Using deep learning to explore the impacts of street-view green space on school myopia prevalence: a multicenter, cross-sectional study

Abstract

Previous studies used the Normalized Difference Vegetation Index (NDVI) to assess green space and showed its potential protective effect on myopia. However, this top-down measurement of green space cannot fully reflect the eye-level greenness perception. This study used the Vision Greenness Index (VGI) to assess visual greenness from street-view images and compare its effect on myopia with NDVI. This multicenter, cross-sectional study investigated 69,051 children from 146 schools in five cities in Hubei, ranging from kindergarten to high school. Visual greenness was calculated from street-view images with DeepLab v3 and NDVI was calculated from Landsat 8 satellite observations. Regression models were used to analyze the relationship between school myopia and environmental factors. The partial correlation coefficient between VGI within a 5,000-m buffer zone and myopia was − 0.343 (P = 0.038) and − 0.326 (P < 0.001) in Models 1 (all schools) and 2 (schools except high schools), respectively, whereas NDVI showed no significant results. In Model 3 (kindergartens), the partial correlation coefficients between VGI and NDVI within a 5,000-m buffer zone were − 0.675 (P = 0.029) and − 0.380 (P = 0.032), respectively. In Model 4 (elementary schools), the partial correlation coefficients between VGI and NVDI within a 5,000-m buffer zone were − 0.310 (P = 0.027) and − 0.088 (P = 0.043), respectively. Model 5 (high schools) did not demonstrate any correlations between the identified factors and the prevalence of myopia during high school. VGI demonstrated a more significant relation to the myopia than NVDI. Visual greenness was found to be related to myopia in kindergartens and elementary schools rather than in high school, indicating its more significant association with a lower prevalence of myopia in early school ages than in high school.