Agricultural and Forest Meteorology 275 (2019) 100–113

Contents lists available at ScienceDirect

Agricultural and Forest Meteorology

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

Incorporating machine learning with biophysical model can improve the T

evaluation of climate extremes impacts on wheat yield in south-eastern
Australia
Puyu Fenga,b, Bin Wangb, De Li Liub,c, Cathy Watersd, Qiang Yue,a,f,
⁎

a
School of Life Sciences, Faculty of Science, University of Technology Sydney, PO Box 123, Broadway, NSW 2007, Australia
b
NSW Department of Primary Industries, Wagga Wagga Agricultural Institute, Wagga Wagga, NSW 2650, Australia
c
Climate Change Research Centre and ARC Centre of Excellence for Climate Extremes, University of New South Wales, Sydney, NSW 2052, Australia
d
NSW Department of Primary Industries, Orange Agricultural Institute, NSW 2800, Australia
e
State Key Laboratory of Soil Erosion and Dryland Farming on the Loess Plateau, Northwest A&F University, Yangling 712100, Shaanxi, China
f
College of Resources and Environment, University of Chinese Academy of Science, Beijing 100049, China

ARTICLE INFO ABSTRACT

Keywords: Accurately assessing the impacts of extreme climate events (ECEs) on crop yield can help develop effective
Extreme climate events agronomic practices to deal with climate change impacts. Process-based crop models are useful tools to evaluate
Wheat yield climate change impacts on crop productivity but are usually limited in modelling the effects of ECEs due to over-
APSIM simplification or vague description of certain process and uncertainties in parameterization. In this study, we
Random forest
firstly developed a hybrid model by incorporating the APSIM model outputs and growth stage-specific ECEs
Hybrid model
indicators (i.e. frost, drought and heat stress) into the Random Forest (RF) model, with the multiple linear
regression (MLR) model as a benchmark. The results showed that the APSIM + RF hybrid model could explain
81% of the observed yield variations in the New South Wales wheat belt of south-eastern Australia, which had a
33% improvement in modelling accuracy compared to the APSIM model alone and 19% improvement compared
to the APSIM + MLR hybrid model. Drought events during the grain-filling and vegetative stages and heat events
immediately prior to anthesis were identified as the three most serious ECEs causing yield losses. We then
compared the APSIM + RF hybrid model with the APSIM model to estimate the effects of future climate change
on wheat yield. It was interesting to find that future yield projected from single APSIM model might have a
1–10% overestimation compared to the APSIM + RF hybrid model. The APSIM + RF hybrid model indicated
that we were underestimating the effects of climate change and future yield might be lower than predicted using
single APSIM informed modelling due to lack of adequately accounting for ECEs-induced yield losses. Increasing
heat events around anthesis and grain-filling periods were identified to be major factors causing yield losses in
the future. Therefore, we conclude that including the effects of ECEs on crop yield is necessary to accurately
assess climate change impacts. We expect our proposed hybrid-modelling approach can be applied to other
regions and crops and offer new insights of the effects of ECEs on crop yield.

1. Introduction induced increases of extreme climatic events (ECEs) are recognised as

the major threat to crop production (Trnka et al., 2014; Watson et al.,
As the global population and living standards increase, demand for 2017; Wheeler and Von Braun, 2013). In recent decades, ECEs have
stable foods such as wheat is expected to increase by 60% towards the resulted in increased yield losses around the world (Lesk et al., 2016).
middle of 21st century (Alexandratos and Bruinsma, 2012; Godfray and For example, in south-eastern Australia, drought and co-occurring heat
Toulmin, 2010). Sustainably improving crop production is urgently stress reduced the agricultural Gross National Product by around 30%
needed to meet this demand. However, the ongoing impacts of climate in 1994, 2002 and 2006 (Kirono et al., 2011). Therefore, accurately
change will increase the risk of meeting this demand for crop produc- estimating current and future ECEs-induced yield losses is urgently
tion (Howden et al., 2007; IPCC, 2014). In particular, climate change- needed to assess the sustainability of our agricultural production.

⁎
Corresponding author at: State Key Laboratory of Soil Erosion and Dryland Farming on the Loess Plateau, Northwest A&F University, Yangling 712100, Shaanxi,
China.
E-mail address: yuq@nwafu.edu.cn (Q. Yu).

https://doi.org/10.1016/j.agrformet.2019.05.018
Received 22 November 2018; Received in revised form 15 May 2019; Accepted 20 May 2019
Available online 25 May 2019
0168-1923/ © 2019 Elsevier B.V. All rights reserved.
P. Feng, et al. Agricultural and Forest Meteorology 275 (2019) 100–113

ECEs are usually defined as atypical precipitation, temperature, and learning approach (Shalev-Shwartz and Ben-David, 2014), which
other weather factors compared to their historical distributions (IPCC, usually result in better performance compared to conventional linear
2012). During a crop growth cycle, various ECEs are likely to occur and regression models (Everingham et al., 2015; Feng et al., 2018; Jeong
cause varying degrees of yield losses. In this study, we focused on three et al., 2016). However, a major limitation of statistical models is that
agro-climatic extremes, i.e. drought, heat stress, and frost, which are they usually only provide a simple evaluation of impacts, rather than
widely considered to have great impacts on crop growth and yield provide a deeper understanding of physiological constraints required to
(Guarin et al., 2018). Drought is currently the main constraint to crop inform adaptation strategies (Roberts et al., 2017). Thus, results from
yield in rainfed systems. Drought-induced insufficient water supply can statistical models might sometimes be vague and ambiguous in aiding
adversely impact crop growth in all growing stages, by restraining root targeted development of adaptive practices.
growth and leaf expansion during the vegetative growth stage, and In recent years, the value of combining both process-based crop
limiting photosynthesis, carbon allocation and yield formation during models and statistical models is gaining recognition. Pagani et al.
reproductive stages (Chaves et al., 2002). The relationship between (2017b) incorporated outputs from the sugarcane model Canegro
heat stress (high temperatures above specified thresholds) and yield (Inman-Bamber, 1991) and agro-climatic indicators into multiple linear
losses have also been identified in numerous studies (Innes et al., 2015; regressions to reproduce recorded yield. Their results showed that the
Pagani et al., 2017a; Semenov and Shewry, 2011). This relationship combined model increased prediction accuracy by about 20% com-
could be potentially explained by a number of mechanisms, including pared to each individual model. Everingham et al. (2016) obtained si-
decreased net photosynthesis (Rezaei et al., 2015), increased main- milar higher levels of accuracy by combining APSIM model and random
tenance respiration rates (Innes et al., 2015), and accelerated plant forest (RF) algorithm. Guzmán et al. (2017) combined DSSAT model
development (Stratonovitch and Semenov, 2015). Crops are most vul- (Jones et al., 2003) and support vector regression model for a com-
nerable to heat stress at reproductive stages, especially from anthesis to prehensive assessment of groundwater variability to demonstrate that
maturity. A single, isolated extreme heat event at anthesis can con- the hybrid model performed better in characterizing groundwater
siderably reduce grain yield, but a continuous period of heat stress can variability. In evaluating the effects of ECEs on wheat yield, the existing
lead to almost total yield loss (Porter and Semenov, 2005). As with state-of-art studies are still based on either crop models (Cammarano
drought and heat, frost (low temperatures below specified thresholds) and Tian, 2018) or statistical models (García-León et al., 2019). Our
can also reduce crop growth in all growing periods from the seedling study will firstly use the combination of the two kinds of models, i.e.
stage to harvest. For example, leaves of wheat seedlings are vulnerable crop models and machine learning techniques, to explore new insights
to extreme cold conditions and may wither (Fuller et al., 2007). When of the effects of ECEs on wheat yield.
the wheat inflorescence is forming but prior to flowering, frost events Australian wheat production is crucial to global food security, be-
can result in sterile flowers, decreasing grain number (Barlow et al., cause Australia is one of the world’s major grain exporters (Grundy
2015). et al., 2016). The New South Wales (NSW) wheat belt is the main wheat
Two distinct methods have been widely used to examine climate- production area of south-eastern Australia, accounting for 27% of the
yield relationships, i.e. process-based crop models and statistical national production (www.abares.gov.au, 2013-14). However, inter-
models. Process-based crop models have been developed to account for annual wheat yields in this area are highly variable. Compared to the
the complex interactions between the local environment, the crop long-term mean, up to 1 t·ha−1 yield loss has occurred frequently over
genotype, and management practices (Chenu et al., 2017). In recent the past three decades (http://www.abs.gov.au/Agriculture). There is a
years, crop models have been widely used to characterise the effects of very definite possibility that recurrent extreme events drive the inter-
historical and future ECEs on crop yield in multiple regions around the annual variability in wheat yields (Hughes et al., 2015). Moreover,
world (Cammarano and Tian, 2018; Harrison et al., 2014; Jin et al., drought and heat stress are projected to increase due to the changing
2017; Lobell et al., 2015). While process-based crop models can provide climate (BOM and CSIRO, 2016). In this study, we combined the APSIM
a comprehensive understanding of the timing, frequency and intensity model (Holzworth et al., 2014) and RF algorithm to build a hybrid
of ECEs on crop growth (Watson et al., 2017), they have limitations. model to evaluate impacts of ECEs on wheat yields. The main objectives
Some of the limitations relate to over-simplification or vague descrip- were to 1) develop a hybrid model to reproduce historical observed
tion of certain process and uncertainties in parameterisation, which can wheat yields in the NSW wheat belt, 2) quantify the relative importance
lead to inaccurate results (Eitzinger et al., 2013). These limitations are of growth stage-specific drought, heat, and frost events in determining
especially obvious in simulating ECEs (Barlow et al., 2015). For ex- wheat yields, and 3) compare the yield differences projected by the
ample, heat stress impacts are particularly poorly captured in crop APSIM alone and the hybrid model under future climate change.
models (Fischer, 2011; White and Hoogenboom, 2010). For example,
most crop models simulate the effects of high temperature on leaf se- 2. Materials and method
nescence and stem carbohydrate accumulation and distribution, rather
than directly model damage to reproductive organs and processes. This 2.1. Study sites
raises uncertainty over the application of crop models to properly ac-
count for yield losses resulting from ECEs and the validity in assessing The NSW wheat belt (Fig. 1) is located in south-eastern Australia,
long-term impacts of ECEs under climate change (Schauberger et al., with its western border bounded by the semi-arid interior. It accounts
2017). for nearly 30% of the total areas planted to wheat in Australia (www.
Statistical models use various regression methods to link historical abares.gov.au, 2013-14), making it important in terms of both domestic
yields to historical climate data which are then used to make predic- and international food security (Ray et al., 2015). Generally, wheat is
tions about yields under altered climate conditions (Schlenker and mainly grown under rainfed conditions and the typical growing season
Roberts, 2009). They are easy to handle and relatively easy to compute. is May to November (Gomez-Macpherson and Richards, 1995).
With the increasing availability and improved quality of observed data, The NSW wheat belt is characterized by variable topography and
statistical models usually have a high level of accuracy (Folberth et al., climatic conditions. There is an east-west gradient in both elevation and
2019; Innes et al., 2015). Moreover, newly emerging machine learning precipitation/temperature. The eastern part of the wheat belt consists
algorithms may improve the ability of statistical models to explore of mountains with elevation up to 1100 m and the western areas are
climate-yield relationships (Chlingaryan et al., 2018). Machine learning mainly plains. Average growing season temperature ranges from 8.3 °C
algorithms are capable of disentangling the effects of co-linear climate in the south-east to 17.1 °C in the north-west and the average growing
variables and analysing hierarchical and nonlinear relationships be- season precipitation ranges from 171 mm in the south-west to 763 mm
tween the predictors and the response variable through an ensemble in the south-east in 1961–2000 (Wang et al., 2017a). In addition, the

