Spatial interpolation of PM2.5 contaminant using Ordinary Kriging and Support Vector Machine (#600)

Authors

Pastén, Felipe

Irarrázabal, Constanza

Blazquez, Carola A.

Jiménez, Raquel

Abstract

Exposure to air pollution such as particulate matter less than 2.5 micrometers (PM2.5) can produce different types of disease. This study uses mobile measurements of PM2.5 contaminant due to wood burning during winter nights in the conurbation of Temuco and Padre Las Casas in southern Chile. The geostatistical tool Ordinary Kriging (OK) and machine learning Support Vector Machine (SVM) are employed to estimate an interpolated surface of PM2.5 in this conurbation. Overall, the results using OK indicate spatial variability of PM2.5 concentrations in the conurbation with high values toward the west and east areas of Temuco and some smaller areas of Padre Las Casas. However, the results of spatial interpolation with SVM vary depending on the method used to select the covariates. The best covariate selection for the SVM includes variables related to residential density and local roads within different buffer sizes. Cross-validation analysis suggests that OK outperforms the SVM algorithm when estimating the PM2.5 surface. In addition, the aforementioned results vary depending on the level of aggregation of the data. The results from this study may be used by authorities to implement environmental actions in areas with high PM2.5 concentrations, and properly allocate resources to reduce air pollution in these areas. Future research should include the implementation of other types of machine learning techniques and the use of additional variables that may impact the generation of PM2.5 from wood burning.