Performance of four crop model for simulations of wheat phenology, leaf growth, biomass and yield across planting dates

Robustness of four wheat simulation model were tested with 2-year field experiments of three cultivars across a wide range of sowing dates in two different climatic regions: Faisalabad (semi-arid) and Layyah (arid), in Punjab-Pakistan. Wheat growing season temperature ranged from -0.1°C to 43°C. The wide series of sowing dates was a unique opportunity to grow the wheat in an environment which temperatures varies from -0.1°C to 43°C. The CERES-Wheat, Nwheat, CROPSIM-Wheat and APSIM-Wheat model were calibrated against the least-stressed treatment for each wheat cultivar. Overall, the four models described performance of early, optimum and late sown wheat well, but poorly described yields of very late planting dates with associated high temperatures during grain filling. The poor accuracy of simulations of yield for extreme planting dates point to the need to improve the accuracy of model simulations at the high end of the growing temperature range, especially given the expected future increases in growing season temperature. Improvement in simulation of maximum leaf area index of wheat for all models is needed. APSIM-Wheat only poorly simulated days to maturity of very and extremely late sown wheat compared to other models. Overall, there is a need of improvement in function of models to response high temperature.


Introduction
Hot and dry region are expected to be particularly vulnerable to climate change associated yield losses associated with increased temperature. Temperature and heat fluctuation negatively affects the morphological, physiological and yield contributing factors of crops. For example, an increase in temperature at flowering stage may cause pollen sterility in crops [1]. Increasing temperature affects wheat growth and development, resulting in smaller grains at 25-35˚C due to shorter grain filling duration and reduced photosynthetic efficiencies at temperatures above 30˚C. Asseng et al. [2] reported 6% reduction of global wheat yield with 1˚C increase of temperature during the most comprehensive analysis to date using crop simulation models.
Crop models have a higher uncertainty in their simulation at elevated temperature due to their incorporated structures and functions [3]. Reducing the uncertainty surrounding the a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 quantification of climate change impacts in models is a major concern of crop modelers [4][5][6][7]. For better simulation of crop responses under increasing temperature and CO 2, is also important for crop modelers to improve models for better use in changing temperature and heat stress. Though model results are more sensitive to stress during reproductive stages as compared to vegetative stages [8], simulations could use improvement to reflect heat stress at both stages [9]. Currently, most studies of the impacts of heat stress on crop growth are performed in growth chambers [10,11], temperature gradient tunnels [12] or temperature free-air controlled enhancement (T-FACE) [13,14] systems. Data of FACE, FATE, and glassware experiment are usually used for formulation of models used to quantify the climate change impacts [7]. These techniques may not accurately represent real responses in field crop production systems. The results of controlled environments cannot be extrapolated to natural field conditions owing to natural differences in solar radiation, wind and evaporation requirements [15]. Craufurd et al. [16] suggested that crop science experiments were urgently required to evaluate and improve crop models of heat stress that is likely under future climate impact projections. Similarly, Liu et al. [17] tested four different models in pot experiments (phytotrons) but suggested that models should be tested against field experiment results.
