A copula model for non-Gaussian multivariate spatial data (Q1755126)

Consider a $p$-dimensional random process defined at locations $s\in\mathbb{R}^d$. The main task is to model dependence both within each variable and between different variables. The proposed model is based on the measurement of the multivariate random process \[ W_i(s)=Z_i(s)+\alpha_{i0}^U\varepsilon_0^U+\alpha_{i}^U\varepsilon_i^U- \alpha_{i0}^L\varepsilon_0^L-\alpha_{i}^L\varepsilon_i^L, \] $i=1,\dots,p$, $s \in \mathbb{R}^d$, where $Z_i(s)$ are cross-correlated Gaussian processes, $\alpha_{i0}^U, \alpha_{i}^U,\alpha_{i0}^L, \alpha_{i}^L$ are nonnegative parameters and $\varepsilon_0^U, \varepsilon_i^U,\varepsilon_0^L, \varepsilon_i^L$ are independent from each other and from all $Z_i(s)$ common factors with unit exponential distribution that do not depend on $s$. This model allows one to analyze spatial data with strong joint dependence in the tails and spatial data with reflection asymmetry. It can be seen as an extension of the linear model of coregionalization where the cross-covariance function can be constructed using a sum of independent processes. Notice that the same cross-covariance structure of the model can be obtained for various sets of parameters $\alpha_{i0}^U, \alpha_{i}^U,\alpha_{i0}^L, \alpha_{i}^L$. Therefore, the properly chosen copula model for the joint distribution of $W_1(s),\dots,W_p(s)$ allows one to determine the tail properties using repeatedly measured data. \par The likelihood of the model is obtained in a simple form. Spatial data can be easily interpolated at locations where values of variables are unknown. The model is applied to temperature and pressure data, the results are compared with those provided by the popular model from multivariate geostatistics.