SENSITIVITY OF PREDICTED AGRO-ECOSYSTEM VARIABLES TO ERRORS IN WEATHER INPUT DATA

摘要

Statistically interpolated weather station data, outputs from climate reanalyses, and results from downscaled general circulation model (GCM) simulations are widely used to drive a variety of agro-ecosystem model applications, including regional-and national-scale crop modeling. In this study, we compared these three types of gridded weather datasets (total of nine datasets) with actual point-level weather station observations and analyzed the biases in predicted ecosystem variables of evapotranspiration (ET), crop grain yield, soil organic carbon (SOC) change, and soil N2O emissions using the process-based DayCent ecosystem model. As a reference system, we defined continuous corn cropping systems for three different regions in the U.S. Our results suggested that the predicted ecosystem variables can be highly sensitive to the sources of weather input data. Interpolated weather data from the PRISM and Daymet data products provided relatively accurate estimations of important ecosystem variables. Compared with the bias-corrected constructed analogs (BCCA) method, GCM results downscaled with the multivariate adaptive constructed analogs (MACA) method performed better for agro-ecosystem simulations under climate change; for datasets using BCCA, the rainfall frequency was positively biased and likely caused models to significantly underestimate solar radiation. For regional climate change studies that use a baseline simulation (historical period) for comparison, we suggest including the uncertainty of the baseline due to biases in the weather data in addition to the uncertainty in the projected weather data.