101
P. Feng, et al. Agricultural and Forest Meteorology 275 (2019) 100–113

experimental records to calibrate and validate our model. A brief de-

scription of the 29 study sites, including climate and annual mean
wheat yield, is shown in Table 1.

2.2. Climate data

Historical (2008–2017) daily climate data (rainfall, maximum and

minimum air temperature, and solar radiation) for the 29 study sites
were downloaded from Scientific Information for Land Owners patched
point dataset (SILO-PPD, https://silo.longpaddock.qld.gov.au/)
(Jeffrey et al., 2001).
Future (2020–2100) climate data were obtained for each of the 29
sites from 34 different global climate models (GCMs, Table 2). The
monthly climate data from these GCMs are provided by different cli-
mate modeling institutions all over the world. Detailed descriptions of
these GCMs can be found at the Coupled Model Inter-comparison Pro-
ject phase 5 (CMIP5, https://cmip.llnl.gov/cmip5/). Generally, raw
GCMs are at coarse temporal (monthly) and spatial (100–300 km grid
spacing) resolutions and therefore cannot be directly used to feed site-
based crop models. Here we used a statistical downscaling method
(NWAI-WG, (Liu and Zuo, 2012)), to downscale the monthly gridded
data simulated by raw GCMs to daily climate data for each of the 29
sites. This approach has been frequently used in recent climate change
Fig. 1. Locations of the 29 study sites in the New South Wales wheat belt in research (Liu et al., 2017; Wang et al., 2015, 2017b). In addition, two
south-eastern Australia.
representative concentration pathways (RCP4.5 and RCP8.5) are
available in the GCMs dataset and were utilized in this study.
climate is characterized by large inter-annual variability mainly due to
El Niño Southern Oscillation (Murphy and Timbal, 2008; Power et al.,
2.3. In-situ trial data
1998). We used 29 sites that are listed in the Grains Research and
Development Corporation National Variety Trials (GRDC-NVT, http://
The GRDC-NVT is a national program of comparative crop variety
www.nvtonline.com.au/) and also located in the wheat belt. These sites
testing with standardized trial management, data generation, collection
are scattered throughout most of the wheat belt (Fig. 1) to represent the
and dissemination. Crop variety trial data from the GRDC-NVT have
range of agro-climatic zones across this area. We used NVT datasets
been frequently used by the scientific community in recent years
because these trials were conducted in recent years and had detailed
(Dreccer et al., 2018; Shen et al., 2018; Zheng et al., 2012). In the

Table 1
A brief description of the 29 study sites used in the study, including location, Soil No.(details at http://www.asris.csiro.au/), GSR (mm), GST (°C), HY, HDR and
AMWY (t ha−1).
ID Site Soil No. GSR GST HY HDR AMWY

1 Beckom 543 298 12.9 2009-2016 – 3.12

2 Bellata 83 307 15.9 2009-2011,2013-2016 – 3.77
3 Bullarah 126 252 17.1 2009,2010,2013,2016 – 3.48
4 Canowindra 703 342 12.6 2008,2011,2013,2014,2017 – 3.73
5 Condobolin 688 290 12.3 2008,2010-2014,2016 2016 2.66
6 Coolah 868 507 10.9 2008-2012,2014-2016 2008-2010 4.82
7 Coonamble 247 267 15.4 2008-2014,2016,2017 – 3.44
8 Galong 545 392 10.9 2008,2010,2011,2013-2016 2016 4.52
9 Gerogery 176 386 12.2 2008-2010,2013,2014,2017 2017 5.15
10 Gilgandra 249 313 14.3 2008-2013,2015-2017 – 3.35
11 Goonumbla 193 331 13.6 2008-2013,2015,2016 – 4.08
12 Lockhart 539 312 12.9 2008,2009,2011-2013,2015 2016 3.53
13 Mayrung 538 251 13.1 2010,2016 – 5.64
14 Merriwa 868 308 13.3 2008-2016 – 3.91
15 Merriwagga 696 233 14.2 2010,2012,2015,2016 – 3.13
16 Narromine 686 289 14.5 2016 – 5.36
17 North Star 237 270 16.3 2010,2011,2016 – 4.16
18 Nyngan 246 253 15.5 2011-2017 – 2.80
19 Oaklands 186 306 12.8 2008-2017 – 3.60
20 Quandialla 693 317 13.1 2008,2010-2016 – 3.91
21 Spring Ridge 127 314 14.4 2008,2012,2013,2015,2016 2009-2012 4.47
22 Temora 913 305 12.2 2010,2011,2013-2016 – 4.16
23 Trangie 683 273 14.7 2010-2013,2015,2017 2008 2.97
24 Tulloona 865 263 16.7 2009-2013,2015,2016 – 3.41
25 Wagga wagga 498 364 12.1 2010-2016 2010-2012 3.52
26 Walgett 1016 239 16.3 2014-2016 – 3.17
27 Willbriggie 697 249 13.7 2010 – 5.66
28 Wongarbon 685 347 13.5 2008-2011,2016 2008 3.54
29 Yenda 697 265 13.8 2015,2016 – 4.22

Note: GSR: growing season rainfall; GST: growing season temperature; HY: harvest year; HDR: heading date record year; AMWY: and annual mean wheat yield.