The primary objective of this study was to evaluate the four crop models (CERES-Wheat, Nwheat, CROPSIM-Wheat and APSIM-Wheat) with phenology, leaf area maximum, above ground biomass and yield of field grown wheat across early to extremely late planting time. Seed bed was well prepared to sow wheat using seed rate 100 kg ha -1 at 22 cm row to row distance. Nitrogen, phosphorus and potassium were applied were applied at the rate of 120, 85 and 60 kg ha -1. Irrigation water was applied without giving water stress to crop. All other crop husbandry practices were kept same. Soil, crop management, crop phenology, leaf growth, yield, and weather data of both locations were collected following the standard procedures and methods.
DSSAT-CERES-Wheat. The DSSAT-CERES-Wheat model [18] under the shell of Decision Support System for Agro-technology Transfer (DSSAT v4.7) is the most cited wheat model; it has been tested and evaluated around the globe e.g., [19][20][21][22][23]. CERES-Wheat has been widely used for exploring agronomic options, breeding preferences, edaphic factors and climatic factors. This model has the capacity to simulate the developmental stages of wheat; growth of leaves, stem and grains; and biomass based on light interception and stresses.
APSIM-Wheat. The Agricultural Production Systems Simulator (APSIM) for wheat (v7.8) is an Australian based wheat model which has the ability to simulate the soil transformations such as nitrogen, water, crop residue, crop growth, development and their interactions [24]. APSIM-Wheat has been evaluated around the globe under different soils, climate, temperatures, CO 2 , planting dates, water, plant populations and cultivars [25][26][27][28][29]. APSIM-wheat simulates wheat growth on a daily time-step [30] by calculating thermal time from the difference between base temperature and 3-hourly crown temperatures derived from the daily maximum and minimum temperatures. The thermal time is then accumulated to determine the phonological development of the crop. The biomass accumulation is based on radiation use efficiency (RUE). Biomass partitioning rates to different plant parts vary with crop development stage and re-translocation begins at the stage of starting grain filling [28].
DSSAT-Nwheat. Recently, Nwheat has been embedded in DSSAT v4.7 as APSIM-Nwheat model. This model has been tested under the shell of APSIM-Nwheat in many environments for temperature, carbon dioxide, nitrogen and water transformation in soil [29][30][31][32][33]. Transpiration efficiency was increased directly 1 to 1.37 with doubling CO 2 from 350 to 700 ppm. DSSAT-Nwheat uses the same input data set as CERES-Wheat but requires more cultivar coefficients are needed to calibrate. A heat stress function was introduced in Nwheat by Asseng et al. [4] based on CERES-Wheat model [34].
DSSAT-CROPSIM-Wheat. DSSAT-CROPSIM-wheat is an integrated model in DSSAT v4.7 that simulates wheat development, growth and morphological parameters based on single plant then converts into whole plant population. Phenological stages are calculated on the concept of "Biological Days" a time measure that equates to chronological days under optimum conditions. It mainly simulates the major phenological stages as given in Zadoks' scale. Biomass is accumulated through intercepted radiation and distributed largely based on demand. Critical crop stresses are always considered during simulation of wheat under low or high temperature, which may cause plant death. Similarly low temperature at anthesis may cause sterility and reduction in final number of grains.

