Spatial resolution of Normalized Difference Vegetation Index and greenness exposure misclassification in an urban cohort

BACKGROUND: The Normalized Difference Vegetation Index (NDVI) is a measure of greenness widely used in environmental health research. High spatial resolution NDVI has become increasingly available; however, the implications of its use in exposure assessment are not well understood. OBJECTIVE: To quantify the impact of NDVI spatial resolution on greenness exposure misclassification. METHODS: Greenness exposure was assessed for 31,328 children in the Greater Boston Area in 2016 using NDVI from MODIS (250 m(2)), Landsat 8 (30 m(2)), Sentinel-2 (10 m(2)), and the National Agricultural Imagery Program (NAIP, 1 m(2)). We compared continuous and categorical greenness estimates for multiple buffer sizes under a reliability assessment framework. Exposure misclassification was evaluated using NAIP data as reference. RESULTS: Greenness estimates were greater for coarser resolution NDVI, but exposure distributions were similar. Continuous estimates showed poor agreement and high consistency, while agreement in categorical estimates ranged from poor to strong. Exposure misclassification was higher with greater differences in resolution, smaller buffers, and greater number of exposure quantiles. The proportion of participants changing greenness quantiles was higher for MODIS (11-60%), followed by Landsat 8 (6-44%), and Sentinel-2 (5-33%). SIGNIFICANCE: Greenness exposure assessment is sensitive to spatial resolution of NDVI, aggregation area, and number of exposure quantiles. Greenness exposure decisions should ponder relevant pathways for specific health outcomes and operational considerations.

Automated summary

This study assessed the impact of different spatial resolutions of the Normalized Difference Vegetation Index (NDVI) on exposure misclassification of greenness. The results showed that coarser resolution NDVI led to higher greenness estimates, but the exposure distributions were similar. Continuous estimates showed poor agreement, while agreement in categorical estimates varied from poor to strong.