4.4 Perspective of applying sophisticated crop models in data
assimilation
Compared with applying simpler models like the crop model SAFY over
large scales (Kang and Özdoğan, 2019), it is a challenge to use
sophisticated crop models like GECROS for practical applications. For
instance, GECROS is distinguished by a highly detailed photosynthetic
process (Yin and Struik, 2017) and the related photosynthetic parameters
were determined based on our previous study (Wang et al., 2022) and
preset in this study, although the uncertainty still exists (Fig. S3).
Also, the states updated in DA should be more carefully chosen, due to
the complex feedback of the updated states in the sophisticated crop
models that incorporate complex feedback mechanisms among biological
processes. For instance, updating LAI in the crop model APSIM directly
contributes to the improved performance of crop growth and yield
formation (e.g., Zhang et al., 2022), due to its physiological mechanism
that LAI influences the biomass accumulation by directly controlling the
intercepted solar radiation in the canopy (Monsi and Saeki, 2005).
However, as LAI is determined by both carbon and N status in the crop in
the crop model GECROS (Yin et al., 2000), updating LAI alone hardly
generated feedback for crop growth and yield forecast (results not
shown) and thus updating states of W leaf andN leaf were incorporated together in this study as
well (Fig. 7a). Similarly, Ines et al. (2013) indicated that updating
states like W leaf and specific leaf area, which
feedback to LAI, might reduce the sensitivity of EnKF to model bias of
the crop model DSSAT.
On the other hand, simple models have their own weaknesses. Due to the
simplified physiological process, the key parameters in simple crop
models like SAFY tend to be not only site-specific, but also
year-specific (Claverie et al., 2012; Kang and Özdoğan, 2019). As the
year-specific parameters should be calibrated with the actual in-season
meteorological data, forecasting of crop growth and end-of-season yield
is very uncertain (Kang and Özdoğan, 2019). In contrast to this, without
any year-specific parameter calibration, GECROS performed reasonably
well in the validation year (Table 6), especially forW above (R 2> 0.87, NRMSE = 0.28). By only updating remotely
sensed leaf traits, the performance of the simulatedW above improved and its NRMSE at maturity
further decreased to 0.25 (Fig. 7b-c). Relying on the integrated
simulation of physiological processes relating to N dynamics,N above and N grain were
simulated by GECROS and further updated more accurately in EnKF than
that of open-loop (Fig. 7b-c), which probably forms a better basis for
predicting traits like grain quality (Ma et al., 2022).