Model input data
Models require input data that describe daily weather, cultivar growth and development characteristics, management events, and soil characteristics. The minimum weather data requirements are daily temperature (minimum and maximum), solar radiation, rainfall, and station information (longitude and latitude). The models use different genetic coefficients for a cultivar such as: vernalization requirement, photoperiod sensitivity, thermal time requirement, kernel number per biomass, kernel growth rate, maximum stem dry weight, and phyllochron interval. Vernalization and photoperiod affect phenology between emergence and floral initiation. Grain yield potential is controlled by a coefficient of kernel number per ear and maximum kernel growth rate. Leaf appearance is associated with degree day accumulation by the phyllochron parameter. Main soil inputs include initial soil water content, lower and drained upper limits, saturated water content, water drainage and runoff coefficients, rooting growth factors, first stage evaporation, and soil albedo. Crop management information incudes planting date and depth, plant population, fertilizer and irrigation application rates and dates, as well as measured or estimated initial soil water and nitrogen content [35].

Model calibration and genetic coefficients
Calibration is the process of adjusting each model's parameters to reflect local conditions. Four models were calibrated with the 15 th November planting during 2013-14 at Faisalabad. This is a necessary step to ensure models provide useful information about the system of interest. It is also necessary to obtain genetic coefficients to represent any new cultivars used in a given modeling study. All four models have different genetic coefficients, which were adjusted as described in Table 1. Some soil parameters were also adjusted in the process of model parameterization. Each of the four models for each cultivars was calibrated using data collected from the least-stressed planting date 15 th November, 2013-14 at Faisalabad Genetic coefficients for local cultivars were not available. So, crop specific parameters were estimated through iteration approach [36] and comparison of simulated and observed data. First, crop specific parameters regarding crop phenology were estimated then growth and yield related genetic coefficients were determined in all crop models ( Table 1). The rest of the model parameters were taken from the original model documentations. Subsequently, calibrated models were applied to the remaining treatments of 2013-14 and 2014-15 at Faisalabad and Layyah. These genetic coefficients may be further used by environmentalist, crop breeders and geneticists for exploration of wheat cultivars under semiarid and arid environment of South Asia especially Pakistan.

Model evaluation
To check the accuracy of the model simulations, models were evaluated with the data recorded during both seasons 2013-14 and 2014-15 at site of Faisalabad and Layyah except the 15 th November planting date used for calibration. The output of models were compared using statistical metrics normalized root mean square error (NRMSE). NRMSE evaluates the average relative deviation between observed and simulated values in percentage. The output variable, index of agreement (d), is a dimensionless and bounded measure originally provided by Willmott [37] and commonly used to compare the match of observed and simulated data [38][39][40].
Simulation performance was evaluated by calculating the statistical indices below, where, Pi is simulated grain yield and Oi is observed grain yield.

Models calibration
CROPSIM-Wheat, CERES-Wheat, Nwheat and APSIM-Wheat model were calibrated (Table 2) to further evaluation and improvement suggestions.
Overall behavior and performance of models. Percentage difference (PD) allows a relative ranking of the performance of the models during calibration for days to anthesis and maturity, biological and grain yield and LAI maximum variables are presented below for three genotypes i.e.

Evaluations and validation
The performances of calibrated CROPSIM-Wheat, CERES-Wheat, Nwheat and APSIM-Wheat models were evaluated with independent data sets obtained from field-grown wheat   Table 3.      (Fig a, b, c and d), Days to Maturity (Fig e, f, g and h), Maximum LAI (Fig i, j, k and l), Above ground biomass (Fig m, n, o and p) and grain yield (Fig q, r, s and

