An interpretable stratified weighted network for uncovering spatially heterogeneous correlations between the built environment and nighttime urban vitality

Abstract

Understanding the correlations between the built environment and nighttime urban vitality (NTUV) is crucial for stimulating consumption and promoting sustainable urban development. However, existing models, influenced by spatial stratified and local heterogeneity, often suffer from local overfitting or underfitting, which makes it difficult to effectively capture the spatially heterogeneous correlations. In this study, we propose an interpretable stratified weighted network (SW-Net), which extends the traditional geographically weighted regression by introducing deep learning techniques. Specifically, the correlations are decomposed into shared patterns and local fluctuations. These components are estimated by the spatial stratification block and weight localization block in SW-Net, respectively, to model spatial stratified and local heterogeneity. SW-Net was applied to two case areas in Xiamen and Fuzhou and strongly outperformed eight baseline models. The results mainly show that (1) consumption activities are the key drivers of NTUV; (2) transportation infrastructure can facilitate nighttime human mobility and enhance NTUV; (3) nighttime cultural and tourism activities hold substantial potential to boost NTUV in both cities; and (4) a higher degree of land use mix does not necessarily lead to greater NTUV. The findings in this study can offer insights for advancing geographical relationship modeling and provide targeted policy suggestions for optimizing the built environment and enhancing NTUV.