Spatial analysis of hourly surface air temperature - Entering the gate to spatiotemporal statistical modelling

Most of the current climate grid datasets are bounded at a time resolution of one day. Particularly in the case of surface air temperature, with its pronounced sub-daily variation, this bound restricts the utility of datasets for environmental modelling. Processes with a non-linear temperature dependence, such as the melting of snow or the transpiration of plants, cannot be accurately represented with daily mean values only. Therefore, advancing into the sub-daily time resolution is highly desirable. But it is also challenging: In order to predict realistic temporal evolutions for unobserved locations, one needs to integrate observations from a whole period, not just one instant. This poster presents early results from a new project that aims at harnessing methods of spatiotemporal statistical modelling for the development of an hourly temperature grid dataset over Switzerland.

We propose and experiment with a dynamic spatiotemporal statistical model that is conceived in the framework of dynamical linear models. Conceptually, it is an extension of kriging with external drift, familiar in spatial climatology, but with time-varying and serially correlated trend coefficients. The coefficients represent characteristic temporal variations in the temperature field, such as a diurnal cycle with a gradually varying amplitude/phase, or the lifting/sinking of a temperature inversion, in response to observations. We apply the method to hourly temperature data from the operational station network in Switzerland and investigate its performance for selected weather episodes (incl. sunny summer days, a winter-time inversion period). Our results hint to more/less successful configurations of the model for hourly temperature gridding in complex terrain. Comparison against the sequential application of a purely spatial analysis of the data shows some added value of the spatiotemporal approach, i.e. a benefit of borrowing information both over space and time. We anticipate that this benefit will increase as we enhance the complexity of the model configuration later in the project.