Comparison of interpolation methods for mapping climatic and bioclimatic variables at regional scale

Climatic data and bioclimatic indexes have been used to study plants, animals and ecosystem distribution. GIS-based maps of climatic and bioclimatic data can be obtained by interpolating values observed at measurement stations. Since no single method can be considered as optimal for all observed regions, a major task is to propose comparisons between results obtained using different methods applied to the same data set of climate variables. We compared three methods that have been proved to be useful at regional scale: 1 - a local interpolation method based on de-trended inverse distance weighting (D-IDW), 2 - universal kriging (i.e. simple kriging with trend function defined on the basis of a set of covariates) which is optimal (i.e. BLUP, best linear unbiased predictor) if spatial association is present, 3 - multilayer neural networks trained with backpropagation (representing a complex nonlinear fitting). Long-term (1955-1990) average monthly data were obtained from weather stations measuring precipitation (201 sites) and temperature (102 sites). We analysed twelve climatic variables (temperature and precipitation) and nine bioclimatic indexes. Terrain variables and geographical location have been used as predictors of the climate variables: longitude, latitude, elevation, aspect, slope, continentality and estimated solar radiation. Based on the root mean square errors from cross-validation tests, we ranked the best method for each variable data set. Universal kriging with external drift obtained the best performances for seventeen variables of the twenty-one analysed, neural network interpolator has proven to be more efficient for three variables and D-IDW for only one. Based on these results, we used the universal kriging estimates to produce the climatic and bioclimatic maps aimed at defining the bioclimatic envelope of species.