102
P. Feng, et al. Agricultural and Forest Meteorology 275 (2019) 100–113

Table 2 2.4. APSIM descriptions

List of 34 GCMs under RCP4.5 and RCP8.5 future climate scenarios used in this
study for statistical downscaling outputs of the 29 sites over the New South APSIM (Agricultural Production System sIMulator) version 7.7
Wales wheat belt in south-eastern Australia. Details of the 34 GCMs can be (http://www.apsim.info/) (Holzworth et al., 2014) was used to simu-
found at https://cmip.llnl.gov/cmip5/availability.html.
late historical and future wheat phenology, final biomass and yield at
Model ID Name of GCM Abbr. of GCM Institute ID Country the 29 sites. The APSIM wheat module was developed in Australia and
has been applied to numerous studies across the Australian wheat belt
1 ACCESS1-0 AC1 CSIRO and BoM Australia
(Asseng et al., 2011; Chenu et al., 2013; Lobell et al., 2015). In parti-
2 ACCESS1-3 AC2 CSIRO and BoM Australia
3 BCC-CSM1-1 BC1 BCC China cular, the wheat module has been shown to adequately simulate a
4 BCC-CSM1-1-m BC2 BCC China number of processes at a daily time step, including phenological de-
5 BNU-ESM BNU GCESS China velopment, biomass accumulation, and yield formation, for multiple
6 CanESM2 CaE CCCMA Canada
cultivars, soil moisture and nutrient status, weather conditions, and
7 CCSM4 CCS NCAR USA
8 CESM1-BGC CE1 NSF-DOE-NCAR USA
farming management practices (Jin et al., 2017). Phenological devel-
9 CESM1-CAM5 CE2 NSF-DOE-NCAR USA opment is determined by temperature and cultivar features as men-
10 CESM1-WACCM CE5 NSF-DOE-NCAR USA tioned above. Biomass accumulation is determined by both radiation
11 CMCC-CM CM2 CMCC Europe interception and soil water limitation, while yield formation is calcu-
12 CMCC-CMS CM3 CMCC Europe
lated based on a simple assimilate partitioning rule (Tao et al., 2017).
13 CNRM-CM5 CN1 CNRM-GAME France
14 CSIRO-Mk3-6-0 CSI CSIRO-QCCCE Australia In the APSIM wheat module, several simplified descriptions of cer-
15 EC-EARTH ECE EC-EARTH Europe tain processes have been defined to regulate the effects of ECEs on crop
16 FIO-ESM FIO FIO China growth (Barlow et al., 2015). Frost and heat events are incorporated by
17 GISS-E2-H GE1 NASA GISS USA stress functions which can lead to leaf senescence, while drought (water
18 GISS-E2-H-CC GE2 NASA GISS USA
19 GISS-E2-R GE3 NASA GISS USA
stress) events are defined as functions that can restrain leaf expansion
20 GFDL-CM3 GF2 NOAA GFDL USA and biomass accumulation. Detailed descriptions of these functions
21 GFDL-ESM2G GF3 NOAA GFDL USA have been described by Zheng et al. (2014). It should be noted that
22 GFDL-ESM2M GF4 NOAA GFDL USA these functions are mostly simple and linear. Moreover, many other
23 HadGEM2-AO Ha5 NIMR/KMA Korea
damages of ECEs on crop growth are not considered in APSIM, such as
24 INM-CM4 INC INM Russia
25 IPSL-CM5A-LR IP1 IPSL France frost and heat-induced sterility around anthesis. In other words, APSIM
26 IPSL-CM5A-MR IP2 IPSL France simplifies several processes and ignores some limitations on yields,
27 IPSL-CM5B-LR IP3 IPSL France which may result in poorly modelling ECEs (Barlow et al., 2015).
28 MIROC5 MI2 MIROC Japan
29 MIROC-ESM MI3 MIROC Japan
30 MIROC-ESM-CHEM MI4 MIROC Japan 2.5. APSIM simulations
31 MPI-ESM-LR MP1 MPI-M Germany
32 MRI-CGCM3 MR3 MRI Japan The four varieties used in this study are available in the APSIM
33 NorESM1-M NE1 NCC Norway
variety bank. In addition, there are more than 800 soil profiles in the
34 NorESM1-ME NE2 NCC Norway
APSoil database (Dalgliesh et al., 2006) available for Australian agri-
cultural areas. Most of these soils have already been parameterized for
present study, we used in-situ wheat trial data (2008–2017) for the 29 modelling wheat. We finally selected 28 soil profiles (Table 1) that are
sites across the wheat belt. These data include sowing date, heading geographically closest to study sites. A same soil was used in Coolah
date (only available at several sites and years, Table 1), and yield for and Merriwa sites. Detailed hydraulic properties of all these soils can be
dozens of wheat varieties. In addition, soil nutrient status (including found at http://www.asris.csiro.au/.
total nitrogen, phosphorous, organic carbon, pH, and conductivity) and For the calibration dataset (2008–2017), we set up the APSIM si-
fertilization practice (including date and fertilizer type) were also mulations strictly according to the NVT trial data (variety, sowing date,
available. We chose four varieties, i.e. Suntop, Sunvale, Ventura, and soil nutrient status and fertilization practice). The output yield and
Wallup, which are widely cultivated across the wheat belt and have also phenology data were then directly used to compare with observed data.
been well parameterized on the vernalization and photoperiod response For the climate change impacts (1961–2100), APSIM simulations were
in the APSIM model. The parameters for vernalization sensitivity were driven by the 34 downscaled GCMs for each of the 29 sites. The four
2.0, 2.8, 1.5, and 1.5 for each variety respectively; and for photoperiod varieties were simulated for each site and for each GCM. For the
sensitivity, 3.5, 3.0, 3.0, and 3.5 respectively. In the APSIM-Wheat management options, a sowing window starting on the 1st of May and
module, photoperiod and vernalization are two important factors that ending on the 30th of June was used with the option “must sow”. The
determine wheat phenology (Keating et al., 2003). As a result, the four fertilizer at sowing was 100 kg/ha of urea (equivalent to 46 kg/ha of N)
selected varieties tended to have different dates of growing stages under each future climate scenario. All other options were left as the
which would result in differences in responses to growth stage-specific defaults.
ECEs. Thus, using the four varieties will enable a comprehensive eva- Effects of elevated CO2 concentration were considered in simula-
luation of impacts of various ECEs on wheat yields. In addition, as each tions of future scenarios. In APSIM, CO2 influences plant growth
variety was only available in parts of years, we eventually collected 516 through regulating transpiration efficiency, radiation use efficiency,
trials data. The yield data did not show obvious skewed distribution and critical leaf nitrogen concentration. However, APSIM has no
(Fig. A1) and can be used for regression analysis. In addition, as these function to generate time-varying values of CO2 concentration. Thus,
data were experimental data and very recent, no significant technolo- we added a function to APSIM so that it could calculate yearly atmo-
gical trend was detected after examination. Thus, no de-trending spheric CO2 concentrations through empirical relations of calendar year
method was applied to remove factors (e.g., changes in management (Liu et al., 2014, 2017). For RCP4.5 scenario, the atmospheric CO2
practices, pesticide application) not reproduced by the modelling so- concentration was calculated by:
lution (Pagani et al., 2017b). 0.000075326 × y 0.16276
[CO2]year = 650.18 + 727.97
0.00018747
0.00022299
y2

× (y 2045)3 (1)

103
P. Feng, et al. Agricultural and Forest Meteorology 275 (2019) 100–113

Table 3
List of extreme climate events used in this study. Heat events were calculated at FI, F, and SGF stages. Frost events were calculated at EJ
and FI stages. Drought events were calculated at EJ, FI, F, and SGF stages. Thus, totally 9 weather extreme indicators were used in this
study.
Extreme event Description Growth stage

Heat Number of days with daily maximum temperature > 27 °C FI, F, SGF
Frost Number of days with daily minimum temperature < 0 °C EJ, FI
Drought Number of days with ARID > 0.6 EJ, FI, F, SGF

Note: EJ: end of juvenile; FI: floral initiation; F: flowering; SGF: start of grain filling; ARID: Agricultural Reference Index for Drought.

