Forecasting of air pollution at unmonitored sites

We address the problem of forecasting air-pollution at a site where there is no monitoring station by constructing data-driven models. We assume synchronous measurements of pollution are available at other sites in the vicinity, and that the spatial correlation carries information relevant for prediction. A Gaussian Process type spatial model is assumed, for interpolating pollution, and the time variation of the hyperparameters of the GP model is considered. An illustration of the method on synthetic data is presented.