An autoencoder-based framework for analyzing regional variations in urban green space demand

Urban green spaces are pivotal for mitigating environmental challenges and enhancing urban livability, yet existing methods for assessing demand often neglect multidimensional interactions and nonlinear relationships. This study introduces an autoencoder-based framework to analyze regional variations in urban green space demand, integrating ecological and social indicators—land surface temperature (LST), carbon dioxide concentration, and population density—through a deep learning approach. Focusing on Chengdu’s central urban area, we employed Gaussian two-step floating catchment area (Ga2SFCA) methods to quantify demand across accessibility, heat island mitigation, and carbon sequestration, followed by autoencoder-driven feature extraction and k-means++ clustering. Results revealed distinct spatial heterogeneity: carbon sequestration demand concentrated in high-emission urban cores, heat island mitigation demand peaked in peripheries with elevated LST, and accessibility deficits dominated densely populated zones. The autoencoder outperformed traditional PCA, achieving a reconstruction error of 4.71 × 10⁻⁵ versus PCA’s 3.01 × 10⁻³, and captured nonlinear interactions among variables through interpretable latent features. Our framework provides a spatially refined, data-driven tool for optimizing green space allocation, addressing climate resilience, and prioritizing equity in urban planning. This work advances sustainable urban development by unifying ecological and social dimensions, offering actionable insights for policymakers to balance resource constraints with growing environmental pressures.