Professional Documents
Culture Documents
Ecological Indicators
journal homepage: www.elsevier.com/locate/ecolind
A R T I C LE I N FO A B S T R A C T
Keywords: Agriculture productions play significant roles in economic development. Extreme weather events, especially
Precipitation drought under climate change conditions, can affect future crop production. Nowadays, researchers are trying to
SPEI apply modeling approaches for estimating future changes on amounts of crop yields. This study employed pooled
SPI panel data to simulate the most effective meteorological drought indices, economic and meteorological variables
Diurnal temperature
on rainfed wheat yield. The observation period was 1990–2016 for several meteorological data, besides SPI
Durbin Watson test
Chow test
(Standardized Precipitation Index) and SPEI (Standardized Precipitation Evapotranspiration Index) drought
indices in monthly and yearly scales. The available economic variables during the study period were yearly
guaranteed wheat prices (Rial/kg) and area under cultivation (ha). In this research, first, the most effective
variables were selected according to the efficiency criteria and stepwise regression. Then by using pooled panel
data, a relation was estimated between yield and the independent variables. Finally, with future downscaled
variables, the amount of wheat yield was determined for the next 20 years (2019–2038). The GFDL- ESM2M and
MIROC5 models under RCP45 and RCP85 were run, and MIROC5 under RCP45 was selected as the best model,
for the evaluation period. The results revealed that guaranteed wheat prices, yearly precipitation and sunshine
hours, the area under cultivation, and SPI of October were identified as the most effective variables on wheat
yield through the Panel model. By using the projection weather variables and the pooled panel model, we
achieved that the amount of rainfed wheat yield would be increased over two next decades at Mashhad,
Sabzevar, and Torbat H. locations.
1. Introduction 2007). Agriculture in the northeast of Iran has been usually affected by
drought events. Many researchers have assessed the possible impact of
Climate change and global warming have significant impacts on them on crop yields using basic simulation models and statistical
plant growth and crop yield productivity all over the world. One of the methods, and most recent assessments have yielded that arid and semi-
most critical factors causing the yield variability is climate (Chaves, arid regions are highly vulnerable to climate change (IPCC, 2014).
2001), and climate change will shift growing conditions and thus im- In arid and semi-arid areas, such as Iran which average annual
pact future crop production. Agriculture plays a crucial role in eco- precipitation is approximately 250 mm, crops are mostly rainfed, and
nomic development and poverty reduction (World Bank, 2005). Con- the sowing date of plants depends on the rain; hence rainfall variability
siderable attention has been dedicated to agricultural effects of climate is critical for agricultural yields. In this country, rainfed agriculture is
change. As a consequence of climate change, drought usually issues likely to be the most sensitive sector to climate change (Nassiri et al.,
from precipitation deficit. 2006). Wheat and rice are Iran’s two main food staples. Wheat is har-
Drought is a severe and complicated natural hazard that sig- vested on more than 6 million hectares; of these, about 2.170 million
nificantly impacts agricultural yields, and its occurrence will adversely hectares are irrigated, and approximately 3.830 million hectares are
affect wheat yield at the critical growth stages of wheat such as til- rainfed (Agriculture Statistics, 2013).
lering, jointing, anthesis, and grain filling (Xue et al., 2006; IPCC, Wheat has traditionally been one of the three most essential crops
⁎
Corresponding author.
E-mail addresses: salehnia61@gmail.com (N. Salehnia), n.salehnia@um.ac.ir (N. Salehnia).
https://doi.org/10.1016/j.ecolind.2019.105991
Received 22 May 2019; Received in revised form 25 November 2019; Accepted 4 December 2019
Available online 16 December 2019
1470-160X/ © 2019 Published by Elsevier Ltd.
N. Salehnia, et al. Ecological Indicators 111 (2020) 105991
Fig. 1. The study area with three synoptic stations located in Khorasan-e Razavi province, northeastern Iran, Middle East, with an area of 118,851 km2.
