Fast Spatiotemporal Weighted Regression and its Applications

Keywords: spatiotemporal weighted regression, spatial analysis, spatial heterogeneity, spatiotemporal heterogeneity
Abstract Type: Poster Abstract

Abstract

Spatiotemporal Weighted Regression (STWR) is an extension of Geographically Weighted Regression (GWR) to explore geospatial processes with both spatial and temporal heterogeneity. It employs a new concept of temporal distance, which is the numerical difference rate between an observation point and a regression point over a time interval, rather than the time interval as temporal distance as in other geographically and temporally weighted regression models. Meanwhile, it utilizes a new spatiotemporal kernel in the form of a weighted average. These new configurations allow it to better capture local spatiotemporal heterogeneity, especially for those processes with spatial heterogeneity in the rate of change of the response variable. Despite these surprising improvements, model calibration becomes time-consuming and may even be impossible to solve. We use matrix partition and cache-to-MPI approach to calibrate the STWR in parallel. This can shorten the time for model calibration, especially in the face of large-scale spatiotemporal data points. Through simulation experiments and application in real meteorological environments and air quality data, the results show that this method is feasible and effective. This parallel improvement enables STWR to be used for large-scale spatiotemporal data points.