Classifying type-specific green space using satellite imagery and LiDAR to assess associations with diabetes prevalence

Source abstract

Abstract Urban green space is often measured by using greenness indices, such as the Normalized Difference Vegetation Index, but these indices do not distinguish among green space types. This study integrated 3-metre spatial resolution multispectral imagery and LiDAR-derived normalized Digital Surface Models (nDSM) to classify type-specific green spaces across Mississauga, Ontario, and examine their associations with diabetes prevalence. We tested eight combinations of spectral bands, vegetation and water indices, and structural data using an object-based random forest and found that the best-performing model included all the bands, indices, and nDSM, achieving 88% overall classification accuracy. The classification identified deciduous trees, coniferous trees, shrubs, grass, and marsh/wetland, which were then linked to census tract-level diabetes prevalence using local and neighbourhood beta regression models by sex and age group, and adjusted for socioeconomic covariates. Deciduous trees, coniferous trees, and shrubs were more consistently associated with lower diabetes prevalence, particularly among adults aged 45 years and older. NDVI-based models showed weaker associations with diabetes prevalence after adjustment, while type-specific models retained associations. These findings suggest that type-specific green space and structure provide more informative measures of exposure than greenness alone. Integrating high spatial resolution multispectral imagery and LiDAR provides a methodological framework for health research examining relationships between urban green space and chronic diseases.