Interpolation of Hourly Surface Air Temperature—Application of a Spatio-Temporal Statistical Model in Complex Terrain

At present, grid datasets of in situ climate observations are mostly constructed in daily time resolution. An extension to sub-daily resolution is desirable, particularly for surface air temperature, in order to represent diurnal variations. Here, we devise and investigate a method to produce hourly temperature grids in a region of complex topography (Switzerland). The method, a dynamic linear model (DLM), extends a model widely used in spatial climatology (kriging with external drift, KED), into a spatio-temporal statistical model, while preserving its flexibility to adapt to the specifics of a region. We configure the method with covariates for known mesoscale effects (inversions, cold-air pooling, valley- and lake effects), assuming quasi-harmonic evolutions in the related coefficients. The method is applied to a wintertime inversion episode and a summertime high-pressure episode, with data from 259 stations. We compare DLM to a sequentially applied KED and evaluate the predictions in cross-validation. The results demonstrate that DLM can accurately reproduce complex spatio-temporal variations, including the gradual build-up of an inversion and the distinct temperature patterns between day- and night-time. The mean absolute error ranges from 0.65°C in flat and hilly terrain (both episodes) to 1.6°C in mountain valleys in winter. DLM yields temporally more continuous effect coefficients than KED, which corroborates the expected advantage from spatio-temporal modelling. However, the prediction errors of DLM are not significantly smaller, except in experiments where the station data was artificially reduced. It seems that the spatio-temporal approach had little advantage over a spatial model in the data-rich conditions of our application. However, its advantage may be more prominent in conditions with fewer observations or with more covariates. DLM is a versatile method for developing sub-daily climate datasets and this study provides valuable insight on its adoption to the climatic conditions of a specific region.