All three crop models performed well in simulating winter wheat yield (Table 3, Fig. 2). Across all sites, years, and soil water conditions, index of agreement (d) exceeded 0.9 for all models (d = 0.97 [SWAT], 0.96 [WEPS], 0.93 [UPGM]). Normalized objective function (NOF) demonstrated slightly decreasing performance from SWAT (0.16) to WEPS (0.23) to UPGM (0.32). WEPS and UPGM generally overpredicted yield (positive RE = 10.5% [WEPS], 19.6% [UPGM]) whereas SWAT slightly underpredicted yield (negative RE = -3.4%), particularly for Greeley Wet. Observed yield for Greeley Dry was similar over the 3 years (3270–3762 kg/ha), but all models simulated greater yield differences (SWAT: 3073–4046; WEPS: 2883–5159; UPGM: 3271–5631 kg/ha). Even without calibration, all models demonstrated skill in simulating wheat yield across the full range of water conditions spanning 10 crop-years. However, no clear benefit was evident for the HU approach to simulating yield used by WEPS and UPGM over the HI approach used by SWAT.
Table 3
Model performance statistics.
| SWAT | WEPS | UPGM |
Variable | RE (%) | d | NOF | RE (%) | d | NOF | RE (%) | d | NOF |
Yield | -3.4 | 0.97 | 0.16 | 10.5 | 0.96 | 0.23 | 19.6 | 0.93 | 0.32 |
AG biomass | -20.2 | 0.86 | 0.35 | -11.6 | 0.91 | 0.30 | -13.7 | 0.87 | 0.33 |
Canopy height | 49.4 | 0.44 | 0.59 | -41.6 | 0.60 | 0.46 | -33.6 | 0.68 | 0.38 |
Harvest index | 37.7 | 0.43 | 0.48 | 24.5 | 0.61 | 0.36 | 35.0 | 0.54 | 0.47 |
RE = relative error (%), d = index of agreement, NOF = normalized objective function, AG = aboveground. |
All three crop models performed well in simulating final aboveground biomass (Table 3, Fig. 3). Performance in simulating final biomass declined slightly compared to yield performance for all models, but d exceeded 0.86 and NOF was less than 0.35 for all models. WEPS (RE = -11.6%, d = 0.91, NOF = 0.30) performed best, followed by UPGM (-13.7%, 0.87, 0.33) and SWAT (-20.2%, 0.86, 0.35). For biomass simulation, WEPS and UPGM, which used an HU partitioning approach outperformed SWAT with its standardized LAI-curve approach. This may point to a benefit in accumulating HUs as attenuated by water stress in simulating crop biomass.
Multiple in-season observations at the Greeley site demonstrated that WEPS and UPGM closely simulated aboveground biomass throughout the growing season, whereas SWAT results were consistently lower and underpredicted observed values (Fig. 4). In cases where errors developed in biomass simulation, they generally occurred late in the season. Final biomass was substantially underpredicted by all models for two site years: Greeley Wet 2009 (Figs. 3, 4) and Drake 2006 (Figs. 3, 5). For Greeley Wet 2009, in-season data shows that all models performed reasonably well up to day 149, but simulated biomass plateaued after about day 180 while observed biomass increased up to day 199 (Fig. 4a). Other treatment-years at Greeley showed an end-of-season plateau in observed biomass that was captured by simulated results. This may indicate that the models are not capturing a periodic climate or management influence on biomass accumulation. For example, the underpredicted Greeley 2009 and Drake 2006 site years were the warmest and driest years during the grain-filling period (1 May to 30 June) at their respective sites during the study period (Table 2). In some cases, UPGM better tracked early-season while WEPS better captured end-of-season biomass accumulation (Greeley Wet 2011, Drake 2004) or vice versa, with WEPS better tracking early-season biomass accumulation (Greeley Wet 2010, Greeley Dry 2011), which suggests the need to calibrate timing of phenological development stages with biomass accumulation (Figs. 4, 5).
SWAT simulates daily accumulated total biomass; WEPS and UPGM also simulate daily biomass partitioned to leaves, stems, straw (leaves + stems), and seed heads (Fig. 6) as well as roots (not assessed in this study). SWAT consistently underpredicted total biomass compared to WEPS and UPGM except for three occasions: UPGM Greeley Wet 2009 (Fig. 6a) and WEPS Greeley Dry 2011 (Fig. 6f) at the very end of the season, and WEPS Drake 2004 HRU 23 (Fig. 6g) for most of the season.
The considerable effect of water stress on simulated timing of developmental stages by WEPS and UPGM is also evident in Fig. 6, as represented by the five vertical lines shown in each diagram for the simulated date of J, S, H, AS, and M stages. WEPS and UPGM differ in the magnitude and timing of biomass partitioning leading to the differences between the biomass accumulation curves. Biomass data collected near day 150 for 8 of the 10 site years often were well simulated for seed heads (WEPS, except Fig. 6c; and UPGM, except Fig. 6c,i), straw (WEPS, except Fig. 6a,b,g; and UPGM, except Fig. 6a,b,e), and total biomass (WEPS, except Fig. 6g; and UPGM, except Fig. 6a,b,c,d,e,i).
The wheat canopy was consistently simulated by SWAT to achieve maximum height regardless of site and climate factors, whereas UPGM and WEPS better reflected observed heights (Fig. 7). Simulated final canopy height increased in rough agreement with observed heights (d = 0.60 [WEPS], 0.68 [UPGM]), although height was underpredicted by both WEPS (RE = -42%) and, to a lesser extent, UPGM (-34%) (Table 3, Fig. 7). The SWAT method using a HU-adjusted fraction of maximum canopy height was accurate only for non-limiting conditions, such as the fully irrigated Greeley Wet conditions of 2009 and 2011 (Figs. 7, 8). Overall, SWAT significantly overpredicted height (RE = 49%) (Table 3, Fig. 7c). Multiple in-season observations at the Greeley site demonstrated that all models simulated a mid-season increase followed by late-season plateau (Fig. 8). Although WEPS and UPGM allow simulation of season-specific canopy height development, both models have difficulty capturing both timing and magnitude of the growth without further adjustments to parameters.
By design, HI in SWAT is relatively stable (here, about 0.50) across environmental conditions, while HI in WEPS and UPGM better predicts the observed values (Fig. 9). HI ranged from 0.23 to 0.60 across sites and years, so WEPS and UPGM, which can account for this variability, have a clear advantage. All models generally overpredicted HI. WEPS (RE = 25%, d = 0.61, NOF = 0.36) performed best, followed by UPGM (35%, 0.54, 0.47) and SWAT (38%, 0.43, 0.48) (Table 3, Fig. 9). Interestingly, SWAT had the greatest overprediction of HI (RE = 38%), which combined with the greatest underprediction of aboveground biomass (-20%) and resulted in the most accurate prediction of yield (-3.4%). WEPS and UPGM do not use HI in calculation of yield, so a similar comparison is not relevant.
No LAI observations were made, so model performance can only be assessed by comparison among models (Fig. 10). For SWAT, the timing varied slightly among plots and years, but the pattern was similar with a consistent peak LAI of 3.0, because SWAT was not responding to variable water conditions. The magnitude and timing of LAI varied only slightly among site years for SWAT but varied considerably for WEPS and UPGM in response to site conditions. In many cases, WEPS and UPGM had greater peak LAI than SWAT, sometimes exceeding a reasonable maximum LAI of 5. Examining leaf senescence after peak LAI, all three models overpredicted final LAI.