Prädiktion der Ozeantemperatur im räumlichen und zeitlichen Verlauf mit Hilfe dynamischer linearer Modelle

Atmospheric and oceanic processes representing important components of the global climate system display variability over both space and time. Applied scientific analysis may be based upon two competing strategies, physically derived deterministic modelling versus statistical approaches, from which the latter is utilized in the present case. Since observations typically constitute large data sets that often are spatially and temporally incomplete and exhibit complicated interactive structural relationships, traditional space-time methods are of limited use. Direct specification of the joint space-time covariance structure often is not possible due to the existence of spatial non-stationarities and nonseparable space-time interaction. In this paper dynamic linear (state-space) models are developed instead, that model the temporally dynamic structure in an autoregressive framework and additionally feature a spatially descriptive component. In order to handle large observational areas, dimension reduction of the spatial field is achieved by use of empirical orthogonal functions. The method is applied to a data set of measurements of the sea surface temperature in the Northwest European Shelf during 1983-1992. The observed point measurements are predicted to a grid of about 20km grid size (1/3^o in east-west direction and 1/5^o in north-south direction) by application of the Kalman filter. Unlike other similar spatiotemporal state-space formulations, the presented approach does not demand for temporally fixed measurement locations. Moreover it allows for a dynamic incorporation of a (large-scale) trend component and an efficient underlying step of parameter estimation is involved.