Evaluating uncertainty introduced to process-based simulation model estimates by alternative sources of meteorological data

Models that represent biophysical processes in hydrology, ecology and agricultural systems, when applied at specific locations, can make estimates with significant errors if meteorological input data are not representative of the sites. This is particularly important where the estimates from the models are used for decision support, strategic planning and policy development, due to the impacts of introduced uncertainty. This paper investigates the impacts of meteorological data sources on a cropping systems simulation model's estimate of crop yield, and quantifies the uncertainty that arises when site-specific weather data are not available. In the UK, as elsewhere, many meteorological stations record precipitation and air temperature, but very few also record solar radiation, a key driving input data set. The impacts of using on-site observed precipitation and temperature with estimated solar radiation, and off-site entirely observed meteorological data were tested on the model's yield estimates. This gave two scenarios: on-site observed versus partially modelled data; and on-site observed versus substitute data from neighbouring sites, for 24 meteorological stations in the UK. The analysis indicates that neighbouring meteorological stations can often be an inappropriate source of data. Of the 24 stations used, only 32% of the nearest neighbours were able to provide the best substitute off-site data. On-site modelled data provided better results than observed off-site data. The results demonstrate that the range of alternative data sources tested are capable of having both acceptable and unacceptable impacts on model performance across a range of assessment metrics, i.e. on-site data sources each produced yield over- or under-estimate errors greater than 2 t ha^-1. A large amount of uncertainty can be introduced to the model estimates due to the data source. Therefore, the applications of models that represent biophysical process where meteorological data are required, need to include the quantification of input data errors and estimate of the uncertainty that imperfect data will introduce.