Aim: Modelling complex environmental and ecological processes over large geographic areas is challenging, particularly when basic research and model development for such processes has historically been at the local scale. Moving from local toward global analysis brings up numerous issues related to data processing, aggregation, tradeoffs between model quality and data quality, and prioritization of data collection and/or compilation efforts. We studied these issues in the context of modelling emissions of N 2O (a potent greenhouse gas) from agricultural soils. Location: Global. Methods: We developed metamodels of the DeNitrification-DeComposition (DNDC) model, a mechanistic model that simulates greenhouse gas emissions from agricultural soils, to estimate global N 2O emissions from maize and wheat fields. We ran DNDC for a diverse sample of global climate and soil types, and fitted the model output as a function of (sometimes simplified) model input variables, using the random forest machine learning algorithm. We used the metamodels to estimate global N 2O emissions from maize and wheat at a very high spatial resolution ( c. 1 km 2) and examined the effects of different approaches of using soil data as well as the effects of spatial aggregation of soil and climate data. Results: The average coefficient of determination ( R2) between holdout data (DNDC output not used to construct the metamodel) and metamodel predictions was 0.97 for maize and 0.91 for wheat. The metamodels were sensitive to soil properties, particularly to soil organic carbon content. Global emission estimates with the metamodel were highly sensitive to the spatial aggregation and other forms of generalization of soil data, but much less so to aggregation of climate data. Main conclusions: Using a simplified metamodel with data of high spatial resolution could produce results that are more accurate than those obtained with a full mechanistic model and lower-resolution data.