For RCP8.5, it was fitted by: values and outliers (Elavarasan et al., 2018). Moreover, the RF can
3 approximate functions with both linear and non-linear relations and
267.78 1.6188 × y y 2010
[CO2]year = 1034.3 + 53.342
+ 21.746 × can also identify the relationship between the response and a variable,
4.0143 + 100
y5.2822 which is conditional upon other variables (Hoffman et al., 2018). Given
y 1911 3 that the effects of ECEs on crop yields are often nonlinear (Lobell et al.,
+ 100.65 × 2011; Schlenker and Roberts, 2009), RF is expected to perform well in
100 (2)
assessing the nonlinear relationship. Our previous studies (Feng et al.,
where y is the calendar year from 1900 to 2100 (i.e. y = 1900, 1901, 2018, 2019; Wang et al., 2018) have demonstrated that the RF model
…, 2100). usually performed better than many other machine learning techniques,
in agricultural-based applications. In addition, RF mode is able to
2.6. Climate extremes indicators provide the relative importance of each predictor in determining re-
sponse variable. Therefore, in the present study, we intended to take
In APSIM, wheat cultivation is divided into 11 stages, i.e. sowing, advantage of RF to enhance the ability of APSIM in simulating the ef-
germination, emergence, end of juvenile (EJ), floral initiation (FI), fects of ECEs on wheat yields.
flowering (F), start of grain filling (SGF), end of grain filling, maturity, We also used the MLR to build the hybrid model with the APSIM
harvest rips, and end crop. In this study, we took EJ, FI, F, and SGF into model. MLR is a commonly used regression method to model the linear
consideration, as they represent the four main growing stages. In this relationship between the independent variables and the dependent
study, we intended to evaluate impacts of three kinds of ECEs (Table 3) variable. It is considered to be the extension of ordinary least-squares
at the four main growing stages on wheat yields. The indicators for heat that involves more than one explanatory variable. It is easy to under-
and frost are simple counts of days with maximum/minimum tem- stand and implement, but usually limited in disentangling the nonlinear
peratures above/below fixed thresholds (Tashiro and Wardlaw, 1989; relationships between the predictors and the response.
Zheng et al., 2012). The impact of water deficit was assessed using the
ARID - Agricultural Reference Index for Drought (Woli et al., 2012).
This drought index is a simple, general, soil-plant-atmosphere metric. It 2.8. Hybrid-modelling approach
usually performs better than many other drought indices in agricultural
drought evaluations (Woli et al., 2013). Fig. 2 illustrates the processes of combining the APSIM and RF (or
MLR) models in our study. First, the APSIM wheat module was run to
Ti simulate wheat phenology, biomass and yield based on NVT trial da-
ARIDi = 1
ETo, i (3) tasets. The outputs of phenology date were then used as references for
the calculation of the 9 indicators at the four growth stages. Lastly,
where i represents the ith day, Ti is the transpiration during the ith day
APSIM simulated biomass and the 9 indicators were applied as pre-
(mm d−1), and ETo,i is the reference evapotranspiration on the ith day
dictors in RF (or MLR) for estimating wheat yield. In this study, we
(mm d−1). When calculating ARID, ETo,i is assumed to be equal to
proposed the RF (or MLR) model as an external modification which was
potential evapotranspiration and can be estimated using the Priestley
expected to help improve the performance of APSIM model in simu-
and Taylor (1972) method. Ti is estimated through a macroscopic
lating the effects of growth stage-specific ECEs.
modeling approach which is based on the water content. Detailed de-
We performed the RF model using the R package “randomForest”
scriptions and calculation processes can be found in Woli et al. (2012).
(Liaw and Wiener, 2002). Two parameters were needed to be de-
ARID values fall between 0 and 1. Values higher than 0.6 are usually
termined before the implement of the model, i.e. ntree (the number of
recognized as high water stress, thus we chose 0.6 as the threshold to
trees to grow in the forest) and mtry (the number of randomly selected
evaluate daily drought condition.
predictor variables at each node). We set the ntree as the default values
To calculate these indicators, we first ran APSIM simulations and
of 500. While for mtry, it could affect the model accuracy. As the dataset
obtained wheat phenology information, including dates and duration of
was not large, we adopted a trial and error analysis to determine the
wheat growing stage. Then, according to the phenology information,
value of mtry. Values of 1–10 were tried and 5 was chosen finally as it
stage-specific ECEs indicators were calculated or counted out. After
leaded to a little higher model accuracy. The relative importance of
calculation, we found that heat events rarely happened at the EJ stage
variables was assessed through the “%IncMSE” metric in the RF model.
and frost events rarely happened at the F and SGF stages during the
In addition, the MLR model was performed using the R package “Rattle”
historical period (2008–2017). Thus, we eventually selected 9 extreme
(Williams, 2011).
events at the four growing stages (Table 3).

2.7. Statistical models 2.9. Model performance assessment

RF (random forest) is a nonparametric and ensemble learning al- The NVT trial data (516 trials, 2008–2017) were used to calibrate
gorithm originated from classification and regression trees (Breiman, the models. The output yields from the APSIM model were directly used
2001). It is a nonparametric technique which builds multiple decision to compare with the observed data. While for the RF model, a 10-fold
trees and combines them together to obtain a prediction. Thus, the RF cross validation approach was applied to the 516 data. The coefficient
usually presents good accuracy in spite of the presence of missing of determination (R2) and root mean square error (RMSE) were used for

104
P. Feng, et al. Agricultural and Forest Meteorology 275 (2019) 100–113

Fig. 2. Diagram of the input and output per model for the APSIM + RF (or MLR) hybrid model applied in this study. EJ: end of juvenile; FI: floral initiation; F:
flowering; SGF: start of grain filling; RF: random forest; MLR: multiple linear regression.

model evaluation following: date (Fig. 3). Agreement between simulations and observations was
2 described by the root mean square error (RMSE), the coefficient of
n
i=1
(Oi O¯ )(Pi P¯ ) determination (R2) and the slope of the regression lines. As shown in
R2 = n
(Oi O¯ )2
n
(Pi P¯ ) 2 Fig. 3a, the simulated flowering dates were consistent with observed
i=1 i=1 (4) dates, with an RMSE of 5.01 days (R2 = 0.82, y = 0.72x+76.71,
n P < 0.01), suggesting that the APSIM was able to provide a satisfactory
(Oi Pi )2
RMSE = i=1 estimation of wheat flowering dates. As the flowering stage was gen-
n (5) erally viewed as the indicative stage of the entire wheat phenology, it
Where n is the number of samples, Oi and Pi denote observed and si- was likely that APSIM could also provide fairly good estimations of the
mulated values, and Ō represents the mean of observed values. other three growth stages. This laid a foundation for our subsequent
Generally, the model with higher R2 and lower RMSE is considered to calculation of stage-specific ECEs indicators. For wheat yields, the
be the more accurate model. model was able to explain 61% of the variation and the RMSE was 0.86
t ha−1. This was a common and acceptable result for large-scale crop
3. Results model simulations (Jin et al., 2017), even though the accuracy was not
high. In general, inaccurate simulations were due to the absent or rough
3.1. Model performance assumptions around certain factors, including pest, diseases, and
weather extremes as we mentioned above. In subsequent analysis, we
The performance of the APSIM wheat module was evaluated by managed to increase the accuracy through improving the ability of si-
comparing simulated and observed wheat grain yield and flowering mulating ECEs.

Fig. 3. Comparison of observed and APSIM simulated values of grain yield and flowering date from 2008 to 2017 at the 29 sites across the New South Wales wheat
belt. Totally 516 yield data and 47 flowering date data were collected from National Variety Trials of Australia (see text for more detail about this dataset). Dashed
lines are the 1:1 ratio line. Red lines are the linear regression fit.

105
P. Feng, et al. Agricultural and Forest Meteorology 275 (2019) 100–113

We used the MLR model and the RF model as external modification Fig. 4c shows the time series of the three kinds of wheat yields, i.e.
on the APSIM model and created two hybrid models to predict wheat observed, APSIM simulated, and the two hybrid models simulated, from
yield. Compared to the APSIM model alone, both hybrid models showed 2008 to 2017. The observed yields ranged from 3.0 t ha−1 to 5.2 t ha−1,
higher accuracy in reproducing the observed yields (Fig. 4). The with the greatest inter-annual variability. In general, all models’ si-
APSIM + RF hybrid model explained 81% of the variation in observed mulations successfully captured the temporal pattern of the observed
yields, an increase of 33% compared to the APSIM model. It also re- wheat yields. However, the APSIM model tended to slightly over-
duced the RMSE by 0.32 t ha−1. Moreover, the slope of the regression estimate the yields in almost every year. This was particularly evident
function was close to 1.0, meaning that the APSIM + RF hybrid model in 2008, 2010, and 2013, where an overestimate of up to 0.5 t ha−1
was unbiased in simulation of wheat yields for our study area. While for occurred. These overestimations can be attributed to an underestimate
the APSIM + MLR hybrid model, its model accuracy increased slightly of the effects of ECEs on wheat yields. Through incorporating ECEs
compared to the APSIM model alone and was far below the APSIM + indicators, the APSIM + RF hybrid model succeeded in making the si-
RF hybrid model. Thus, the external modification using the RF model mulated yields more consistent with the observed yields.
with ECEs indicators could greatly improve the performance of the
APSIM model.

Fig. 4. Comparison of observed, APSIM simulated, APSIM + MLR simulated, and APSIM + RF simulated wheat yields from 2008 to 2017 at the 29 sites across the
New South Wales wheat belt. (a) observed vs. APSIM + MLR hybrid model simulated. (b) observed vs. APSIM + RF hybrid model simulated. (c) time series of the
four kinds of yields. In (c), error bars indicate the standard deviation from yields at the 29 sites.

106
P. Feng, et al. Agricultural and Forest Meteorology 275 (2019) 100–113

Fig. 5. Number of extremes climate events occurred from 2008 to 2017. Values for each year were averaged values of the 29 study sites.

