This paper aims to provide a wavelet analysis for spatio-temporal data which are observed on irregularly spaced stations at discrete time points, where the spatial covariances show serious non-stationarity caused by local dependency. A specific example that is used for the demonstration is US precipitation data observed on about ten thousand stations in every month. By a reinterpretation of Whittle likelihood function for stationary time series, we propose a kind of Bayesian regression model for spatial data whose regressors are given by modified Haar wavelets and try a spatio-temporal extension by a state space approach. We also propose an empirical Bayes estimation for the parameters, which is regarded as a spatio-temporal extension of Whittle likelihood estimation originally defined for stationary time series. We conduct the extended Whittle estimate and compare mean square errors of the forecasts with those of some benchmarks to evaluate its goodness for the US precipitation data in August from 1987-1997.
