|
Abstract:
|
Numerical models are mathematical deterministic models used to predict spatio-temporal processes. For model evaluation and calibration, it is necessary to solve the spatial misalignment between the predictions yielded by numerical models, given in terms of averages over grid cells, with observations collected at sites. Here, we propose a multivariate spatio-temporal model that allows to downscale to point level multiple outputs from a numerical model, exploiting the possible correlation between the different outputs. The model regresses the observational data on the numerical model outputs using correlated Gaussian processes as coefficients in the linear regression. We explore different ways to characterize the correlation among the Gaussian processes. As an example, we apply our model to ozone and PM 2.5 concentration for the Eastern US.
|