3.2. Effects and projected changes of ECEs wheat yield to different ECEs at a same stage were different. For ex-
ample, the F stage was more vulnerable to heat rather than drought,
The historical occurrence of ECEs is shown in Fig. 5. Drought was while the SGF stage was more vulnerable to drought rather than heat.
the most commonly occurring ECE during the historical period. In The shaded area (Fig. 6) denotes the 10th to 90th percentile of each
general, the four growth stages, i.e. EJ, FI, F, and SGF, usually last for variable in the calibration dataset. All lines have a sharp change in the
72–88, 37–48, 6–9, and 25–34 days respectively for the four cultivars. shaded area, meaning that a small change of each ECE may have a large
Thus, nearly one quarter, half, and half of the EJ, F, and SGF stages impact on wheat yield. As shown in the boxplots in Fig. 6, drought
respectively experienced drought conditions. While for heat, it com- events were projected to have small increase in occurrence at the four
monly occurred around and post anthesis. The SGF stage was most study stages, but heat events were likely to increase significantly. In
vulnerable to heat stress and up to two-thirds of this stage might be particular, under RCP8.5, heat events during the end of the 21st century
under heat threat. Frost events mainly occurred during the EJ stage and may double compared to the baseline period. Frost events were likely to
a few frost events also occurred at the FI stage. Fig. 4b and Fig. 5 show decrease, by > 10 days at the EJ stage and by 1-4 days at the FI stage.
that there was an obvious and direct relationship between wheat yields However, the decrease of frost events at the EJ stage might also cause
and the occurrences of ECEs. Wheat yields were much lower during yield reductions. Thus, in general, more yield losses were indicated as a
years with a higher occurrence of ECEs, such as 2012 and 2017. result of changes in future ECEs.
According to the above results, the RF model could potentially im-
prove the performance of the APSIM model in simulating the effects of 3.3. Differences between APSIM projected and the hybrid model projected
ECEs. We then obtained the relative importance (percentage values in future wheat yields
Fig. 6) and the marginal effect (lines in Fig. 6) of each predictor from
the RF model. The trend of the line, rather than the actual values, de- As the RF model performed better in improving the performance of
scribes the nature of the dependence between the response and the the APSIM model compared to MLR. We then used the APSIM model
predictor variables. All ECEs, except frost during the EJ stage, had and the APSIM + RF model to evaluate the impacts of future climate
negative effects on wheat yield. Drought events occurred at the SGF and change on wheat yield in the study area. Projected changes in simulated
EJ stages had high importance values, meaning that they were more wheat yield from the APSIM model and the hybrid model for two of the
harmful to wheat growth. The third was heat events at the FI stage. study sites are shown in Fig. 7 (other sites in Fig. A2). We calculated
Thus, even though heat events were more common during the SGF changes of simulated wheat yields for each site and found that trends
stage (Fig. 5), they tended to cause more yield losses if occurred at the from APSIM-simulated wheat yield differed in different sites. However,
FI stage. Frost events at the EJ stage had positive effects on wheat yield. the APIM + RF hybrid model-simulated yield was projected to decrease
This may be due to that winter wheat plant is capable of withstanding in all study sites. For example, Fig. 7 shows the APSIM-simulated yields
extreme cold before the initiation of flowering (Fowler and Carles, were projected to decrease by 0.5–3% at Bellata but increase by 1.5–3%
1979). Moreover, the plant also requires enough exposure to cool at Mayrung. This might be due to different soil conditions and climate
temperature for jarovization, which will affect subsequent growth and projections at the two sites. However, the APSIM + RF hybrid model-
development (Robertson et al., 1996). In addition, the responses of simulated wheat yields were projected to decrease at both sites.

107
P. Feng, et al. Agricultural and Forest Meteorology 275 (2019) 100–113

Fig. 6. Partial dependence of wheat yield change on extreme climate events and projected changes in each event under RCP4.5 and RCP8.5. The random forest model
could provide partial dependence of the change in the response (blue lines) for selected predictors, when accounting for the average effect of all other driver
predictors. The blue lines are smoothed representations of the response, with fitted values (model predictions) for the calibration data. The trend of the line, rather
than the actual values, describes the nature of the dependence between the response and predictors. The shaded area denotes calibration data between the 10th and
90th percentile. The percentages values denote the relative importance of each predictor generated from the random forest model. The box plots indicate the
occurrences of extreme climate events during the baseline (1961–2000) period and two future periods (2041–2060 and 2081–2100) based on the 34 downscaled
GCMs. Box boundaries indicate the 25th and 75th percentiles across GCMs, whiskers below and above the box indicate the 10th and 90th percentiles. The black lines
within each box indicate the multi-model median. EJ, FI, F, and SGF indicate end of juvenile, floral initiation, flowering, and start of grain filling, respectively.

According to the multi-model ensemble mean values (2041–2060), the model and the hybrid model showed that the APSIM + RF hybrid
APSIM + RF hybrid model-simulated wheat yields were 4% and 3% model is better at reproducing historical wheat yields. Using the RF
lower than the baseline levels at Bellata and Mayrung, respectively. algorithm as an external modification on the APSIM model outputs
Moreover, the yield reductions magnified over time. The differences appears to improve the performance of the individual APSIM model
between the two models were mainly caused by different responses to (Fig. 3 and Fig. 4). Moreover, the RF model also outperformed the MLR
ECEs. Climate change might result in various trends of the APSIM model. Everingham et al. (2016) conducted a similar study through
projections at different sites, but the ECEs changes, especially the in- incorporating the biomass simulated by the APSIM model and several
crease of heat and drought events (Fig. 6), were most likely to reduce climate indices into the RF algorithm to simulate sugarcane yield and
the yields to lower levels. also obtained a high R2 of ˜0.8. The most likely explanation is that the
external statistical model may help improve the performance of the
crop model by simulating the effects of these selected climate indices.
4. Discussion
Using the crop model outputs as predictors may also improve the ability
of the statistical model to consider more agro-climatic processes.
The comparison between the results obtained from the APSIM

108
P. Feng, et al. Agricultural and Forest Meteorology 275 (2019) 100–113

Fig. 7. Projected changes in simulated wheat yield from the APSIM model and the APSIM + RF hybrid model for two of the study sites. Changes were estimated
between two future periods (2041–2060 and 2081–2100) and the baseline period (1961–2000) under RCP4.5 and RCP8.5 based on the 34 downscaled GCMs. Box
boundaries indicate the 25th and 75th percentiles across GCMs, whiskers below and above the box indicate the 10th and 90th percentiles. The black lines and
crosshairs within each box indicate the multi-model median and mean respectively.

Keating and Thorburn (2018) proposed the possible research trend of climate change are needed to be assessed. In our study, we found that
blending mechanistic and empirical statistical models. The approach drought events are expected to have a slight increase, while heat events
outlined in this paper can be viewed a feasible method which may be around anthesis and grain-filling periods will likely be more common.
easily extended to other wheat growing regions to obtain new insights Semenov and Shewry (2011) obtained similar results in a study in
to guide agricultural practices. Europe and demonstrated that wheat plants are expected to suffer more
In recent years, researchers have been concerned with ECEs because heat stress than drought in the future. Given the severe and rapid da-
of their remarkable damage to crops (Lesk et al., 2016; Rezaei et al., mage caused by heat stress around anthesis, a single day increase in
2015). It was reported that nearly one-quarter of all damage and losses heat events is likely to cause great wheat yield losses. On the other
in the agriculture sector are caused by ECEs in many countries (FAO, hand, while fewer frost days are expected because of global warming,
2015). As ECEs may impact during different stages of a crop cycle and we found that frost events during vegetative stages had a positive effect
cause varying degrees of yield losses, it is necessary to identify the most on wheat yield across our study region. Thus, in general, ECEs changes
harmful ECEs in a particular region. In our study, we found that will result in future increased risks for wheat production in the study
drought events during grain-filling and vegetative stages have the area.
greatest adverse effects on wheat yields across a wide range of agro- Given the more unfavorable weather conditions in the future, the
climatic zones. Drought events around anthesis may have relatively possible trend of future wheat yields is frequently discussed among
small effects on wheat yields. One reason is probably that yield for- researchers. The most commonly used method to assess climate impacts
mation mainly occurs at the grain filling stage (Royo et al., 2006). is a combination of process-based crop models and GCMs. For example,
Moreover, the number of drought events occurring around anthesis is Qian et al. (2016) reported that an average increase in wheat yields of
relatively low. A short-term drought event is not able to severely reduce 26–37 % to be expected during 2041–2070 compared to a baseline
the final yield (Clarke et al., 1992). period (1971–2000) in the Canadian Prairies using the CERES-Wheat
In contrast, the impacts of heat and frost events on wheat yield are model and two GCMs. Wang et al. (2017b) conducted a study in Aus-
usually direct and rapid (Barlow et al., 2015). We found heat events tralia using the APSIM model and 11 GCMs to demonstrate that there
around anthesis (FI_Heat and F_Heat), result in a large negative impact would be a decrease in the yield in the south-eastern Australian wheat
on wheat yields, even though they may only occur over a few days in belt throughout the 21st century. The majority of these previous and
the growing season, which is consistent with Balla et al. (2009)’s and similar climate change impacts studies only used one single crop model.
Hays et al. (2007)’s studies. This is mainly because the sensitivity of the However, model comparison studies, especially for climate change
wheat plant to various ECEs varies with different growth stages studies, have emphasized the limited ability of crop models to account
(Hlaváčová et al., 2018). During anthesis, the awns or spikes start to for ECEs. Sánchez et al. (2014) observed that a number of crop models
emerge from the flag leaf as the grain begins to form. However, even a adequately predict mean yields, but are less able to predict extreme low
single frost or heat event may cause sterility and aborted grains, thereby yields, due to their inability to handle ECEs. Hochman et al. (2013)
reducing the number of grains in the inflorescence. While during the reported that the APSIM model poorly accounted for ECEs such as se-
grain-filling period, as grains have formed, the adverse impacts from vere frost and was also overly optimistic about water limited yield
heat events will be reduced (Barlow et al., 2015). impacts in some seasons and locations. Therefore, it is questionable
The ECEs that have occurred during the historical period have al- whether crop model-based projections accurately reflect the direction
ready resulted in large yield losses, so their likely changes induced by and magnitude of the effects of climate change on yield.

