Spatially modelling temperature normals in the Rocky Mountains with kriging and cokriging estimators using ANN produced secondary information

Multivariate geostatistical procedures have been used to interpolate/extrapolate climate fields in mountainous regions using strictly station data and spatially exhaustive secondary information, such as DEM derived topographic data One problem often encountered in such endeavours is the non-linearity of climate-secondary variable relationships and the variability of those associations in space. Poor overall correlations between the climate-secondary variables tend to minimize the influence of the secondary data in multivariate geostatistical analysis. This study used collocated cokriging, modified residual kriging, and kriging with external drift to spatially predict air temperature fields in southern Alberta and SE British Columbia, Canada. The primary data were station air temperature normals, while the secondary data were air temperatures predicted using ANN analysis. The independent variables used in ANN were topographic attributes derived from digital elevation models (e.g., elevation) and spatial indicator variables, such as windward/leeward location, chinook zones, etc. Data mining tools were also assessed for their ability to objectively delineate these indicator spatial variables. This study compared the results of using ANN produced secondary data versus those generated using linear techniques.