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).