(corn, rice, and wheat) used in the world (FAOSTAT, 2010; Zhou, prefectures over the 1977–1996 period, have been examined by Chen
2010). It is crucial to simulate the yield of wheat under various stress and Chang (2005). The results of Chen and Chang (2005) not only il-
conditions like high air temperature and low precipitation to prepare lustrated the sensitivity of crop yield distributions in response to
response strategies to future climate change. The issue of how to in- weather changes in this area but also identified the atmospheric con-
crease wheat yields has been a critical research question for agrono- ditions, which control the yield distributions. The determinants of
mists and farmers. Crop yield is a function of weather variables and wheat output per hectare in Ethiopia during 2011– 2013 by using a
economic factors (Cabas et al., 2010). To investigate the meteorological panel data approach were explored. The model identified specific
and economic variables affecting rainfed wheat yield, different ap- contributors to wheat yields that include farm management techniques,
proaches, models, and methods have been applied. Several simulation weather (e.g., rainfall), water availability, and policy intervention
models have been developed for estimating wheat yield, such as BISm (Mann and Warner, 2017).
model by Snyder et al. (2004), AquaCrop developed by FAO (FAO, This study focuses on determining the most effective meteorological
2012), WOFOST, and DSSAT by Jones et al., 2003. Different meth- and economic variables on rainfed wheat yield. In this regards, three
odologies for simulating impacts of climate change on crop production specific goals were identified: 1) associations between drought indices,
were revealed by White et al. (2011), they reviewed 221 peer-reviewed meteorological-economic variables, and wheat yield in northeast of Iran
papers that used crop simulation models to examine diverse aspects of known as Khorasan-e Razavi province, through pooled panel data, 2)
how climate change might affect agricultural systems. Finally, they comparison of the performance of statistical downscaling method over
could categorize 221 types of researches to six subject areas namely, two CMIP5 models under two RCPs scenarios during the evaluation
target crops and regions; the crop model(s) used and their character- period (2006–2016), then selecting the best CMIP5 model under the
istics; sources and application of data on [CO2] and climate; impact RCP scenario for each station and obtaining the future projections of
parameters evaluated; assessment of variability or risk; and adaptation the selected weather variables, and 3) assessment future wheat yield
strategies. amount over the next decades through the selected CMIP5 model and
Many other studies attempted to determine the most effective the chosen RCP scenario.
weather variables on crop yields, and predicted the amount of crop
yields, under climate change conditions. We have summarized some of 2. Materials and methods
these researches which have implemented in all different countries
around the world (Supplementary File 1). 2.1. Study sites and weather data
The primary model used in this study was Panel data. Panel data
have become widely available in both the developed and developing In this study, we examined the climate and economic variables af-
countries. Panel data refer to a data set containing observations of fecting wheat yield at three major agricultural locations (Mashhad,
various phenomena over multiple periods (Chen and Chang, 2005; Sabzevar, and Torbat Heydarieh (Torbat H.)) in Khorasan-e Razavi
Hsiao, 2014). Unlike regression data, with panel data, we observe many province located in northeastern Iran (Fig. 1). This region has a semi-
subjects over time, and unlike time-series, with panel data we find arid climate and lies between 33° and 38°N latitude and 55° and 61°E
many topics, therefore, it has two dimensions: spatial (cross-sectional) longitude, with an area of 118,851 km2. Table 1 presents the phy-
and temporal (time series) (Darmofal, 2015). This concept is performed siocharacteristics of the study locations, during the baseline weather
in the various field of researches, especially in crop science. Coelli record (1990–2016). The data were collected from the meteorological
(1996) used panel data on rice production in Vietnam to investigate the station at each location. Homogenization and quality control of weather
total factor productivity (TFP) growth. The impact of weather on the data were performed by the national meteorological organization of
yields of seven major crops in Taiwan based on pooled panel data for 15 Iran (www.weather.ir) before the release of such data to users.
2
N. Salehnia, et al. Ecological Indicators 111 (2020) 105991
⎝ x − (7)
⎣ ⎦
SPEI is a meteorological drought index that consists of the pre-
In the last step, with the value of F(x), the SPEI can be estimated as
cipitation data (mm) and the evapotranspiration data (mm/day). The
the standardized values of F(x). The SPEI equation is calculated by Eq.
evapotranspiration rate from a reference surface is called the reference
(8):
crop evapotranspiration or reference evapotranspiration and is denoted
as ETo. We followed the FAO Penman-Monteith approach to calculate C0 + C1 W + C2 W 2
SPEI = W −
the reference evapotranspiration (ETO) (Allen et al., 1998). The ETo 1 + d1 W + d2 W 2 + d3 W 3 (8)
where W = −2Ln (P ) for P ≤ 0.5 and P is the probability of exceeding
Table 2
All the input variables at three stations. a determined D value. The constants are C0 = 2.515517,
C1 = 0.802853, C2 = 0.010328, d1 = 1.432788, d2 = 0.189269 and
Variable Scale Period d3 = 0.001308.
