Multiscale heteroscedastic multivariate spatio-temporal models

Technological advances are generating massive multivariate spatio-temporal datasets. However, modeling of such highly structured multivariate datasets is not trivial because computations may scale cubicly with spatio-temporal dataset size. Here, we propose a fast and scalable Bayesian dynamic multivariate multiscale spatio-temporal framework that accounts for the spatio-temporal and inter-variables correlation structures. Our framework builds on recently proposed Bayesian dynamic multiscale spatio-temporal models for multivariate data. Here we extend these models to the heteroskedastic case when the covariance matrices at the finest resolution level vary spatially. Finally, we apply our proposed framework to model tropospheric temperatures at different altitudes. Our results indicate persistent long-term trends for temperatures at the upper and middle tropospheric altitudes.