Geostatistical modelling of air temperature in a mountainous region of Northern Spain

Air temperature is one of the most important factors affecting vegetation and controlling key ecological processes. Air temperature models were compared in a mountainous region (Asturias in the North of Spain) derived from five geostatistical and two regression models, using data for January (coolest month) and August (warmest month). The geostatistical models include the ordinary kriging (OK), developed in the XY plane and in the X, Y and Z-axis (OKxyz), with zonal anisotropy in the Z-axis (variogram fitting procedure developed in this study), and three techniques that introduce elevation as an explanatory variable: ordinary kriging with external drift (OKED) and universal kriging, using the ordinary least squares (OLS) residuals to estimate the variogram (UK1) or the generalised least squares (GLS) residuals (UK2). The OKED, UK1 and UK2 techniques were more satisfactory than OK in terms of standard prediction error and mean absolute error, which were inferior by 1 °C, but OKxyz improved the results obtained with those techniques. Moreover, OKxyz, OKED, UK1 and UK2 improved slightly the results of a regression model with UTM coordinates and elevation data as independent variables in terms of bias (R1); whereas a complex regression model, which includes altitude, latitude, distance to the sea and solar radiance as independent variables (R2), showed better results in terms of mean absolute error, under 0.16 °C for both months. A second validation carried out with stations discarded for the interpolation showed a greater similarity between the efficiency of R2 and the geostatistical techniques.