The Dynamical Model Representation of Convolution-Generated Spatio-Temporal Gaussian Processes and Its Applications
By: Yutong Zhang, Xiao Liu
Potential Business Impact:
Tracks pollution changes from wildfires.
Convolution-generated space-time models yield an important class of non-separable stationary Gaussian Processes (GP) through a sequence of convolution operations, in both space and time, on spatially correlated Brownian motion with a Gaussian convolution kernel. Because of its solid connection to stochastic partial differential equations, such a modeling approach offers strong physical interpretations when it is applied to scientific and engineering processes. In this paper, we obtain a new dynamical model representation for convolution-generated spatio-temporal GP. In particular, an infinite-dimensional linear state-space representation is firstly obtained where the state transition is governed by a stochastic differential equation (SDE) whose solution has the same space-time covariance as the original convolution-generated process. Then, using the Galerkin's method, a finite-dimension approximation to the infinite-dimensional SDE is obtained, yielding a dynamical model with finite states that facilitates the computation and parameter estimation. The space-time covariance of the approximated dynamical model is obtained, and the error between the approximate and exact covariance matrices is quantified. We investigate the performance of the proposed model through a simulation-based study, and apply the approach to a real case study utilizing the remote-sensing aerosol data during the recent 2025 Los Angeles wildfire. The modeling capability of the proposed approach has been well demonstrated, and the proposed approach is found particularly effective in monitoring the first-order time derivative of the underlying space-time process, making it a good candidate for process modeling, monitoring and anomaly detection problems. Computer code and data have been made publicly available.
Similar Papers
Multivariate Low-Rank State-Space Model with SPDE Approach for High-Dimensional Data
Methodology
Speeds up air quality predictions by 93%.
Gaussian Process State-Space Modeling and Particle Filtering for Time Series Decomposition and Nonlinear Signal Extraction
Methodology
Learns hidden patterns in messy data.
Generative Modeling of Clinical Time Series via Latent Stochastic Differential Equations
Machine Learning (CS)
Predicts patient health with better accuracy.