Fig 1. Framework of the developed methodology in this study that applies
systematically analyzed errors of observations and simulations in the
data assimilation procedure to enhance crop status forecast. Field
experiments in two successive years were conducted for the acquisition
of the necessary dataset to validate this method. Daily weather data
served as the forcing input of the crop model GECROS and all field
observations in the first year were averaged for model calibration.
Before conducting the calibration procedure, the parameters in GECROS
were fixed or treated as uncertain and the error model that describes
the uncertainty of crop model simulation was assumed. The uncertain
parameters in GECROS and parameters in the error model were determined
simultaneously by an efficient Markov Chain Monte Carlo approach
(DREAM_zs). To further improve the forecast ability of the crop model,
the in-season observations of leaf traits in the second year were
incorporated by the commonly used data assimilation procedure, Ensemble
Kalman Filter (EnKF), which integrates the sequential observations into
crop model simulations of crop growth processes. Two types of
observations were collected. The first one was from field destructive
measurements. The second type was from the remote sensing predictions,
which were regressed from the machine learning method of Gaussian
Process Regression (GPR). The uncertainties of field measurements were
derived from the replications in the 2nd year’s field
experiment, while that of remote sensing predictions were estimated from
the GPR model itself. With the systematically analyzed uncertainties of
crop model simulations and observations, in-season leaf traits at the
canopy level, leaf weight (W leaf), leaf nitrogen
(N) content (N leaf) and leaf area index (LAI)
were updated directly by EnKF, and other crop carbon and N statuses,
including aboveground biomass (W above), grain
weight (W grain), aboveground N content
(N above) and grain N content
(N grain), were updated accordingly as well. The
performance of the updated states in the second year was evaluated by
their in-situ measurements, accordingly.