How well does APSIM NextGen simulate wheat yields across Australia using gridded input data? Validating Continental-Scale Crop Model Simulations

Processed-based models are increasingly being used with gridded soil and weather data; however, their validation is often on small, sampled datasets, calling into question their accuracy when extrapolated to larger scales with more uncertain input data. Here, we benchmarked the accuracy of the APSIM Next Generation (APSIM-NG) crop model for wheat yield predictions on multiple trials from 2005 to 2022 using data from the largest independently coordinated national trial network in the world (Australia's Grain Research and Development Corporation's National Variety Trials). This gauged the validity of APSIM application with gridded soil and weather data across the 60 million ha of the Australian grain belt over time. Overall, results indicated that APSIM-NG can track the spatial-temporal changes in wheat yields from the studied period at the national, state, or agroecological level – nationally: R2 = 0.5, d = 0.83, a bias of 6.9%, non-significant monotonic trend (p>0.05), and significant Genetic X Environment effect (p<0.05) – indicating acceptable model performance in the face of substantial variability in the observed data. There was notable variation in temporal, year-by-year R2 varying from 0.19 to 0.65, as well as spatial model performance, with R2 varying from 0.33 to 0.61 between the states of Queensland and South Australia, respectively. There are numerous causal factors associated with such reduced performance levels in crop yield simulation using big spatial gridded input datasets in comparison with more detailed experimental datasets; for example, the R2 = 0.89 from the internal APSIM-Wheat validation dataset. We discuss these factors in detail herein, but they include poor cultivar association, local input parameter variability from both gridded climate (patchy rainfall, variations in particularly cold temperatures due to micro-landscape elevation) and gridded soil datasets, as well as the sometimes-undetected impacts of pests and diseases in large-scale datasets, which APSIM does not simulate. Future efforts to reduce these uncertainties will no doubt help bridge the gap in performance between crop simulation using detailed experimentally measured input data and that from gridded datasets. Our analysis indicates this could help models explain an additional 39% (R2 = 0.89–0.5) of the observed variability.