Surface ozone (O3) pollution is one of the major threats to global food security due to the detrimental effects of ozone exposure on crops (Ainsworth et al., 2012; Avnery et al., 2011b; Leung et al., 2020; Long et al., 2005; Tai et al., 2014; Tai and Val Martin, 2017). In the United States alone, crop loss due to tropospheric O3 costs more than USD 5 billion annually (Ainsworth et al., 2012; Avnery et al., 2011a; Van Dingenen et al., 2009).

Soybean is one of the main staple crops for human consumption; it also serves as an important source of animal feed. It is a cheap source of proteins, and therefore soybean products are consumed around the world. The impact of O3 on soybean physiology and growth has been studied extensively (Ainsworth et al., 2012; Betzelberger et al., 2012; Dermody et al., 2008; Morgan et al., 2003). Crop yield losses to tropospheric O3 have been quantified using model projection and experiments. The National Crop Loss Assessment Network and European Open Top Chamber programmes have established the air quality guideline, which derived dose–response relationships from comparable experimental data. These campaigns provided critical information such as the O3 response relationship and estimated yield loss due to O3 damage that enabled regional projections of O3 effects on crop yields (Fuhrer, 2009). However, open top chambers modify plant response to O3 due to the “chamber effects” which create microclimates (Elagöz and Manning, 2005) and environmental differences between the chamber and open air micrometeorology in which yield loss is underestimated (Van Dingenen et al., 2009). Recently the introduction of free-air concentration enrichment (FACE) technology avoids the artefacts from enclosed chambers, and O3 fumigation was adapted to FACE facilities (Agathokleous et al., 2017; Paoletti et al., 2017). The application of FACE experiment on crops took place in China (Zhu et al., 2011) and the USA, including experiments with soybean at the SoyFACE facility in Champaign, Illinois (Morgan et al., 2004; Betzelberger et al., 2010, 2012).

Crops are a significant component of the land surface; e.g. croplands and pasturelands represent 12 % and 26 % of the global terrestrial land, respectively (Van den Hoof et al., 2011). Moreover, the phenology of crops is very different from that of natural vegetation and is characterised by high growth, turnover rate, and strong seasonality. It is thus necessary to include a crop-specific parameterisation scheme to improve simulations of land surface fluxes and regional climate in agroecosystems (Van den Hoof et al., 2011). The Joint UK Land Environment Simulator with crops (JULES-crop) is a crop parameterisation (Osborne et al., 2015) within the land surface model JULES (Best et al., 2011; Clark et al., 2011). Global simulations have been performed with JULES-crop for rice, wheat, maize, and soybean (Osborne et al., 2015). These four crop types contribute more than 70 % of human calorie intake (Ray et al., 2013). JULES-crop includes routines representing growth, development, and harvesting of crops driven by the overlying meteorological inputs. In JULES-crop, four new prognostic variables have been added: crop development index (DVI), root carbon (Croot), harvest carbon (Charv), and reserve carbon (Cresv). DVI controls the duration of the crop growing season in four distinct stages – sowing, emergence, flowering, and maturity – and it determines when changes in carbon partitioning occur (Osborne and Hooker, 2011). Croot, Charv, and Cresv are the carbon pools for roots, harvested organs (e.g. grains of cereal, fruits, and root), and stem reserves, respectively. Carbon pools for stem and leaves are determined from the existing prognostic variables, leaf area index (LAI) and canopy height. In Osborne et al. (2015), global runs of maize, wheat, soybean, and rice were carried out using JULES-crop. Site runs were performed at four FLUXNET sites with soybean–maize rotation: Bondville (US-Bo1), Fermi (US-IB1), and Mead (US-Ne2 and US-Ne3). Simulated yield was compared against country and global FAO crop yields. Osborne et al. (2015) used generic representations for each of the crops in their global study. For the plant parameters that are needed outside the crop model such as leaf nitrogen and leaf respiration parameters, these are set to those of the C3 or C4 grass functional types. Osborne et al. (2015) suggested that these parameters could be tuned to be more crop specific to improve fit to observations. These JULES parameters have been calibrated against observations for maize, using data from the Mead FLUXNET sites in Nebraska (Williams et al., 2017). However, to date, these parameters have not been calibrated to soybean data.

There are many crop models developed by institutions/organisations around the world. Most are designed for application to an individual field up to the regional scale and do not include O3 impacts on vegetation. Supplement Table S1 compares a selection of land surface models which include crop tiles and have the functions to model climate impact on crop productivity. JULES-crop is of particular interest because it is a development of the global land surface component JULES of the Met Office numerical weather prediction and climate models and contains a detailed representation of plant physiological processes at sub-diurnal timescales, including consideration of O3 effects on natural vegetation, thus making it suitable for this study. JULES-crop has been accepted into the JULES trunk with the intention to be coupled with the Hadley Centre Global Environment Model (HadGEM) in the near future. HadGEM is recognised as one of the best performing climate models with smaller errors than typical climate models (Gleckler et al., 2008; Knutti et al., 2013).

