Agriculture is a human economic system with a direct link to the physical environment. As a result, climate change associated with an elevated concentration of greenhouse gases will alter agricultural productivity. The size of the impact will depend on the degree to which farmers can adapt to these changes. To fully evaluate the impact of climate change on agricultural production, one must assess the physical effects of and adaptation to climate change. Our biophysical modeling system uses information from the physical and social environment to estimate corn and wheat yield for a sample of counties across the United States. Forecasts of climate change (physical environment) are used to estimate the physical impact of climate change on productivity. The effectiveness of farmer adaptation is explored through modifications to the social environment and planting decisions.
General circulation models (GCMs) have emerged as the leading tools for forecasting climate change. However, the spatial resolution of these models is not adequate to assess the impact of climate change on local level ecosystems. In recent years, methods have been developed to spatially interpolate GCM forecasts from model grid points to the local level. These methods fall into two general classes: (1) statistical down-scaling techniques, or (2) meso-scale meteorological models nested within GCM grids. Both approaches have significant drawbacks. Due to the complicated spatial and temporal structure of common meteorological data, statistical methods require a high degree of complexity to achieve a reasonable degree of accuracy. To improve the spatial resolution of GCMs using nested meso-scale models, vast computing time on state-of-the-art supercomputers is needed.
This paper describes a new methodology to spatially interpolate GCM forecasts using Artificial Neural Networks (ANNs). ANNs learn to represent highly non-linear relationships between input and output vectors through a highly interconnected set of simple computational units known as processing elements. In this analysis, we use multi-layer feedforward ANNs with sigmoidal non-linearities. Operating under a supervised training strategy, these ANNs utilize low resolution information representative of GCM forecasts as input vectors to make predictions for point locations at the local level (output vectors). Predictions are made for temperature and precipitation fields on a daily time step and are compared to estimates from traditional point estimation techniques. This new methodology has two main advantages over recent spatial interpolation techniques. First, ANNs represent the complicated spatial and temporal structure of temperature and precipitation fields with a relatively simple network architecture. Secondly, ANNs reduce the computational time needed to spatially interpolate GCM forecasts relative to nested meteorological models.