A Geographically Weighted Regression Analysis of the Underlying Factors Related to the Surface Urban Heat Island Phenomenon

Abstract

This study investigated how underlying biophysical attributes affect the characterization of the Surface Urban Heat Island (SUHI) phenomenon using (and comparing) two statistical techniques: global regression and geographically weighted regression (GWR). Land surface temperature (LST) was calculated from Landsat 8 imagery for 20 July 2015 for the metropolitan areas of Austin and San Antonio, Texas. We sought to examine SUHI by relating LST to Lidar-derived terrain factors, land cover composition, and landscape pattern metrics developed using the National Land Cover Database (NLCD) 2011. The results indicate that (1) land cover composition is closely related to the SUHI effect for both metropolitan areas, as indicated by the global regression coefficients of building fraction and NDVI, with values of 0.29 and −0.74 for Austin, and 0.19 and −0.38 for San Antonio, respectively. The terrain morphology was also an indicator of the SUHI phenomenon, implied by the elevation (0.20 for Austin and 0.09 for San Antonio) and northness (0.20 for Austin and 0.09 for San Antonio); (2) the SUHI phenomenon of Austin on 20 July 2015 was affected by the spatial pattern of the land use and land cover (LULC), which was not detected for San Antonio; and (3) with a local determination coefficient higher than 0.8, GWR had higher explanatory power of the underlying factors compared to global regression. By accommodating spatial non-stationarity and allowing the model parameters to vary in space, GWR illustrated the spatial heterogeneity of the relationships between different land surface properties and the LST. The GWR analysis of SUHI phenomenon can provide unique information for site-specific land planning and policy implementation for SUHI mitigation.

Spatial Patterns of Land Surface Temperature and Their Influencing Factors: A Case Study in Suzhou, China

“…While global regression techniques have widely been used in SUHI analysis, the SUHI distribution is recognized as spatially varying and appears to be non-stationary across space. To address the spatial heterogeneity in LST, and its spatially varying relationships with land cover, local regression methods, such as geographically weighted regression, are suitable choices [9,23].…”

“…These studies substantially improved our understanding of the spatial patterns of both LST and SUHI, and helped further explore the relationships between LST and its influencing factors.Identification of the dominant factors affecting LST is important in understanding LST dynamics, and is helpful in alleviating the SUHI effect and improving the quality of the urban environment. LST is closely linked to land-use patterns, land coverage, landscape structures, and land-use configuration [1,[22][23][24]. Rapid urban growth has led to major land-use pattern change, altering the underlying surface [25,26].…”

Impacts of Land Cover/Use on the Urban Thermal Environment: A Comparative Study of 10 Megacities in China

“…Taking into account the synthetic circumstances of land cover composition (LCC) and population density (PD), we conducted the global-based (OLS) and local-based (GWR) analysis to explore the relationship between LST-LCC-PD in each megacity. Ample research on the UHI phenomenon has verified that the GWR model is superior for explaining the formation of SUHI than are other regression models from a local perspective [4,31,32]. Referring to the optimal observation scale selection of spatial regression analysis, we created a 1 × 1 km fishnet grid cell to provide sampling units because of the minimization of spatial dependence and autocorrelations, as well as the reservation of sufficient pattern information [32][33][34].…”

“…Ample research on the UHI phenomenon has verified that the GWR model is superior for explaining the formation of SUHI than are other regression models from a local perspective [4,31,32]. Referring to the optimal observation scale selection of spatial regression analysis, we created a 1 × 1 km fishnet grid cell to provide sampling units because of the minimization of spatial dependence and autocorrelations, as well as the reservation of sufficient pattern information [32][33][34]. Here, the average SUHII values of each grid were extracted as the dependent variables, by subtracting the average NDLST value of non-IS pixels from the NDLST value of each pixel within each study region [31,77].…”

Quantifying the impact of trees on land surface temperature: a downscaling algorithm at city-scale

“…Several studies have focused specifically on the effects of trees, comparing temperatures in a site with trees with those of a nearby site without trees (Bowler, Buyung-Ali, Knight, & Pullin, 2010). The most used approach for investigating the role of urban trees on land surface temperature (LST) is based on multivariate spatial statistical models in which the normalized difference vegetation index (NDVI) is used as an explanatory variable (see Weng, Lu, & Schubring, 2004;Yuan & Bauer, 2007;Guo et al, 2015;Bonafoni, Anniballe, Gioli, & Toscano, 2016;;Zhao, Jensen, Weng, & Weaver, 2018). According to the literature, urban morphology also has an important role in shaping urban heat island phenomena; the relationship between LST and urban morphology has recently been analyzed using the Sky View Index (SVI) calculated based on 3D city models, which are derived from LiDAR data (Kokalj, Zakšek, & Oštir, 2011;Chun & Guldman, 2014;Tan, Lau, & Ng, 2016;Nakata-Osaki, Souza, & Rodrigues, 2018).…”