An Adaptive Gradient GWR Model for China's Building Carbon Emissions

ABSTRACT

Accurately quantifying nonstationary correlations between geographic variables underpins spatial process analysis, yet traditional geographically weighted regression (GWR) with isotropic neighborhoods cannot depict complex anisotropic spatial patterns. This study develops an adaptive gradient geographically weighted regression (AgGWR) to better model anisotropic and nonstationary spatial associations. Relying on local coefficient surface gradients, AgGWR establishes elliptical anisotropic neighborhoods with point‐wise adaptive geometric shapes, offering clear geometric interpretability for anisotropy and fine‐scale spatial analysis. A robust estimation algorithm is further integrated to suppress noise and outliers, boosting model stability. Multi‐scenario simulation experiments verify AgGWR's superiority over standard GWR: It raises average R ² by 4%, cuts RMSE by 26%, and reduces AIC by 120. Applied to China's urban building carbon emissions, AgGWR captures directional heterogeneity in emission driving factors, attaining an R ² of 0.85, alongside 3.37% lower RMSE and 19.4% smaller AIC relative to conventional GWR.