in multivariable regression method

for network of 505 during the two periods of training

and prediction periods in annual scale. In this graph,

vertical vector indicates true prediction index and hori-

zontal vector indicates false prediction index. Results

show that highest ef ciency of the model deals with the

time in which precipitation is predicted in normal or a

higher range. It has less accuracy in low precipitation

months.

Figure 10 shows precipitation data average of Aph-

rodite network, raw model data and post-processed data

in multivariable regression method for the network of

505 over the prediction period (years 2002-2007). The

results indicated that there is a signi cant difference

between raw predicted output and post-processed out

of the model; in a way that post-processed model pre-

diction has proper consistency with network Aphrodite

data and this appropriation and consistency have better

results in high precipitation months.

Behrooz Sari Sarraf etal.

192 ANALYSIS OF POST-PROCESSING METHOD FOR DYNAMIC MODELS OUTPUT BIOSCIENCE BIOTECHNOLOGY RESEARCH COMMUNICATIONS

FIGURE 14. Zoning SPI index for December (autumn), right: Aphrodite data and left: post-processed model

data

FIGURE 13. Zoning SPI index for September (summer), right: Aphrodite data and left: post-processed model

data

Table 5. The capability of MRI-CGCM3 model in predicting seasonal SPI indices in the network of 505 during the period

of 2001 to 2007

Network

Accuracy of classi ed prediction %

Correlation of SPI model prediction with the

Observed

Before Post-processing After Post-processing Before Post-processing After Post-processing

505 60.9 78.4 0.068 0.46

Networks Average 61.7 79.5 0.061 0.49

Figures 11 to 14 show zoning of SPI drought index

modi ed using Aphrodite network rainfall and the rainfall

predicted by MRI-CGCM3 model subject to post-processing

process. In the aforesaid  gures, SPI index can be seen for

the three months periods leading to March (winter), June

(spring), September (summer) and December (autumn).

Table 5 indicates the rate of increase in prediction

accuracy of SPI index by the MRI-CGCM3 model after

performing statistical post-processing process in Aph-

rodite network cell of 505 for the studied region. In the

second and third columns of the table, the prediction

accuracy rate of seasonal drought index is given as

per modi ed classi cation of Maki et al.; and in two

last columns, the correlation value of drought indices

before and after performing post-processing process was

inserted on model output.

The above table shows the accuracy of the drought

index prediction in the period 2001 to 2007 for the cells

Behrooz Sari Sarraf etal.

BIOSCIENCE BIOTECHNOLOGY RESEARCH COMMUNICATIONS ANALYSIS OF POST-PROCESSING METHOD FOR DYNAMIC MODELS OUTPUT 193

of network 505 as well as the mean of 59 existing Aph-

rodite networks available in the studied region; in other

words, there are 12 predictions for SPI index of each

year. According to this table, the inter-correlation of SPI

values calculated from not post-processed precipitation

of MRI-CGCM3 model to network 505 was 0.068 that

after post-processing, it reached to 0.46; for networks

average, the amount improved from 0.061 to 0.49. Also,

prediction accuracy of this index according to the clas-

si cation presented in Table 5 for not post-processed

SPI index for the network of 505 samples was 60.9%

and for post-processed SPI was 78.4%; this amount for

the studied networks improved from 61.7% to 79.5%

indicating the promotion of prediction accuracy of this

index valued at 22.3% after performing statistical post-

processing process.

CONCLUSION

Concerning the importance having access to seasonal

predictions and also the prediction of annual conditions

for future months in North West of Iran and especially

for three provinces of West Azerbaijan, East Azerbaijan,

and Ardebil; they mostly cover catchment areas of the

Lake Uremia bearing special importance in this regard.

Also, management of energy and water resources in the

said region is highly affected by climate conditions; this

research tends to apply multivariable regression method

for post-processing the output of the seasonal prediction

of MRI-CGCM3 model on the aforesaid region toward

promoting accuracy of monthly predictions and also,

drought index. Toward realization of this, multivariable

regression method was networked on 20 model indices

and applying Aphrodite data which has less temporal

and spatial errors than observed data of stations. Num-

ber of Aphrodite network points in the aforesaid region

is 59 points with spatial distance of 0.5 geographical

degrees, having highest accuracy after the validating

station based data; therefore, their application for the

studied region with its special topographical conditions

may increase post-processing accuracy.

