Deep Learning on Satellite Imagery Reveals Landscape Drivers of Metabolic Syndrome

Abstract Background: Understanding how environmental contexts influence metabolic syndrome (MetS) is limited by conventional exposure metrics that fail to capture complex, visually perceived neighborhood features. We aimed to develop a satellite-based deep learning framework to characterize fine-scale environmental exposures and examine their associations with MetS. Methods: We developed VISTAnet to extract latent geographic features from 49,050 high-resolution satellite images and identified eight landscape types through unsupervised clustering. 1,34,675 adults from Deqing, China, were included in this study. For each participant, environmental exposure was quantified as the Isochrone Landscape Proportion (ILP) within a 15 min walking isochrone around the residence. Associations between ILP and MetS were evaluated using mixed-effects logistic regression models. Stratified analyses were conducted by age, sex, educational level, employment status, and community type, and sensitivity analyses were conducted to assess the robustness of the findings. Results: Greater exposure to nature-dominated environments was associated with lower odds of MetS. Participants in the highest exposure quartile had significantly lower odds of having MetS for Dense Vegetation (adjusted odds ratio [aOR] 0.818, 95% CI 0.728–0.918), Wetland (0.941, 0.901–0.983), and Dry Farmland (0.838, 0.798–0.881). In contrast, higher exposure to Dense Residential Area (1.101, 1.058–1.145) and Industrial Zone (1.203, 1.151–1.257) was associated with increased odds of MetS. Findings were robust in sensitivity analyses. Conclusions: Nature-dominated landscapes were linked to lower MetS risk, whereas dense residential and industrial environments were linked to higher risk. This framework may support scalable environmental health assessment and health-oriented urban planning.