Precipitation (mm) Monthly (May, Jun., and Oct.) 1990–2016 The SPI is widely used as a meteorological drought index (Angelidis
-Yearly et al., 2012). Since continuous long-term data of at least 30 years is
Wind speed (m/s) Monthly (May, Jun., and Oct.) 1990–2016 required to compute SPI (Mckee et al., 1993), therefore we used
Relative Humidity (%) Monthly (May, Jun., and Oct.) 1990–2016 30 years (1987–2016) of precipitation over case study area for calcu-
Mean of temperature (°C) Monthly (May, Jun., and Oct.) 1990–2016
SPI Monthly (May, Jun., and Oct.) 1990–2016
lating it, then we have separated the monthly SPI data over 1990–2016.
DTR (°C) Yearly 1990–2016 SPI is usually computed by fitting the gamma probability distribution
Nhour (hour) Yearly 1990–2016 (Edwards and McKee, 1997) to the observed precipitation data, as
SPEI Yearly 1990–2016 follows:
The area under cultivation (ha) Yearly 1990–2016
Guaranteed wheat prices (Rial/ Yearly 1990–2016 1
G (x ) = x α − 1e−x β (x > 0)
kg) β α Γ(α ) (9)
DTR: Diurnal Temperature Range. where x > 0 is the amount of precipitation, α > 0 is a shape
3
N. Salehnia, et al. Ecological Indicators 111 (2020) 105991
The related information, the details of characteristics of each model, and the processes for downloading original models are available at: http://www.ipcc-data.org/sim/gcm_monthly/AR5/Reference-Archive.html.
tion, and α, β, and Γ(α) computed as following (Eqs. (10) and (11)),
where n is the number of observations:
Horizontal resolution
−
1 ⎛ 4A ⎞ x − ∑ ln(x )
α= ⎜1 + 1+ ⎟, β = , whereA = ln(x ) −
4A ⎝ 3 ⎠ α n (10)
(lat × lon)
1.40*1.40
∞
2.5*2.0
Γ(α ) = ∫0 y α − 1e−ydy (11)
SPI values range from −2 (extremely dry) to +2.0 (extremely wet),
Grid size
256*128
also SPI is in “normal condition” when 0.99 > SPI > −0.99. Details
144*90
about the SPI computation can be found in McKee et al. (1993), and
Guttman (1999). For computing the drought indices, the Drought
Model abbreviation
Monitor and Prediction (DMAP) software tool (AgriMetSoft, 2018a,b)
was used.
GFDL-ESM2M
MIROC5
2.3. GCMs and statistical downscaling method
NOAA GFDL
models. In this study, we used the Delta method (Bennett et al., 2014).
Institute ID
The Delta technique was selected because it is widely used for down-
MIROC
scaling GCMs outputs (Maraun et al. 2010; Themeßl et al. 2011). Delta
method was calculated through Eqs. (12) and (13), as follows:
ables. Parameters β1, β2, β3, and βn are to be estimated, and β0 re-
presents the model intercept. Data for three stations covering the
Modeling center (group)
4
N. Salehnia, et al. Ecological Indicators 111 (2020) 105991
dependent variable, using a set of independent variables (Lobell and period (2006–2016), the prediction performance of the models was
Burke, 2010; Das et al., 2018) (21 variables in Table 2). Linear re- compared based on the mean absolute error (MAE) (MacLean, 2005),
gression is performed by adding or removing independent variables on and Nash–Sutcliffe efficiency (NSE), as defined in Eqs. (21) and (22).
each iteration. Stepwise regression procedure was used for the selection The mean absolute error (MAE) Nash–Sutcliffe efficiency (NSE),
of significant variables (O’Gorman and Woolson, 1991).
