Aggregation of Gridded Crop Model Outputs and Validation Across Scales.

For most climate vulnerability, impact, and adaptation (VIA) analyses in agriculture or related sectors, researchers and the typical stakeholders that they target for their information products, require measures of crop yields and climate impacts at decision-relevant environmental and geo-political scales (e.g. counties, states, nations, watersheds, and river basins). Though detailed biophysical crop models are likely the most promising way to quantify these yield and impact measures, these models are generally focused on reproducing phenology and productivity from detailed field-scale measures and have typically been less tested and less successful at reproducing larger scale measures. We present here results from a gridded version of the CERES-Maize model run for the conterminous US at 5 arcminute spatial resolution using historical observation- and reanalysis-based weather products spanning 1980-2009. We further present a methodology for aggregating gridded results to decision relevant spatial scales and compare these aggregations with statistics from the USDA National Agricultural Statistics Service (NASS) by developing metrics at various scales for comparative validation of the model and the input data products. We show that the accuracy of the simulations, as measured by RMSE and time-series correlations with NASS statistics, improves with scale as we aggregate from the grid cell to the county, state, and national level. To explore the variability in crop yield measures introduced by the differences in commonly used historical weather products, we simulate maize over the conterminous US using 4 different historical data products spanning 1980-2009, comprised of different combinations of observation- and reanalysis-based data products (NCEP CFSR, NOAA CPC, and NASA SRB). Finally, we use the multi-scale NASS RMSE and time-series correlation measures, along with temperature and precipitation indices, to assess the applicability of each data product for agricultural yield and impact simulations and estimate their accuracies at a variety of spatial scales. We focus specifically on the additional error introduced in the simulations by using reanalysis-based precipitation data in place of observation-based products, with the goal of extending this analysis globally, including regions where observation-based products are scarce or non-existent.