Discussion
Many studies evaluated and improved crop models under high temperatures e.g. [27,41] in order to better simulate the results of climate change impacts on crops. Liu et al. [17] tested four models (DSSAT-CERES-Wheat, DSSAT, Nwheat, APSIM-Wheat, and Wheat-Grow) under heat stress conditions in phytotrons at grain filling and anthesis stages and highlighted the need for improving model simulation of grain yield and its components through field experimentation. Precise simulation of wheat development is the first step for accurate simulation of biological and grain yield, as well as their components [42]. Genetic characteristics, photoperiod and temperature are the main determinants of crop stages, but temperature is a major determinant of phenological stages [43]. Asseng et al. [2] also reported that the phenology of wheat is mainly regulated by temperature. Four models of wheat simulated development stages like anthesis and maturity with NRMSE ranging 12.3-35.8% and 11.1-59.6%, respectively across both locations and growing seasons. Simulated phenology of four models varied from observed due to different simulation function of four crop models. Large variation among the models because of different assumptions for parameter functions [9] like which cardinal temperature. CROPSIM-Wheat, CERES-Wheat, Nwheat, and APSIM-Wheat predicted days to anthesis closely. However, CROPSIM-Wheat, CERES-Wheat and Nwheat in DSSAT 4.7 [44], predicted the days to maturity similarly while APSIM-Wheat showed increased days to maturity with delayed planting due to its incorporated functions of photoperiod, cardinal temperature and low temperature sensitivity. APSIM-Wheat model empirically calculates of mean crown temperature to determining thermal time from daily maximum and minimum temperature, and calculates temperature stress by daily mean temperature [24].
A good model integrates all crop parameters and the effect of stresses on these parameters for final grain yield [9]. Liu et al. [17] also pointed to the need to improve the heat response of APSIM-Wheat. CROPSIM-Wheat performed comparatively better, providing good simulation of days to anthesis and maturity and total above ground biomass. CROPSIM-Wheat's grain yield calculation method and cultivar coefficients also contributed to the good model performance. This model simulated the yield on the basis of tillering following 2.5 leaves at main stem, and grain numbers are determined by the function of the difference between the above ground biomass and at the end of anthesis stage and earlier stage [45]. Performance of CERES-Wheat was not as good at simulating phenology, but its NRMSE ranged from 35.30% to 62.60% across sites and years. Overall, CERES-Wheat over simulated grain yield; the model showed less sensitivity to increasing temperature after anthesis. CERES-Wheat simulation of days to anthesis and maturity did not show an effect of high temperature during grain filling stage on grain size and filling duration as in the field. Liu et al. [17] similarly reported that CERES-Wheat underestimates heat effects on grain filling duration. Models calculation of grain numbers at flowering stage and grain size as a function of grain growth rate and biomass partitioning at the reproductive stage [46] have been recommended for modification to better reflect heat stress effects [7,47] Among four models of our study, Nwheat best simulated biomass, with NRMSE ranging from 21.6%-26.8%, followed by CROPSIM-Wheat (16.6%-38.5%), APSIM-Wheat-Wheat 24.1-40.8% and CERES-Wheat (25.3%-72.50%). Nwheat biomass outputs were more sensitive to heat stress effects than CERES-Wheat and CROPSIM-Wheat [35]. APSIM-Wheat simulation performance was reliable but tended to overestimate biomass and its components like days to maturity and grain yield. In particular, later planting dates were associated with increased days to maturity, days to biomass accumulation, and consequently over simulation of biomass. A likelier driver of biomass overestimation is leaf area over simulation; days to maturity were over simulated in all environments but leaf area was over simulated in the same cases in which biomass was over simulated.
RMSE of all four models were averaged at two locations during two years to check the performance for simulation of days to anthesis (27.93%) and maturity (25.55%), grain yield (48.71%), biological yield (30.55%) and leaf area index (42.57%). As in other studies, we recommend the improvement of the models' response function for simulation of grain yield at high temperature and under heat stress [17,35]. Furthermore, we found room for improvement in simulations of leaf area index in the current models tested [21]. Comparison of mean NRMSE of all models, parameters, locations showed similar response in both experimental years 2013-14 (34.27%) and 2014-15 (35.25%). Models' performance was better at Faisalabad (41.28%) than Layyah (33.96%). Mean NRMSE for all model parameters at both locations during year 2013-14 and 2014-15 showed that Nwheat (33.95%) performed better than CERES-Wheat (36.67%) and APSIM-Wheat (41.991%) due to better response to changing photoperiod, temperature and genetic coefficient.
Evaluation of models showed their level of reliability of simulation under different environments and temperature regimes such Faisalabad and Layyah as well as early and late sowing dates. Hussain et al. [48] reported in review that high temperature under climate change scenario would affect badly to wheat in semiarid and arid environment. These changing impacts of temperature could be offset through breeding and agronomic adaptations. Agronomic adaptations such as efficient irrigation, adjusting planting dates (16 th November ± 10 days) and increasing nitrogen application (10%), could enhance crop yield. Development of virtual cultivars through crop simulation modeling would be a good recommendation for breeder for breeding heat and temperature resistant cultivars.
In crux, multiple models performed well in early (16 th October), optimum (1 st and 16 th November) and late (1 st and 16 th December, 1 st January) sowing, but for very late planting dates (16 th January, 1 st and 16 th February, 1 st and 16 th March) under high temperature, models performance was poor. Performance of models during evaluation was sequenced as CROPSIM-Wheat > Nwheat > CERES-Wheat > APSIM-Wheat. Model performance was least accurate at simulating field data on leaf area index followed by grain yield. These data from later planting date field experiments are important for evaluating model performance at high temperature and can be used to further improve crop models in areas where heat stress is likely.