n
1
MAE =
n
× ∑ |Oi − Pi|
2.5. Statistical tests i=1 (21)
5
N. Salehnia, et al. Ecological Indicators 111 (2020) 105991
Table 5 two models under two RCPs, while precipitation showed a much lower
The coefficients and statistical criteria from the Panel data model. Dependent Pearson correlation (0.29 up to 0.55). In general, the best model for
Variable: LNY, Method: Panel Least Squares, Sample: 1990 2016, Periods in- Nhour and precipitation was MIROC5 under RCP4.5 with the Delta
cluded: 27, Cross-sections included: 3, Total panel (balanced) observations: 81. method, due to both of them had the best standard deviation (0.99 and
Variable Coefficient Std. Error t-Statistic Prob. 0.94) and RMS (1.4 and 0.73 mm) values compared with the other
cases, respectively.
C 10.65775 5.269710 2.022456 0.04670*
As well, in Sabzevar (Fig. 2(b)) and Torbat H. (Fig. 2(c)), the best
LNP 0.103875 0.038625 2.689336 0.0088**
LNPre 0.574854 0.136955 4.197391 0.0001**
model was MIROC5 under RCP4.5, for both weather variables (Nhour
LNNhour −1.191619 0.641554 −1.857395 0.067*** and precipitation). Indeed, for Nhour values in Sabzevar, the range of
LNCult 0.096387 0.055896 1.724405 0.088*** Pearson correlation was about 0.87–0.90, NSE = 0.57–0.59,
LNSPIOct −0.442108 0.118708 −3.724335 0.0004** MAE = 1.19–1.23 h, RMS = 1.44–1.48 h, and the best model was
LNTmeanOct 0.195143 0.337535 0.578143 0.56412
MIROC5 under RCP4.5. Whereas, for precipitation in the selected
R-squared 0.519420 Mean dependent 5.881358
var model (MIROC5 under RCP4.5) with observed data have presented the
Adjusted R- 0.475589 S.D. dependent var 0.488735 Pearson correlation = 0.40, RMS = 0.62 mm, and MAE = 0.40 mm.
squared For Torbat H. location (Fig. 2(b)), the best criteria’s for precipitation
S.E. of regression 0.373622 Akaike info 0.951312
were consist of NSE = 0.13, RMS = 0.7 mm, MAE = 0.47 mm, and the
criterion
Sum squared resid 10.32992 Schwarz criterion 1.158240
Pearson correlation = 0.56, and for Nhour were NSE = 0.64,
Log-likelihood −31.52813 Hannan-Quinn 1.034334 RMS = 1.52 h, MAE = 1.24 h, and the Pearson correlation = 0.89.
criteria. However, GFDL-ESM2M did not show a good representation (under
F-statistic 10.48166 Durbin-Watson stat 2.042527 both RCPs) of precipitation and Nhour data, in comparison to MIROC5.
Prob (F-statistic) 0.000000
* Significant at 5% level. 3.4. Future projections of precipitation, sunshine hours, and SPI
** Significant at 1% level.
*** Significant at 10% level. As shown in Fig. 3, for tracking the changes of total monthly pre-
cipitation and average monthly Nhour, these variables were plotted
3.2.1. Panel data model results during the recent 20 years (1997–2016) against 20 years in the future
The Eq. (23) developed by using the Panel data model, relative to period (2019–2038). In Fig. 3(a), in Mashhad station, there was not any
this equation, the coefficients are presented in Table 5. The results of remarkable change in Jan. and Feb., but there were noticeable differ-
Table 5 indicate that among of six variables only the last one (Tmea- ences in the other months, as, during summer and fall, the average
noct) did not show a significant relationship with yield (P > 0.05, and amount of Nhour would be decreased in both seasons with −7% and
relative to the significance test for coefficient), in Panel data model. So, −17%, respectively. Whereas during Mar. and spring, the amount of
we deleted this variable for the rest of analyzing. Nhour would increase by at least 14 h (27% and 12%, respectively)
compared to the baseline average Nhour, 20 years later. In the Sabzevar
lnYit = β0 + β1 lncultit + β2 lnPreit + β3 lnPit + β4 lnNhourit + β5 and Torbat H. stations, there was a similar condition in this variable
lnSPIoctit + β6 lnTmeanoctit + Uit i = 1⋯3; t = 1990⋯2016 (Fig. 3(c) and (e)). The whole amount of Nhour’s changes during
2019–2038 against baseline period had very slight variations (around
(23)
0), and it would not be changed in yearly scale, but there was a shift in
The Breusch-Pagan/Cook-Weisberg test for heteroscedasticity was the monthly scale in the amounts of Nhour, over the three stations
revealed that Ho: Constant variance, chi2 (6) = 4.87, and (Fig. 3).
