Evaluating the Correlation Between Impacting Factors and Land Surface Temperature via Spatial Regression Model and Random Forest

Abstract

Identifying accurate correlations between land surface temperature (LST) and its contributing factors is crucial for mitigating LST. Current studies, however, lack precise methods to determine key impacting factors for correlation analyses. Additionally, existing dominant analysis methods, which are spatial regression models and random forest (RF), have strengths and weaknesses, with insufficient comparative studies to determine their suitability for LST analysis. This study, informed by comprehensive literature reviews, identifies key factors affecting LST. Using Singapore as the study site, both multiscale geographically weighted regression (MGWR; spatial regression) and RF models were employed to analyze these factors, considering spatial variance and nonlinear correlations. The results demonstrate high accuracy, with R 2 values of 89 percent for MGWR and 81 percent for RF. The analyses reveal that, in terms of impact magnitude, normalized difference built-up index (NDBI) and normalized difference vegetation index (NDVI) significantly affect LST. Regarding the direction of the correlations, NDBI is positively correlated with LST in both models, whereas green view index (GVI) consistently shows negative correlations with LST. Other variables show variable correlations depending on location, reflecting context-dependent impacts. Based on these results, to mitigate LST, NDBI should be decreased by converting builtup areas to green spaces where feasible. Increasing GVI is also recommended to increase the proportion of green space from a street view, thereby lowering LST.