The applied statistical period of this research is a

28-year period (covering from 1980 to 2007) which

is even in both Aphrodite and model data series. The

above-mentioned period can be divided into two periods

of 22-years used for determining multivariable regres-

sion equations for each point of the network and for

different months of the year and also a 6-years period

for presenting prediction and validating predictions with

actual data. Then, results of the prediction of SPI drought

index were applied. The obtained results were vali-

dated using statistical indices and  ndings showed that

application of multivariable regression method in post-

processing model output excluding spatial range, has

higher accuracy in cold and high precipitation seasons

and less accuracy in low precipitation seasons. Also, the

monthly bias value of the precipitation decreased from

the 67mm for before post-processing to 9mm for after

post-processing. This indicates positive effect of apply-

ing post-processing method on model output.

Finally, correlation value of post-processed and not

post-processed output indices were validated using

results in SPI index which increased from 0.061 to 0.45

for mean network points and the accuracy of classi ed

predictions improved from 61.7% to 79.5% indicating

22.5% promotion from post-processing method in model

output. Findings of the present research indicated that

application of post-processing method on model out-

puts may improve accuracy of results for smaller spa-

tial scales; also, application of Aphrodite network data

instead of station based sparse data may lead to more

improved results. Application of post-processed results

in SPI index may lead to codi cation of a comprehen-

sive model in using drought index in drought prediction

context and it provides seasonal predictions for drought

hazard. This may develop future macro-management in

the  eld of climate and drought.

REFERENCES

Azadi M., Shirgholami M., Hojam S., Sahraeean F. (2011) Post

processing WRF model output for daily precipitation in Iran.

Iran-Water Resources Research 7(4):71-80 (In Persian)

Babaeian I., Karimian M., Modirian R.(2013), Statistical post

processing MRI-CGCM3 output for seasonal precipitation fore-

cast over Khorasan-Razavi province. Iranian Journal of Geo-

physics 7(3):119-133 (In Persian).

Ensa Moghadam T., 2007, An Investigation and assessment of

climatological indices and determination of suitable index for

climatological droughts in the Salt Lake Basin of Iran, 14(2):

271-288 (in Persian)

Fathi Margh A., MH.Mahdeian, (2010), Predicting winter pre-

cipitation using indicators beyond the neural barley yields:

a case study for Isfahan, Iran. International Journal of Ayer

2:451-469 (In Persian)

Fattahi A., y.Sedaghatkerdar,(2007), Curve analysis of time and

severity of drought frequency: Case study in south west sta-

tions of Iran, 5(10):77-90 (in Persian)

Finnis J., W. W.Hsieh, H.Lin, and W.Merry eld, (2011), Nonlin-

ear post-processing of numerical seasonal climate forecasts:

Journal of Atmosphere-Ocean, 26, 1-34.

Gheti R., 2007, Statistical post-processing of dynamical surface

air temperature seasonal predictions using the leading ocean-

forecd spatial patterns, MSc Thesis, McGill University, 87pp

Hasanzadeh Y., A. Abdi Kordani, A.Fakherifard, (2012), fore-

casting drought by using Genetic Algorithm and wavelet-

neural network hybrid model. Journal of Water & Wastewater

3:48-59 (In Persian)

Behrooz Sari Sarraf etal.

194 ANALYSIS OF POST-PROCESSING METHOD FOR DYNAMIC MODELS OUTPUT BIOSCIENCE BIOTECHNOLOGY RESEARCH COMMUNICATIONS

Jeffrey S., X. Wei, 2005, Improving week two forecast with

multi-model re-forecast ensemble. Monthly Weather Rviewi,

pp. 1-25.

Johnson C., Swinbank R., (2009) Medium-range multi-model

ensemble combination and calibration. UK met-of ce Techni-

cal Note 517:31

Kamiguchi K., O. Arakawa, A. Kitoh, A. Yatagai, A. Hamada,

N. Yasutomi, (2010), Development of APHRO_JP, the First

Japanese High-resolution Daily Precipitation Product for More

than 100 Years, Hydrol. Res. Lett., 4, 60–64.

Kim H., P.J. Webster, J.A. Curry, (2012a) Seasonal Prediction

Skill of ECMWF System 4 and NCEP CFSv2 Retrospective

Forecast for the Northern Hemisphere Winter, Climate Dynam-

ics Observational, Theoretical and Computational Research on

the Climate System.

