Evaluating the suitability of large-scale datasets to estimate nitrogen loads and yields across different spatial scales

•Poorly monitored areas hinder the application of large-scale catchment models.•Large-scale models identify knowledge gaps across spatial scales when data are scarce.•Large-scale models performance declines in areas with low spatial variability.•A mix of local and large-scale models improves water quality management planning. Decision makers are often confronted with inadequate information to predict nutrient loads and yields in freshwater ecosystems at large spatial scales. We evaluate the potential of using data mapped at large spatial scales (regional to global) and often coarse resolution to predict nitrogen yields at varying smaller scales (e.g., at the catchment and stream reach level). We applied the SPAtially Referenced Regression On Watershed attributes (SPARROW) model in three regions: the Upper Midwest part of the United States, New Zealand, and the Grande River Basin in southeastern Brazil. For each region, we compared predictions of nitrogen delivery between models developed using novel large-scale datasets and those developed using local-scale datasets. Large-scale models tended to underperform the local-scale models in poorly monitored areas. Despite this, large-scale models are well suited to generate hypotheses about relative effects of different nutrient source categories (point and urban, agricultural, native vegetation) and to identify knowledge gaps across spatial scales when data are scarce. Regardless of the spatial resolution of the predictors used in the models, a representative network of water quality monitoring stations is key to improve the performance of large-scale models used to estimate loads and yields. We discuss avenues of research to understand how this large-scale modelling approach can improve decision making for managing catchments at local scales, particularly in data poor regions.