Measuring Neighborhood Socioeconomic Status from Sky and Street: An Ensemble Learning Framework.

Abstract

Fine-scale, geocoded neighborhood socioeconomic status (nSES) data are foundational inputs for urban studies and spatial inequality research, yet remain scarce in emerging economies that together account for over 70% of the global population. Here, we use publicly available satellite and street view imagery to estimate nSES at fine spatial resolution, proposing an ensemble regression framework that unites multi-view images to measure intra-urban inequality. We estimate eight indicators - spanning household registration (hukou), housing, income, employment, and education - as numeric values rather than broad quantile bands, enabling direct cross-neighborhood comparison. Family and individual SES from a representative survey serve as training data. The proposed fusion model outperforms single ensemble regression models and single-view image-based models in mean absolute percentage error. To interpret image-based prediction, we apply SHAP-based feature attribution and modality-contribution analysis, identifying the built-environment features that drive predictions and revealing that satellite and street view imagery play complementary rather than redundant roles. Neighborhood inequalities at traffic analysis zone and township scales in central Beijing are visualized for the first time. We further transfer the trained model to Shanghai, where no comparable public nSES dataset exists, providing a limited proof-of-concept against real-estate data for two of the eight indicators and outlining a replicable pathway for data-scarce cities.