Remote sensing metrics to assess exposure to residential greenness in epidemiological studies: A population case study from the Eastern Mediterranean

Introduction/aims: Application of remote sensing-based metrics of exposure to vegetation in epidemiological studies of residential greenness is typically limited to several standard products. The Normalized Difference Vegetation Index (NDVI) is the most widely used, but its precision varies with vegetation density and soil color/moisture. In areas with heterogeneous vegetation cover, the Soil-adjusted Vegetation Index (SAVI) corrects for soil brightness. Linear Spectral Unmixing (LSU), measures the relative contribution of different land covers, and estimates percent of each over a unit area. We compared the precision of NDVI, SAVI and LSU for quantifying residential greenness in areas with high spatial heterogeneity in vegetation cover. Methods: NDVI, SAVI, and LSU in a 300 m radius surrounding homes of 3,188 cardiac patients living in Israel (Eastern Mediterranean) were derived from Landsat 30 m spatial resolution imagery. Metrics were compared to assess shifts in exposure quartiles and differences in vegetation detection as a function of overall greenness, climatic zones, and population density, using NDVI as the reference method. Results: For the entire population, the dispersion (SD) of the vegetation values detected was 60% higher when greenness was measured using LSU compared to NDVI: mean (SD) NDVI: 0.17 (0.05), LSU (%): 0.23 (0.08), SAVI: 0.12 (0.03). Importantly, with an increase in population density, the sensitivity of LSU, compared to NDVI, doubled: There was a 95% difference between the LSU and NDVI interquartile range in the highest population density quartile vs 47% in the lowest quartile. Compared to NDVI, exposures estimated by LSU resulted in 21% of patients changing exposure quartiles. In urban areas, the shift in exposure quartile depended on land cover characteristics. An upward shift occurred in dense urban areas, while no shift occurred in high and low vegetated urban areas. Conclusions: LSU was shown to outperform the commonly used NDVI in terms of accuracy and variability, especially in dense urban areas. Therefore, LSU potentially improves exposure assessment precision, implying reduced exposure misclassification.