The interactions between genotype, management and environment in regional crop modelling

Biophysical models to simulate crop yield are increasingly applied in regional climate impact assessments. When performing large-area simulations, there is often a paucity of data to spatially represent changes in genotype (G) and management (M) across different environments (E). The importance of this uncertainty source in simulation results is currently unclear. In this study, we used a variance-based sensitivity analysis to quantify the relative contribution of maize hybrid (i.e. G) and sowing date (i.e. M) to the variability in biomass yield (YT, total above-ground biomass) and harvest index (HI, fraction of grain in total yield) of irrigated silage maize, across the extent of arable lands in New Zealand (i.e. E). Using a locally calibrated crop model (APSIM-maize), 25 G x M scenarios were simulated at a 5 arc minute resolution (∼5 km grid cell) using 30 years of historical weather data. Our results indicate that the impact of limited knowledge on G and M parameters depends on E and differs between model outputs. Specifically, the sensitivity of YT and HI to genotype and sowing date combinations showed different patterns across locations. The absolute impact of G and M factors was consistently greater in the colder southern regions of New Zealand. However, the relative share of total variability explained by each factor, the sensitivity index (Si), showed distinct spatial patterns for the two output variables. The YT was more sensitive than HI in the warmer northern regions where absolute variability was the smallest. These patterns were characterised by a systematic response of Si to environmental drivers. For example, the sensitivity of YT and HI to hybrid maturity consistently increased with temperature. For the irrigated conditions assumed in our study, inter-annual weather conditions explained a higher share of total variability in the southern colder regions. Our results suggest that the development of methods and datasets to more accurately represent spatio-temporal G and M variability can reduce uncertainty in regional modelling assessments at different degrees, depending on prevailing environmental conditions and the output variable of interest.