Abstract:
The need for more comparisons among models is widely recognized. This study aimed
to compare three different modelling approaches for their capability to simulate and
predict trends and patterns of winter wheat yield in Western Germany. The three
modelling approaches included an empirical model, a process-based model (LINTUL2),
and a metamodel derived from the process-based model. The models outcomes were
aggregated to general climate zones level of Western Germany to allow for a comparison
with agricultural census data for validation purposes. The spatial patterns and temporal
trends of winter wheat yield seemed to be better represented by the empirical model (R2=
70%, RMSE= 0.48 t ha-1 yr-1, and CV-RMSE= 8%) than by the LINTUL2 model (R2=
65%, RMSE= 0.67 t ha-1 yr-1, and CV-RMSE=11%) and the metamodel (R2= 57%,
RMSE= 0.77 t ha-1 yr-1, and CV-RMSE=13%). All models demonstrated a similar order of
magnitude of yield prediction and associated uncertainties. The suitability of the three
models is context dependent. Empirical modelling is most suitable to analyze and project
past and current crop-yield patterns, while crop growth simulation models are more
suited for future projections with climate scenarios. The derived metamodels are fast
reliable alternatives for areas with well calibrated crop growth simulation models. A
model comparison helps to reveal shortcomings and strengths of the models. In our case,
a performance comparison between the three modelling approaches indicated that, for
simulating winter wheat growth in Western Germany, higher sensitivity to soil depth and
lower sensitivity to drought in the LINTUL2 model would probably lead to better
predictions.
Keywords: Crop growth simulation model, Climate change, Metamodel, Regression
analysis, LINTUL2