Spatiotemporal Analysis of Hourly Surface Air Temperature – an Application in Switzerland

Today, most climate data sets derived by spatial analysis of station data have a maximum time resolution of one day. This is limiting in environmental modelling applications. For example, plant transpiration and snow melt depend on temperature non-linearly. Modelling these processes using daily mean temperatures only is therefore questionable. Our objective is to improve this situation and to advance spatial analysis of temperature into the sub-daily time scale. At this high temporal resolution, surface temperature fields are highly organized not just spatially but also temporally. This requires a new modelling approach that allows information from observations to also flow over time, not just over space like in traditional spatial climate analysis.

An attractive framework for this methodological extension is spatiotemporal (ST) statistical modelling. In this presentation we propose, and experiment with, a dynamic linear model (DLM). A DLM can be thought of as an extension of Kriging with external drift (KED, a Gaussian spatial process), where the trend coefficients are assumed to be serially correlated, i.e. smooth in time. These time series represent the characteristic variations of the temperature field, such as a diurnal cycle with a varying amplitude and phase, or the gradual variation of the height and amplitude of a temperature inversion.

In our study we develop and configure a DLM for surface air temperatures in Switzerland. The model can represent non-linear temperature profiles, such as inversions or well-mixed boundary layers. Our configuration involves several profiles across the domain to account for large-scale variation and across-ridge contrasts of the temperature distribution. We also account for mesoscale cold-air pooling with a tailor-made predictor. All model components are allowed to vary quasi harmonically to account for the diurnal variation.

We apply the DLM to hourly temperature observations from about 200 stations in Switzerland. Experiments are made for a set of test episodes including typical weather conditions like summertime high pressure and wintertime inversion situations. We find that the DLM reproduces physically plausible temperature distributions and evolutions during these weather conditions. The developments are continuous and reveal, for example, the generation, diurnal oscillation and dissolution of an inversion layer over several days. We also find a quite robust performance with respect to erroneous and missing observations. In a comparison with a traditional spatial-only analysis (sequentially applied KED), the DLM exhibits some added value. This benefit will be demonstrated in the presentation. In summary, the DLM framework turns out to be flexible and promising for the development of sub-daily temperature datasets, in Switzerland and likely in other regions of the world.