Prob > chi2 = 0.5606. So, in this model, the H0 hypothesis is ac- For total precipitation, during 2019–2038 the amount of pre-
cepted; that is, the variation is identical. As well, the result of the cipitation will increase over spring, end of summer and in Oct., in three
Durbin Watson test (2.0425) was confirmed that the lack of self-cor- stations, with MIROC5 under RCP45 (Fig. 3(b), (d), and (f)). The sea-
relation (Table 5). sonal changes project the most significant increase in spring and Oct.,
Overall, regarding the results of Eq. (23) with increasing the guar- and the smallest in summer. For example, these changes in Apr. will be
anteed wheat prices, precipitation and area under cultivation, the around 56%, 80%, and 64%, respectively for Mashhad, Sabzevar, and
wheat yield would improve, but the increase in the Nhour, and the Torbat H., for 20 future years. Also, in Oct., during 20 years later, there
SPIoct. would decrease the amount of yield. The average temperature of will be incremental changes in total precipitation, while these changes
Oct. also would not have a significant effect on the yield amount. The are 150 mm, 140 mm, and 152 mm, respectively for Mashhad, Sab-
results show that with the assumption of other factors being constant, zevar, and Torbat H. Overall, 10%, 4%, and 8% increase in precipita-
one percent increase in guaranteed wheat prices, yearly precipitation tion relative to the baseline period for Mashhad, Sabzevar, and Torbat
and area under cultivation then wheat yield increased by 0.1%, 0.57%, H., respectively, using MIROC5 under RCP45. In general, the total
and 0.09% respectively. Increased one percent in Nhour and SPIoct yearly amount of precipitation will increase by around 8% in the two
respectively reduces 1.19% and 0.44% in wheat yield. In addition, the future decades in compare to the 20-years of the baseline period. In the
result of F-statistics reveals that the whole regression is significant; as climate of study area (arid and semi-arid), the rainfall in October, April,
this model is a micro model, the adjusted R-squared (0.48) is accep- and May is very beneficial for wheat growth, so through MIROC5′s
table, too. Akaike, Schwarz and Hannan-Quinn criteria represent that projections, it seems that the amount of rain-fed wheat will increase
the lagged form of the variables should be short-run. during 2019–2038.
For the Mashhad station (Fig. 4(a)), the projection IQR, median
3.3. Statistical downscaling and selecting the best model (−0.23) and mean (0.02) of future SPI is lower than the baseline
period, so it seems that the variation and deviation of SPI in Oct during
For selecting the best model and scenario during the evaluation 20 years later will be lower than the baseline period, whereas the ex-
period (2006–2016), a Taylor diagram was drawn. Fig. 2(a) illustrates treme wet and extreme drought conditions will be higher than the
the monthly evaluation of the precipitation and Nhour over Mashhad baseline period. For the Sabzevar station (Fig. 4(b)), the IQR is shrunk
station. There was a high correlation using observed and downscaled and with a strong skewness to the lower values of SPI – according to the
projected data analysis for Nhour (0.87 up to 0.91), the ranges of NSE median- and presents a stronger shift to extreme drier conditions, in
and MAE were 0.75 up to 0.80 h, and 0.97 up to 1 h, respectively for comparison to the baseline period. The median of SPI data in future
6
N. Salehnia, et al. Ecological Indicators 111 (2020) 105991
Fig. 2. Taylor diagram of stations with two GCMs, under RCP4.5 and RCP8.5 with Delta method for precipitation and Nhour for evaluation period, (a) Mashhad, (b)
Sabzevar, and (d) Torbat H.
