시공간적 변동이 공간·시공간 크리깅 추정에 미치는 영향

Spatial kriging interpolates attributes at unobserved locations based on spatial autocorrelation, while spatio-temporal kriging extends it by incorporating temporal dimensions to reflect spatio-temporal interactions. Previous studies have shown that applying spatio-temporal kriging to spatio-temporal data improves estimation accuracy. However, because spatio-temporal kriging needs to simultaneously model both spatial and temporal autocorrelation, the nature of spatio-temporal variability may influence its accuracy. This study explores the impacts of spatio-temporal variations on the estimations of spatial and spatio-temporal kriging. In addition, it investigates the effects on kriging models that incorporate machine learning. Specifically, the study employs machine learning techniques with varying levels of flexibility, including random forest and boosting, in addition to conventional polynomial regression, to estimate first-order effects, while second-order effects are modeled using kriging. Kriging is applied to estimate nitrogen dioxide (NO2) concentrations in Seoul during 2023, and estimation accuracy is assessed across two time points with distinct spatio-temporal variation. The analysis results show that estimating first-order effects with machine learning improves overall accuracy, with more flexible models yielding higher performance. When spatio-temporal variation is relatively smooth, spatio-temporal kriging performs better, whereas on days with high variability, spatial kriging, which considers only spatial variation at a single time point, demonstrates superior accuracy. Particularly, spatio-temporal kriging tended to overestimate concentrations overall at time points with high variability, while models without first-order effect estimation showed higher accuracy. In addition, the spatial distribution of the estimated concentrations was divided into two patterns depending on the method used to estimate the first-order effect: one exhibiting smooth and gradual trends, and the other showing abrupt and localized variations. This study empirically demonstrates how spatio-temporal variation and first-order estimation methods affect kriging-based prediction accuracy, underscoring the importance of selecting an appropriate kriging method based on the characteristics of spatio-temporal variation.