109
P. Feng, et al. Agricultural and Forest Meteorology 275 (2019) 100–113

In our study, we used the hybrid model which systematically in- 5. Conclusions
corporated the APSIM model output and ECEs indicators through
random forest algorithm to predict future wheat yields and found that a Impacts of heat, frost, and drought events on current and future
single APSIM model may have 1–10% overestimation for future yield wheat yields were analyzed in this study based on a hybrid model
projections compared to the APSIM + RF hybrid model (Fig. A2). The which incorporated the APSIM model output and climate extremes in-
overestimation was mainly caused by underestimating the ECEs-in- dicators into the Random Forest algorithm. The likely changes of the
duced yield losses. This is likely to be a common phenomenon in crop climate extremes were also discussed. The major conclusions are as
model projections, as the most popular crop models poorly account for follows:
impacts of ECEs (Eitzinger et al., 2013). However, as future ECEs are
projected to increase in most part of the world (IPCC, 2012), previous (1) Drought and pre-anthesis heat events are the major climate ex-
projections based on crop models might overestimate the most likely tremes causing current wheat yield losses in the New South Wales
yield level in the future. Thus, whether an increase or a decrease of wheat belt of south-eastern Australia.
wheat yield is projected in a particular region using crop models, the (2) In general, future climate in the New South Wales wheat belt is
most likely achieved yields may be lower due to the underestimation of expected to be more unfavorable. Drought events are projected to
ECEs-induced yield losses. remain at historical levels, while heat events are projected to in-
Appropriate adaptation strategies can be developed in order to crease in the future. Frost events during the vegetative and pre-
maintain and improve wheat yields in the face of current and future anthesis stages will decrease.
increasing ECEs in the NSW wheat belt. Two kinds of strategies are (3) Future yield projections from conventional process-based crop
frequently discussed among researchers, i.e. minimizing and escaping models might have a 1–10% overestimation because of the under-
the adverse effects of ECEs. In terms of minimizing the adverse effects estimation of climate extremes-induced yield losses.
of ECEs, the main approach is to increase the resistance of crops
through breeding. From our study, the heat tolerance traits will be In addition, the combination of process-based crop models and
important. Stratonovitch and Semenov (2015) also reported heat tol- statistical models showed high performance in modelling extreme cli-
erance around flowering in wheat as a key trait for increased yield mate events and is worthy of consideration for future research. We
potential in Europe under climate change. On the other hand, escaping believe this study would provide some useful information for local
the adverse effects of climate extremes is also a potential approach, farmers and policy makers with respect to development of adaptation
which mainly aims to stagger the reproductive stages to avoid suffering strategies in face of increased climate extremes under climate change.
ECEs. Adjusting sowing date is currently the most effective farming
management practice to avoid negative impacts of ECEs. Optimising Acknowledgments
sowing dates can lead to suitable duration of the pre-anthesis period for
accumulating biomass and suitable flowering and grain-filling windows The first author acknowledges the China Scholarship Council (CSC)
without frost, heat, and terminal drought (Bell et al., 2014). Shortening for the financial support for his PhD study. Facilities for conducting this
the whole growth period may also help avoid suffering terminal heat study were provided by the New South Wales Department of Primary
and drought events, but yield may also decline because of short growth Industries and University of Technology Sydney. Thanks to Grains
length. Therefore, more studies that take into account both breeding Research and Development Corporation (GRDC) National Variety Trials
and selection approaches as well as farming practices are required to (NVT), Australia (http://www.nvtonline.com.au/) for their data sup-
maintain and improve wheat yields in face of increasing ECEs. port. We are grateful to the three anonymous reviewers whose com-
ments have greatly improved this manuscript.

Appendix A

Fig. A1. Distribution of observed wheat yields (2008–2017) for the 29 sites across the New South Wales wheat belt.

110
P. Feng, et al. Agricultural and Forest Meteorology 275 (2019) 100–113

Fig. A2. Projected changes in simulated wheat yield per hectare from the APSIM model and the APSIM + RF hybrid model for the study sites. Changes were
estimated between two future periods (2041–2060 and 2081–2100) and the baseline period (1961–2000) under RCP4.5 and RCP8.5 based on the 34 downscaled
GCMs. Box boundaries indicate the 25th and 75th percentiles across GCMs, whiskers below and above the box indicate the 10th and 90th percentiles. The black lines
and crosshairs within each box indicate the multi-model median and mean respectively.

111
P. Feng, et al. Agricultural and Forest Meteorology 275 (2019) 100–113

References Glob. Chang. Biol. 20 (3), 867–878.