Kim H., P.J. Webster, J.A. Curry, V.E. Toma, (2012b,) Asian

Summer Monsoon Prediction in ECMWF System 4 and NCEP

CFSv2 Retrospective Seasonal Forecasts, Climate Dynamics

Observational, Theoretical and Computational Research on the

Climate System.

Kishtawal C.M., S. Basu, F. Patadia, P.K. Thapliyal, (2003), Fore-

casting Summer Rainfall over India Using Genetic Algorithm,

Geophys. Res. Lett., 30(23), 2203, doi: 10.1029/2003GL018504.

Kitoh A., Arakawa O., (2011), Precipitation Climatology over

the Middle East Simulated by the High-Resolution MRI-

AGCM3, Global Environmental Research, 15/2011: 139-146

Krishnamurti, T. K., C. M., Kishtawal, Z., Zhang, T., Larow, D.,

Bachiochi, E., Williford, (2000), Multimodel ensemble forecasts

for weather and seasonal climate, Journal of Climate, NO. 13,

pp. 4196-4216.

Kusunoki S., C. Kobayashi, (2003), Skill Evaluation of Proba-

bilistic Forecasts by the Atmospheric Seasonal Predictability

Experiments, Journal of the Meteorological Society of Japan,

Vol. 81, No. 1, pp. 85-112, 2003 85.

Lee, J. Y., (2003), Data and Model Description, Assessment of

Potential Seasonal Predictability with a Multi-Model Dynami-

cal- Statistical Ensemble System: Ph.D. thesis, Faculty of the

Graduate School of the Seoul National University.

Lim Y.K., S. Cocke, D.W. Shin, J.T. Schoof, T.E. LaRow, J.J.

O’Brien, (2009), Downscaling Large-Scale NCEP CFS to Resolve

Fine-Scale Seasonal Precipitation and Extremes for the Crop

Growing Seasons over the Southeastern United States, Cli-

mate Dynamics Observational, Theoretical and Computational

Research on the Climate System.

McCollor D., Stull R., (2008) , Hydrological accuracy enhance-

ment via post processing of numerical weather forecasts in

complex terrain, Weather and Forecasting 23:131-144

Rasu A.G.G., T. Mahmood, Q. Zaman, S.B. Cheema, (2012), Val-

idation of APHRODITE Precipitation Data for Humid and Sub

Humid Regions of Pakistan, Pakistan Journal of Meteorology,

Vol.9, Issue 17.

Rastgu Z, Azadi M, Hojam S , (2010), Post processing WRF

model output for level 10- m wind velocity and level 2-m tem-

perature by using Palayeh nonlinear Kalman. Journal of Cli-

mate Research 1(3 and 4):93-106 (In Persian)

Shimizukawa J, Chen C Y, Iba K, (2009), Multi-regression

model for peak load forecast in demand side like university

campus, In: Proc. of the International Conference on Electrical

Engineering, 5-9 July, China

Wilks D.S., 2008, Improved Statistical Seasonal Forecasts

Using Extended Training Data, Int. J. Climatol., 28: 1589-1598.

Yasutomi N., A. Hamada, A. Yatagai, (2011), Development of a

Long-term Daily Gridded Temperature Dataset and its Appli-

cation to Rain/Snow Discrimination of Daily Precipitation,

Global Environmental Research.

Yatagai A., K. Kamiguchi, O. Arakawa, A. Hamada, N. Yasu-

tomi, A. Kitoh, (2012), Constructing a Long-Term Daily Grid-

ded Precipitation Dataset for Asia Based on a Dense Network

of Rain Gauges, American Meteorological Society.

Yukimoto S, Adachi Y, Hosaka M, Sakami T, Yoshimura H, Hira-

bara M, Tanaka TY, Shindo E, Tsujino H, Deushi M, Mizuta R,

Yabu S, Obata A, Nakano H, Koshiro T, Ose T, Kitoh A (2012) A

new global climate model of the meteorological research Insti-

tute: MRI-CGCM3-Model description and basic performance.

Journal of the Meteorological Society of Japan 90:23-64

WMO, (2006), Standardized veri cation system (SVS) for long-

range forecasts (LRF), new attachment to manual of GDPFS

(WMO-No. 485), Vol.1, Fina Report of the Joint Expert Teams

on Long-Range Forecasting, ECMWF, April 2006

Zheng, X., and Renwick J. A., (2003), a regression-based

scheme for seasonal forecasting of New Zealand temperature:

Journal of Climate, 16, 1843-1853.