The calibration of O3 damage on soybean would allow land surface and crop models to more realistically and reliably simulate present-day and future O3 damage and subsequently to quantify its economic impacts. The objective of this study is to calibrate soybean representation for JULES-crop, with a particular focus on the response of soybean to O3 exposure.

This paper is organised as follows: Sect. 2 describes the model set-up and observations used for the JULES calibration. Section 3 compares the results from the calibrated JULES runs against independent observations. Section 4 assesses the suitability of the model for modelling soybean under O3 damage and discusses ways of future model improvement.

A flowchart demonstrating the calibration and evaluation procedure is given in Fig. 1. We first tuned the JULES-crop soybean parameterisation at the US-Ne2 and US-Ne3 Mead sites, where three years of soybean physiological and meteorological observations were available, at ambient ozone (Fig. 1, steps 1–5). The three years are 2004, 2006, and 2008 in which soybeans were grown in Mead; maize was grown in other years.

Secondly, to calibrate the JULES ozone damage parameters (Fig. 1, step 6) we made the assumption that there is a negligible damage to crop yield at ambient background levels of O3 at both the SoyFACE and Mead sites. This is consistent with Mills et al. (2007), who reviewed over 700 published papers and conference proceedings and found that O3 level of AOT40 over 3 months of 5 ppm h reduced soybean yield by less than 5 %. Then we calibrated specifically the soybean O3 response using leaf gas exchange measurements from soybean grown under elevated O3 concentrations at SoyFACE.

Finally, we applied JULES-crop newly calibrated for soybean and its O3 sensitivity at the leaf level and evaluated model performance against observed yield and leaf area index from SoyFACE, taken for the full range of rings and cultivars (Fig. 1, step 7).

Results from JULES runs with crop model and ozone damage turned on are shown in Figs. 3 and 4. Figure 3 shows the evaluation of the soybean aboveground biomass carbon for different O3 exposure levels (AOT40) using the O3 damage parameters in Table 4. The model aboveground carbon (solid lines) is compared to the line fitted in Betzelberger et al. (2012) to their aboveground carbon observations. The run with the newly-calibrated parameters overestimated the carbon at ambient ozone levels. One contributing factor could be that water stress is underestimated in the new configuration, since it was not possible to evaluate the response to soil water availability using the Mead site data, so we instead derived a value for fsmc_p0 (parameterised in the calculation of the threshold for water stress; see Table 3) from the literature. We tested the sensitivity to this choice by re-running this configuration with fsmc_p0=0, which represents water-stressed conditions, and this caused a 12 % reduction in aboveground carbon (plots shown in Supplement). In addition, the representation of the soil properties in the JULES SoyFACE run could be improved by calibration to site measurements. In contrast, the Osborne 2015 tuning intersects the line fitted to observed aboveground carbon at zero ozone concentration (partially because of higher water stress) but then shows a sharp decrease from zero to ambient levels, which is not realistic. Note that no observations were taken for below-ambient ozone concentrations at SoyFACE, so this section of the fitted line is an extrapolation. The slope of the aboveground carbon response to increasing ozone concentrations is similar for all three runs and compares very well to the Betzelberger et al. (2012) fitted line.

The yield-O3 response curve in Fig. 4 shows that new parameterisation slightly overestimates yield in the ambient SoyFACE ring, compared to the spread of SoyFACE yield observations from Betzelberger et al. (2012). The Osborne 2015 tuning with high ozone sensitivity is within the spread of measured yield in ambient conditions, but note that the modelled yield has decreased sharply from zero ozone concentration to ambient levels, which is undesirable. The magnitude of the gradient of yield against AOT40 for all three model configurations is within the spread of the observations. However, the slope is underestimated for the new, calibrated run and overestimated for the Osborne 2015 tuning, especially for the range from ambient to 40 ppm h. Recall that ozone concentration modifies net leaf CO2 assimilation rate in JULES and that the model parameters governing this process (Fo3crit, a) are calibrated directly to net leaf CO2 assimilation rate observations from SoyFACE in our new configuration (Sect. 2). Reductions in the modelled net leaf CO2 assimilation rate lead to the reductions in model aboveground biomass, yield, and LAI, which we show in this section. However, Betzelberger et al. (2012) also reported additional impacts of ozone damage, such as changes in leaf absorptance and specific leaf mass, that are not represented in JULES, and therefore our tuning does not account for them. In contrast, the values of Fo3crit and a in the high and low sensitivity versions of the Osborne 2015 tuning simulations (Table 4) were calibrated in Sitch et al. (2007) to yield observations. Therefore, they can be seen as “effective” parameters in these configurations, since they incorporate the effect of the ozone damage processes that are not explicitly represented in JULES.

