Generalized conceptualization of the optimal contextual unit in measuring green space exposure: The role of machine learning

Recent advances in environmental exposure measures have highlighted the optimal contextual unit that minimizes contextual errors in human health outcome modeling. This study provides a generalized articulation of the optimal contextual unit and also the critical role of machine learning for automatic and numerical solutions. We employed a green space health study in Hong Kong as an example to articulate our generalized analytics. 481 participants from Sham Shui Po, Kwun Tong, and Kwai Tsing reported their sociodemographic characteristics, self-reported health status, and home locations. Residential green space exposures were derived from remote sensing data at 3 m spatial resolution and variable buffer sizes around home locations. The optimal buffer sizes around home locations were automatically derived using gradient descent algorithms and overall health outcome modeling. Two optimization strategies were developed. First, the brute force strategy disclosed the optimal buffer size in each community at the theoretically highest spatial resolution. The optimal buffer sizes were found to be 111 m in Sham Shui Po, 198 m in Kwai Tsing, and 375 m in Kwun Tong. Second, a multi-scale strategy further reduced the time cost by 48% to 85%. This study provided a practical solution to find the optimal contextual unit (e.g., buffer zone) for high-quality and effective residential exposure measurements with the highest spatial accuracy and strongest statistical power. Consequently, this study may effectively contribute to a range of environmental health and health geography studies that need accurate exposure measurements of green space or many other environmental factors.