A Bayesian Space–Time Hierarchical Model for Remotely Sensed Lattice Data Based on Multiscale Homogeneous Statistical Units

The Bayesian hierarchical model has the outstanding capacity to combine prior knowledge with observations, as well as to generate rich spatiotemporal patterns. However, this advanced method has limited use in remotely sensed lattice data applications due to the large computational burden and unreasonable statistical inferences based on statistical units with a fixed scale. This paper presents a multiscale spatial homogeneous subdivision method and develops a Bayesian space-time hierarchical model (BSTHM) for remotely sensed lattice data. This can solve the above-mentioned limitations by constructing multiscale spatial homogeneous units with good statistical properties. The quantitative criteria for subdivision are provided. The outcome of the BSTHM is not only more reasonable in theory but also much easier to interpret; meanwhile, the computational efficiency is also considerably improved. This novel approach and its merits are illustrated by a case study that examines the spatiotemporal variation of PM
_2.5
pollution in Asia from 2000 to 2014 using remotely sensed data describing PM
_2.5
annual mean concentrations. Overall spatial pattern, common time trend, and local variation trend were decomposed and quantificationally estimated from an intricate spatiotemporal process.