Hays, D.B., Do, J.H., Mason, R.E., Morgan, G., Finlayson, S.A., 2007. Heat stress induced
ethylene production in developing wheat grains induces kernel abortion and in-
Alexandratos, N., Bruinsma, J., 2012. World Agriculture Towards 2030/2050: the 2012 creased maturation in a susceptible cultivar. Plant Sci. 172 (6), 1113–1123.
Revision. ESA Working paper FAO, Rome, pp. 147 No. 12-03. Hlaváčová, M., Klem, K., Rapantová, B., Novotná, K., Urban, O., Hlavinka, P., Smutná, P.,
Asseng, S., Foster, I., Turner, N.C., 2011. The impact of temperature variability on wheat Horáková, V., Škarpa, P., Pohanková, E., Wimmerová, M., 2018. Interactive effects of
yields. Glob. Chang. Biol. 17 (2), 997–1012. high temperature and drought stress during stem elongation, anthesis and early grain
Balla, K., Bencze, S., Janda, T., Veisz, O., 2009. Analysis of heat stress tolerance in winter filling on the yield formation and photosynthesis of winter wheat. Field Crops Res.
wheat. Acta Agron. Hung. 57 (4), 437–444. 221, 182–195.
Barlow, K.M., Christy, B.P., O’Leary, G.J., Riffkin, P.A., Nuttall, J.G., 2015. Simulating the Hochman, Z., Gobbett, D., Holzworth, D., McClelland, T., van Rees, H., Marinoni, O.,
impact of extreme heat and frost events on wheat crop production: a review. Field Garcia, J.N., Horan, H., 2013. Reprint of “quantifying yield gaps in rainfed cropping
Crops Res. 171, 109–119. systems: a case study of wheat in Australia”. Field Crops Res. 143, 65–75.
Bell, L., Lilley, J., Hunt, J., Kirkegaard, J., 2014. Optimising grain yield and grazing Hoffman, A.L., Kemanian, A.R., Forest, C.E., 2018. Analysis of climate signals in the crop
potential of crops across Australia’s high rainfall zone: a simulation analysis. 1. yield record of sub‐Saharan Africa. Glob. Change Biol. 24 (1), 143–157.
Wheat. Crop Pasture Sci. Holzworth, D.P., Huth, N.I., Zurcher, E.J., Herrmann, N.I., McLean, G., Chenu, K., van
BOM and CSIRO, 2016. State of the Climate. Oosterom, E.J., Snow, V., Murphy, C., Moore, A.D., Brown, H., 2014.
Breiman, L., 2001. Random forest. Mach. Learn. 45, 5–32. APSIM–evolution towards a new generation of agricultural systems simulation.
Cammarano, D., Tian, D., 2018. The effects of projected climate and climate extremes on Environ. Model. Softw. 62, 327–350.
a winter and summer crop in the southeast USA. Agric. For. Meteorol. 248, 109–118. Howden, S.M., Soussana, J.F., Tubiello, F.N., Chhetri, N., Dunlop, M., Meinke, H., 2007.
Chaves, M.M., Pereira, J.S., Maroco, J., Rodrigues, M.L., Ricardo, C.P.P., Osório, M.L., Adapting agriculture to climate change. Proc. Natl. Acad. Sci. U. S. A. 104 (50),
Carvalho, I., Faria, T., Pinheiro, C., 2002. How plants cope with water stress in the 19691–19696.
field? Photosynthesis and growth. Ann. Bot. 89 (7), 907–916. Hughes, L., Steffen, W., Rice, M., Pearce, A., 2015. Feeding a hungry nation: climate
Chenu, K., Deihimfard, R., Chapman, S.C., 2013. Large-scale characterization of drought change, food and farming in Australia. Climate Council Aust.
pattern: a continent-wide modelling approach applied to the Australian wheatbelt– Inman-Bamber, N., 1991. A growth model for sugar-cane based on a simple carbon bal-
spatial and temporal trends. New Phytol. 198 (3), 801–820. ance and the CERES-Maize water balance. South Afric. J. Plant Soil 8 (2), 93–99.
Chenu, K., et al., 2017. Contribution of crop models to adaptation in wheat. Trends Plant Innes, P.J., Tan, D.K.Y., Ogtrop, F.V., Amthor, J.S., 2015. Effects of high-temperature
Sci. 22 (6), 472–490. episodes on wheat yields in New South Wales, Australia. Agric. For. Meteorol. 208,
Chlingaryan, A., Sukkarieh, S., Whelan, B., 2018. Machine learning approaches for crop 95–107.
yield prediction and nitrogen status estimation in precision agriculture: a review. IPCC, 2012. Managing the risks of extreme events and disasters to advance climate
Comput. Electron. Agric. 151, 61–69. change adaptation. Special report of the Intergovernmental Panel on Climate Change.
Clarke, J.M., DePauw, R.M., Townley-Smith, T.F., 1992. Evaluation of methods for J. Clin. Endocrinol. Metab. 18 (6), 586–599.
quantification of drought tolerance in wheat. Crop Sci. 32 (3), 723–728. IPCC, 2014. Climate change 2014: synthesis report. In: Pachauri, R.K., Meyer, L.A. (Eds.),
Dalgliesh, N., Wockner, G., Peake, A., 2006. Delivering soil water information to growers Contribution of Working Groups I, II and III to the Fifth Assessment Report of the
and consultants. Proceedings of the 13th Australian Agronomy Conference. Intergovernmental Panel on Climate Change [Core Writing Team. IPCC, Geneva,
Australian Society of Agronomy Perth 10–14. Switzerland, pp. 151.
Dreccer, M.F., Fainges, J., Whish, J., Ogbonnaya, F.C., Sadras, V.O., 2018. Comparison of Jeffrey, S.J., Carter, J.O., Moodie, K.B., Beswick, A.R., 2001. Using spatial interpolation to
sensitive stages of wheat, barley, canola, chickpea and field pea to temperature and construct a comprehensive archive of Australian climate data. Environ. Model. Softw.
water stress across Australia. Agric. For. Meteorol. 248, 275–294. 16 (4), 309–330.
Eitzinger, J., Thaler, S., Schmid, E., Strauss, F., Ferrise, R., Moriondo, M., Bindi, M., Jeong, J.H., Resop, J.P., Mueller, N.D., Fleisher, D.H., Yun, K., Butler, E.E., Timlin, D.J.,
Palosuo, T., Rötter, R., Kersebaum, K.C., Olesen, J.E., 2013. Sensitivities of crop Shim, K.M., Gerber, J.S., Reddy, V.R., Kim, S.H., 2016. Random forests for global and
models to extreme weather conditions during flowering period demonstrated for regional crop yield predictions. PLoS One 11 (6), e0156571.
maize and winter wheat in Austria. J. Agric. Sci. 151 (6), 813–835. Jin, Z., Zhuang, Q., Wang, J., Archontoulis, S.V., Zobel, Z., Kotamarthi, V.R., 2017. The
Elavarasan, D., Vincent, D.R., Sharma, V., Zomaya, A.Y., Srinivasan, K., 2018. Forecasting combined and separate impacts of climate extremes on the current and future US
yield by integrating agrarian factors and machine learning models: a survey. Comput. rainfed maize and soybean production under elevated CO2. Glob. Chang. Biol. 23 (7),
Electron. Agric. 155, 257–282. 2687–2704.
Everingham, Y., Sexton, J., Robson, A., 2015. A statistical approach for identifying im- Jones, J.W., Hoogenboom, G., Porter, C.H., Boote, K.J., Batchelor, W.D., Hunt, L.A.,
portant climatic influences on sugarcane yields. Proceedings of the 37th Annual Wilkens, P.W., Singh, U., Gijsman, A.J., Ritchie, J.T., 2003. The DSSAT cropping
Conference of the Australian Society of Sugar Cane Technologists 37 (37), 8–15. system model. Eur. J. Agron. 18 (3-4), 235–265.
Everingham, Y., Sexton, J., Skocaj, D., Inman-Bamber, G., 2016. Accurate prediction of Keating, B.A., Carberry, P.S., Hammer, G.L., Probert, M.E., Robertson, M.J., Holzworth,
sugarcane yield using a random forest algorithm. Agron. Sustain. Dev. 36 (2). D., Huth, N.I., Hargreaves, J.N., Meinke, H., Hochman, Z., McLean, G., 2003. An
FAO, 2015. The Impact of Natural Hazards and Disasters on Agriculture and Food and overview of APSIM, a model designed for farming systems simulation. Eur. J. Agron.
Nutrition Security – a Call for Action to Build Resilient Livelihoods. Food and 18 (3-4), 267–288.
Agriculture Organization of the United Nations. Keating, B.A., Thorburn, P.J., 2018. Modelling crops and cropping systems—evolving
Feng, P., Wang, B., Liu, D.L., Xing, H., Ji, F., Macadam, I., Ruan, H., Yu, Q., 2018. Impacts purpose, practice and prospects. Eur. J. Agron.
of rainfall extremes on wheat yield in semi-arid cropping systems in eastern Australia. Kirono, D.G.C., Kent, D.M., Hennessy, K.J., Mpelasoka, F., 2011. Characteristics of
Clim. Change 147 (3-4), 555–569. Australian droughts under enhanced greenhouse conditions: results from 14 global
Feng, P., Wang, B., Liu, D.L., Yu, Q., 2019. Machine learning-based integration of re- climate models. J. Arid Environ. 75 (6), 566–575.
motely-sensed drought factors can improve the estimation of agricultural drought in Lesk, C., Rowhani, P., Ramankutty, N., 2016. Influence of extreme weather disasters on
South-Eastern Australia. Agric. Syst. 173, 303–316. global crop production. Nature 529 (7584), 84–87.
Fischer, R., 2011. Wheat physiology: a review of recent developments. Crop Pasture Sci. Liaw, A., Wiener, M., 2002. Classification and regression by randomForest. R news 2 (3),
62 (2), 95–114. 18–22.
Folberth, C., Baklanov, A., Balkovič, J., Skalský, R., Khabarov, N., Obersteiner, M., 2019. Liu, D.L., Anwar, M.R., O’Leary, G., Conyers, M.K., 2014. Managing wheat stubble as an
Spatio-temporal downscaling of gridded crop model yield estimates based on ma- effective approach to sequester soil carbon in a semi-arid environment: spatial
chine learning. Agric. For. Meteorol. 264, 1–15. modelling. Geoderma 214, 50–61.
Fowler, D.B., Carles, R.J., 1979. Growth, development, and cold tolerence of fall-accli- Liu, D.L., Zeleke, K.T., Wang, B., Macadam, I., Scott, F., Martin, R.J., 2017. Crop residue
mated cereal grains 1. Crop Sci. 19 (6), 915–922. incorporation can mitigate negative climate change impacts on crop yield and im-
Fuller, M.P., Fuller, A.M., Kaniouras, S., Christophers, J., Fredericks, T., 2007. The prove water use efficiency in a semiarid environment. Eur. J. Agron. 85, 51–68.
freezing characteristics of wheat at ear emergence. Eur. J. Agron. 26 (4), 435–441. Liu, D.L., Zuo, H., 2012. Statistical downscaling of daily climate variables for climate
García-León, D., Contreras, S., Hunink, J., 2019. Comparison of meteorological and sa- change impact assessment over New South Wales, Australia. Clim. Change 115 (3-4),
tellite-based drought indices as yield predictors of Spanish cereals. Agric. Water 629–666.
Manage. 213, 388–396. Lobell, D.B., Bänziger, M., Magorokosho, C., Vivek, B., 2011. Nonlinear heat effects on
Godfray, H.C.J., Toulmin, C., 2010. Food security: the challenge of feeding 9 billion African maize as evidenced by historical yield trials. Nat. Clim. Change 1 (1), 42.
people. Science 327 (5967), 812–818. Lobell, D.B., Hammer, G.L., Chenu, K., Zheng, B., McLean, G., Chapman, S.C., 2015. The
Gomez-Macpherson, H., Richards, R., 1995. Effect of sowing time on yield and agronomic shifting influence of drought and heat stress for crops in northeast Australia. Glob.
characteristics of wheat in south-eastern Australia. Aust. J. Agric. Res. 46 (7), Chang. Biol. 21 (11), 4115–4127.
1381–1399. Murphy, B.F., Timbal, B., 2008. A review of recent climate variability and climate change
Grundy, M.J., et al., 2016. Scenarios for Australian agricultural production and land use in southeastern Australia. Int. J. Climatol. 28 (7), 859–879.
to 2050. Agric. Syst. 142, 70–83. Pagani, V., Guarneri, T., Fumagalli, D., Movedi, E., Testi, L., Klein, T., Calanca, P.,
Guarin, J.R., Asseng, S., Martre, P., Bliznyuk, N., 2018. Testing a crop model with extreme Villalobos, F., Lopez-Bernal, A., Niemeyer, S., Bellocchi, G., 2017a. Improving cereal
low yields from historical district records. Field Crops Res. yield forecasts in Europe–the impact of weather extremes. Eur. J. Agron. 89, 97–106.
Guzmán, S.M., Paz, J.O., Tagert, M.L.M., Mercer, A.E., Pote, J.W., 2017. An integrated Pagani, V., Stella, T., Guarneri, T., Finotto, G., Van den Berg, M., Marin, F.R., Acutis, M.,
SVR and crop model to estimate the impacts of irrigation on daily groundwater levels. Confalonieri, R., 2017b. Forecasting sugarcane yields using agro-climatic indicators
Agric. Syst. and Canegro model: a case study in the main production region in Brazil. Agric. Syst.
Harrison, M.T., Tardieu, F., Dong, Z., Messina, C.D., Hammer, G.L., 2014. Characterizing 154, 45–52.
drought stress and trait influence on maize yield under current and future conditions. Porter, J., Semenov, M., 2005. Crop responses to climatic variation. Philos. Trans. Biol.