tends to wet condition than a baseline period, while the mean of SPI period), so in this Section (3.5), the simulated wheat yield was calcu-
tends to the normal condition. The depicted boxplot (Fig. 4(c)) for the lated by the Panel data model’s equation. These results are presented in
Torbat H. station shows the differences between future SPI and baseline Fig. 6 for three stations. The results indicate that in Mashhad station
one. Relative to the boxplots in Fig. 4(c) for future SPI, we can reveal (Fig. 6(a)), the amount of yield would increase in a first and second
that it is shifted to lower values, present a skewness to more negative decade (2019–2028 and 2029–2038) in comparison to the baseline
values, and a significant shrink of the IQR. Finally, in the future period decades (1997–2006 and 2007–2016), and overall mean wheat yields
(2019–2038), the boxplots of the three locations depict that minimum would be increased up to 28% with selected variables. Fig. 6(b) reveals
values of SPI in Oct. move to more negative and drier values of SPI, that in Sabzevar location, the Panel data model forecasts a possible
whereas the variations of IQRs tend to shorter values. It seems that the ~12% reduction in the wheat yield for the first future decade (320 kg/
severe and extreme drought would possibly increase in comparison to ha) relative to the first decade of baseline (360 kg/ha), but there is a
the baseline period. However, the occurrence of wet periods tends to noticeable increment around ~60% for the second decade, through
increase as the median future periods of SPI in Oct., further than zero MIROC5 under RCP45. For Torbat H., there is not any change for the
for three locations (Fig. 4). first future decade vs. baseline one, but in the second decade, there
would be possible an 18% increment for wheat yield amount, in com-
3.5. Simulation wheat yield in future parison to the second baseline decade, with MIROC5 under RCP45.
Consequently, the mean total wheat yields during the future period in
The results show that the Exponential plot has the best outcome, comparison to baseline periods possibly would increase to 30%, 20%,
since R2 = 0.99. Finally, for the future price, the achieved equation in and 10% respectively for Mashhad, Sabzevar, and Torbat H. locations.
Fig. 5 was performed for future years. MIROC5 projected the selected Eyshi Rezaie and Bannayan (2012) obtained similar outputs since they
weather variables under RCP45 (according to the result of evaluation stated that the rainfed wheat would increase 15% during 2010–2039 in
7
N. Salehnia, et al. Ecological Indicators 111 (2020) 105991
Fig. 3. Historical Observations (1997–2016) and downscaled prediction (2019–2038) data of average Nhour and total precipitation period over the three stations.
comparison with the baseline with HadCM3 model under the A2 sce- scenario, we yielded that total precipitation would probably increase in
nario, in Khorasan province. spring and early fall, and gradually decreased Nhour and shift its values
throughout the study region. These changes may lead to an increase in
4. Conclusion rainfed wheat yield, which is demonstrated through the prediction of
rainfed wheat yields by MIROC5 data under RCP45. In conclusion, we
In this study, we have assessed the potential impacts of weather showed that mean total wheat yields during the future period vs.
variables, drought indices, and economic variables on wheat yield at baseline periods would increase to 30%, 20%, and 10% respectively for
three major agricultural locations (Mashhad, Sabzevar, and Torbat H.) Mashhad, Sabzevar, and Torbat H. locations.
in northeastern Iran. We applied different approaches, such as Pooled
Panel data model and Delta statistical downscaling method for pre- CRediT authorship contribution statement
diction rainfed wheat yield. Through the achieved Panel’s equation
with selected downscaled weather data of future, the mean value of the Nasrin Salehnia: Writing - original draft, Methodology,
area under cultivation, and exponential relation for amount guaranteed Conceptualization, Software, Formal analysis, Writing - review &
wheat prices, the amount of future wheat yields were estimated for the editing, Investigation, Resources, Data curation, Validation,
three locations. According to the MIROC5 outputs under RCP45 Visualization. Narges Salehnia: Project administration, Supervision,
8
N. Salehnia, et al. Ecological Indicators 111 (2020) 105991
Fig. 4. The amount of SPI for Oct. during baseline period (1997–2016) versus future period (2019–2038), over three stations. The black horizontal line indicates the
median, and the multiplication sign (×) refers to the mean value.
Fig. 5. The Exponential equation for the amount of Guaranteed Wheat Prices (Rial/kg), during 1990–2017.
9
N. Salehnia, et al. Ecological Indicators 111 (2020) 105991
10
N. Salehnia, et al. Ecological Indicators 111 (2020) 105991
11