Evaluating the effect of 3D urban form on neighborhood land surface temperature using Google Street View and geographically weighted regression

Context: Land surface temperature (LST) directly responds to incoming solar radiation and is strongly influenced by vertical urban structures, such as trees and buildings that provide shade. Conventional LST-planar land-cover assessments do not explicitly address the vertical dimension of the “urbanscape” and therefore do not capture the heterogeneity of solar radiation exposure of planar surfaces adequately. Objectives: To fill this gap, this study compares and integrates novel spherical land-cover fractions derived from Google Street View (GSV) with the conventional planar land-cover fractions in estimating daytime and nighttime LST variations in the Phoenix metropolitan area, AZ. Methods: The GSV spherical dataset was created using big data and machine learning techniques. The planar land cover was classified from 1 m NAIP imagery. Ordinal least square (OLS) and geographically weighted regression (GWR) were used to assess the relationship between LST and urban form (spherical and planar fractions) at the block group level. Social-demographic variables were also added provide the most comprehensive assessment of LST. Results: The GSV spherical fractions provide better LST estimates than the planar land-cover fractions, because they capture the multi-layer tree crown and vertical wall influences that are missing from the bird-eye view imagery. The GWR regression further improves model fit versus the OLS regression (R² increased from 0.6 to 0.8). Conclusions: GSV and spatial regression (GWR) approaches improve the specificity of LST identified by neighborhoods in Phoenix metro-area by accounting for shading. This place-specific information is critical for optimizing diverse cooling strategies to combat heat in desert cities.