112
P. Feng, et al. Agricultural and Forest Meteorology 275 (2019) 100–113

Sci. 360 (360), 2021–2035. J. Exp. Bot. 66 (12).

Power, S., Tseitkin, F., Torok, S., Lavery, B., Dahni, R., McAvaney, B., 1998. Australian Tao, F., Rötter, R.P., Palosuo, T., Díaz-Ambrona, C.G.H., Mínguez, M.I., Semenov, M.A.,
temperature, Australian rainfall and the Southern Oscillation, 1910-1992: coherent Kersebaum, K.C., Nendel, C., Cammarano, D., Hoffmann, H., Ewert, F., 2017.
variability and recent changes. Aust. Meteorol. Magaz. 47 (2), 85–101. Designing future barley ideotypes using a crop model ensemble. Eur. J. Agron. 82,
Priestley, C., Taylor, R., 1972. On the assessment of surface heat flux and evaporation 144–162.
using large-scale parameters. Mon Weather Rev 100 (2), 81–92. Tashiro, T., Wardlaw, I.F., 1989. A comparison of the effect of high temperature on grain
Qian, B., De Jong, R., Huffman, T., Wang, H., Yang, J., 2016. Projecting yield changes of development in wheat and rice. Ann. Bot.-Lond. 64 (1), 59–65.
spring wheat under future climate scenarios on the Canadian Prairies. Theor. Appl. Trnka, M., Rötter, R.P., Ruiz-Ramos, M., Kersebaum, K.C., Olesen, J.E., Žalud, Z.,
Climatol. 123 (3-4), 651–669. Semenov, M.A., 2014. Adverse weather conditions for European wheat production
Ray, D.K., Gerber, J.S., Macdonald, G.K., West, P.C., 2015. Climate variation explains a will become more frequent with climate change. Nat. Clim. Change 4 (7), 637.
third of global crop yield variability. Nat. Commun. 6 (5989) 5989-5989. Wang, B., Liu, D.L., Asseng, S., Macadam, I., Yang, X., Yu, Q., 2017a. Spatiotemporal
Rezaei, E.E., Webber, H., Gaiser, T., Naab, J., Ewert, F., 2015. Heat stress in cereals: changes in wheat phenology, yield and water use efficiency under the CMIP5 mul-
mechanisms and modelling. Eur. J. Agron. 64, 98–113. timodel ensemble projections in eastern Australia. Clim. Res. 72 (2), 83–99.
Roberts, M.J., Braun, N.O., Sinclair, T.R., Lobell, D.B., Schlenker, W., 2017. Comparing Wang, B., Liu, D.L., Asseng, S., Macadam, I., Yu, Q., 2015. Impact of climate change on
and combining process-based crop models and statistical models with some im- wheat flowering time in eastern Australia. Agric. For. Meteorol. 209, 11–21.
plications for climate change. Environ. Res. Lett. 12 (9), 095010. Wang, B., Liu, D.L., O’leary, G.J., Asseng, S., Macadam, I., Lines‐Kelly, R., Yang, X., Clark,
Robertson, M.J., Brooking, I.R., Ritchie, J.T., 1996. Temperature response of vernaliza- A., Crean, J., Sides, T., Xing, H., 2017b. Australian wheat production expected to
tion in wheat: modelling the effect on the final number of mainstem leaves. Ann. Bot. decrease by the late 21st century. Glob. Change Biol. 24 (6), 2403–2415.
78 (3), 371–381. Wang, B., Waters, C., Orgill, S., Cowie, A., Clark, A., Liu, D.L., Simpson, M., McGowen, I.,
Royo, C., Villegas, D., Rharrabti, Y., Blanco, R., Martos, V., García del Moral, L., 2006. Sides, T., 2018. Estimating soil organic carbon stocks using different modelling
Grain growth and yield formation of durum wheat grown at contrasting latitudes and techniques in the semi-arid rangelands of eastern Australia. Ecol. Indic. 88, 425–438.
water regimes in a Mediterranean environment. Cereal Res. Commun. 34 (2-3), Watson, J., Zheng, B., Chapman, S., Chenu, K., 2017. Projected impact of future climate
1021–1028. on water-stress patterns across the Australian wheatbelt. J. Exp. Bot. 68 (21-22),
Sánchez, B., Rasmussen, A., Porter, J.R., 2014. Temperatures and the growth and de- 5907–5921.
velopment of maize and rice: a review. Glob. Chang. Biol. 20 (2), 408–417. Wheeler, T., Von Braun, J., 2013. Climate change impacts on global food security. Science
Schauberger, B., Archontoulis, S., Arneth, A., Balkovic, J., Ciais, P., Deryng, D., Elliott, J., 341 (6145), 508–513.
Folberth, C., Khabarov, N., Müller, C., Pugh, T.A., 2017. Consistent negative response White, J.W., Hoogenboom, G., 2010. Crop Response to Climate: Ecophysiological Models,
of US crops to high temperatures in observations and crop models. Nat. Commun. 8, Climate Change and Food Security. Springer, pp. 59–83.
13931. Williams, G., 2011. Data mining with Rattle and R: the art of excavating data for
Schlenker, W., Roberts, M.J., 2009. Nonlinear temperature effects indicate severe da- knowledge discovery. Springer Science & Business Media.
mages to US crop yields under climate change. Proc. Natl. Acad. Sci. U.S.A. 106 (37), Woli, P., Jones, J.W., Ingram, K.T., 2013. Assessing the Agricultural Reference Index for
15594–15598. Drought (ARID) using uncertainty and sensitivity analyses. Agron. J. 105 (1),
Semenov, M.A., Shewry, P.R., 2011. Modelling predicts that heat stress, not drought, will 150–160.
increase vulnerability of wheat in Europe. Sci. Rep. 1, 66. Woli, P., Jones, J.W., Ingram, K.T., Fraisse, C.W., 2012. Agricultural reference index for
Shalev-Shwartz, S., Ben-David, S., 2014. Understanding Machine Learning: From Theory drought (ARID). Agron. J. 104 (2), 287–300.
to Algorithms. Cambridge University Press. Zheng, B., Chenu, K., Doherty, A., Chapman, S., 2014. The APSIM-Wheat Module (7.5
Shen, J., Huete, A., Tran, N.N., Devadas, R., Ma, X., Eamus, D., Yu, Q., 2018. Diverse R3008). (accessed May 2018). http://www.apsim.info/Portals/0/Documentation/
sensitivity of winter crops over the growing season to climate and land surface Crops/WheatDocumentation.pdf.
temperature across the rainfed cropland-belt of eastern Australia. Agric. Ecosyst. Zheng, B., Chenu, K., Fernanda, D.M., Chapman, S.C., 2012. Breeding for the future: what
Environ. 254, 99–110. are the potential impacts of future frost and heat events on sowing and flowering time
Stratonovitch, P., Semenov, M.A., 2015. Heat tolerance around flowering in wheat requirements for Australian bread wheat (Triticum aestivium) varieties? Glob.
identified as a key trait for increased yield potential in Europe under climate change. Chang. Biol. 18 (9), 2899–2914.

113