Note that we plot AOT40 on the x axis for illustrative purposes only, to be comparable with results presented in Betzelberger et al. (2012) – AOT40 was not used in the JULES run. An alternative would be to plot ring number or ring target concentration. Ideally, we would plot the x axis with the metric phytotoxic ozone dose (POD) for JULES and observed data, which account for the dosage of O3 that get into the stomata of soybean, but this is beyond the scope of the present study.

Figure 5 compares the model and observed LAI at SoyFACE for different O3 concentrations. JULES was able to reproduce LAI seasonality; however, it underestimated the amplitude. The maximum LAI for calibrated JULES peaked around day 240 in September, and observations peaked at DoY 220–230. The peak LAI in the model runs was less than half the observed LAI in all cases. While the Mead model runs also showed a slight underestimation of peak LAI compared to observation (Supplement), the majority of the underestimation of the modelled SoyFACE LAI is due to a difference between the observed relationships between peak LAI and yield at the Mead and SoyFACE sites. At both sites, observed maximum yield increases with observed peak LAI. However, for similar observed yields, the observed SoyFACE yield tends to be higher than the observed Mead LAI. Given that our calibration is based on Mead observations, it is therefore not surprising that our model runs at SoyFACE underestimate peak LAI compared to the SoyFACE observations.

A contributing factor to the different relationship between observed peak LAI and observed yield at SoyFACE compared to Mead could be the different methods used to measure LAI at the Mead sites (which this parameter set was tuned against) and at SoyFACE. At Mead, destructive measurements were taken, whereas at SoyFACE, LAI was measured indirectly, using radiation attenuation through the canopy.

Another plausible contributing factor for the different relationship between observed peak LAI and observed yield at SoyFACE compared to Mead is the row density of the soybean. The SoyFACE row spacing was half that of Mead, so as described above we set the initial carbon to twice that observed at Mead. The denser planting allowed soybean at SoyFACE to reach higher LAI earlier in the growing season. If this also resulted in thinner leaves at the beginning of the season than with the Mead row spacing, then this could explain the difference in the peak LAI to yield relationship between the two sites. Ricaurte et al. (2016) showed that higher sowing density would increase phyllochron in a linear relationship, which results in a higher LAI measured, which is consistent with our study. JULES also does not account for leaf age on leaf assimilation rate – in reality a lower leaf assimilation is observed in the late season associated with leaf ageing, and it is plausible that this could also be affected by row spacing.

Figure 5 also demonstrates that model LAI responds more to ozone concentrations than the observed LAI. One contributing factor is the observed decrease in specific leaf area at SoyFACE in increased ozone (Betzelberger et al., 2012). As mentioned above, this process is not captured by JULES. This issue is particularly pronounced in the Osborne 2015 tuning runs, where the modelled LAI in the ring with target 200 ppb is roughly a third of the peak LAI in the ambient ring.

Climate change and air pollution are a great threat to food production. JULES-crop has been developed to represent crops in the land surface model and allows us to estimate the future climate and air pollution impact on crops. The O3 impact on crops could be quantified with an improved parameterisation to the existing O3 damage scheme for C3 plants. The default soybean biochemical and respiratory parameters in JULES were based on C3 grass parameters. Characteristics of soybean are more similar to a shrub than grass; therefore, parameter calibration is needed to improve the performance of soybean in JULES-crop.

In this paper, the parameters needed to describe soybean in JULES-crop were first revised against observations from the Mead FLUXNET sites to ensure that the crop biochemical and respiratory parameters explicitly represented soybean. Comparison with observations from these sites showed that GPP and LAI were well represented for irrigated soybean at Mead. The O3 damage parameterisation was subsequently calibrated against leaf gas exchange observations from the Soybean Free Air Concentration Enrichment (SoyFACE) facility for the O3 damage, by tuning the sensitivity and critical threshold of O3 damage. On the whole, JULES-crop reproduces the observed negative correlation between yield and O3 exposure. It also reproduced the negative impacts of ozone on LAI and the seasonality of phenology, although the simulated LAI was underestimated at SoyFACE. This method of calibrating soybean could be replicated for other crops once data become available and would contribute to more accurate parameters for crop models. The calibration will be applied to a regional and transient run and eventually the newly calibrated JULES-crop for soybean and its sensitivity to O3 damage, coupled within an Earth system model.