Visual information from street view imagery (SVI) can be used to extract specific exposure data or be linked directly to health outcomes. Building on our phase 1 foundations (Yuan et al.), this project deploys advanced segmentation and object detection to improve exposure modelling and epidemiological research.
During the first phase of Exposome-NL, we mined Google Street View images from all Dutch municipalities with over 50,000 inhabitants. Using a pyramid scene parsing network (PSPNet) pre-trained on the ADE20K dataset, we segmented these images into 150 classes to calculate pixel percentages for features like trees, grass, and roads. This data has already been used to refine environmental exposure estimates and study their health impacts.
Moving forward, we are expanding our framework by implementing zero-shot object detection to identify specific environmental elements, including air pollutants, heat, noise and the food environment. Additionally, we will deploy deep learning to agnostically link images to health datasets, uncovering novel visual features that explain variations in health effects.
The environment we live in has a dominant impact on our health. It explains an estimated seventy percent of the chronic disease burden. Where we live, what we eat, how much we exercise, the air we breathe and whom we associate with; all of these environmental factors play a role. The combination of these factors over the life course is called the exposome. There is general (scientific) consensus that understanding more about the exposome will help explain the current burden of disease and that it provides entry points for prevention and ...
Read More
