Quantifying the uncertainty in spatially-explicit land- use model predictions arising from the use of substituted climate data

There can often exist a significant gap between the sophistication of land use systems models and our ability to provide the required biophysical input data. This can be particularly significant when systems models are used as components for decision support. The lack of spatially-explicit input parameters means that models’ site-specific predictions have potentially large uncertainties that are frequently unquantified. For crop production models, solar radiation (SR) is a significant parameter that is often sparsely measured. In the absence of site-specific data, modellers often substitute data from other sources: the nearest meteorological station; data derived from measurements made on the site (e.g. solarradiation interpolated from temperature and rainfall); or data derived from weather generators. This paper investigates the impact on a land use model of substituting on-site measured data with that from sites at increasing distances. This analysis quantifies the changes in both data bias and model variability introduced by the process of data substitution and forms a baseline against which it is possible to evaluate the alternative methods of data provision. To measure the relationship of changing uncertainty with distance to a data source, a database of observed climate data was created for 24 locations in the U.K. The root mean square error (RMSE) between each location’s SR data was calculated for each day of the year for corresponding years. The CropSyst crop production model was used to calculate yield, for a standardised spring barley scenario, for each year of available climate data for the 24 locations. The RMSE and each site’s yield estimates were then compared to determine the impact. The results show that there can be a complex relationship between the rate of decay in data similarity and impact on model output, governed by factors such as the density of the network of measurement locations and topography. Fundamentally, however, the results show that data substitution methods have a profound impact on the reliability of model results.