ABSTRACT

The geographical detector model (GDM) is based on the spatial variance analysis of geographical strata of variables to assess the association between the independent variables ($X$) and dependent variables ($Y$). The independent variables of the GDM must be discretized into classes. However, current discretization methods employ univariate analysis, which may lead to inaccurate results. The aim of this study was to develop a novel bivariate optimal discretization approach, known as the multiscale discretization (MSD) method. The objective of the MSD method is to determine an appropriate set of thresholds for $X$, thereby minimizing the variance of $Y$ within the spatial partitions determined by the discrete $X$. We successfully applied the MSD method to assess the relationship between the precipitation and enhanced vegetation index on the African continent, as well as the habitat range of pandas in Ya’an County, Sichuan Province, China. The results demonstrate that the MSD is a feasible, robust, and rapid method for converting continuous data into discrete data, with globally optimal discretization results. Furthermore, the MSD method can evaluate the degree of association between $X$ and $Y$ more accurately, and can optimize the results of the GDM.