Gridding of Hourly Surface Air Temperature – Application of a Spatio-Temporal Statistical Model in Complex Terrain

At present, most operational climate grid datasets are available with a temporal resolution of one day. This can be limiting in applications that rely on an explicit representation of the diurnal cycle, such as when modelling temperature-dependent environmental processes (e.g. transpiration). In this study we propose a methodology for the construction of hourly datasets of surface air temperature and present results from its application over Switzerland. The method uses a spatio-temporal (ST) statistical model. Unlike classical interpolation, the ST approach exploits data from a period of times simultaneously, and therefore allows to formally represent the connection of spatial and temporal variations. This is particularly desirable in complex terrain, where the diurnal cycle of temperature has marked topographic imprints. As ST model we use a dynamic linear model (DLM), which is a conceptual extension of kriging with external drift (KED), and offers high flexibility to be configured to the specifics of a region. In our application, for example, the DLM incorporates model components for cold-air pooling, basin-scale inversions, and the lake- and valley effects.

We present results from an application of the method to several multi-day episodes with specific weather conditions and from a continuous application over several months. The results illustrate that the method is capable to represent complex spatio-temporal variations, including the build-up and decay of an inversion and distinct diurnal variations in valleys, the flatland and at lake shores. Comparison of the spatio-temporal DLM to a spatial-only KED shows added value of DLM in terms of enhanced temporal continuity of trend coefficients. However, cross-validation statistics are not significantly different between the two approaches. Further experiments suggest that this is due to the dense station coverage in Switzerland. In regions / times where data is sparser, or when model complexity is substantially increased, spatio-temporal modelling is expected to provide a measurable improvement over spatial-only modelling. Our presentation will illustrate the results of our application in films of the temperature evolution. We will also introduce the extensions, both methodological and technical, implemented for the continuous multi-month application, and quantify the predictive performance of the method in this setting.

Our study provides valuable insight on extending a familiar interpolation concept (KED) into a spatio-temporal model, suitable to produce sub-daily temperature grid datasets.