Estimation of Population Health Indicators Using Satellite Images

In all countries, the collection of data on public health takes place through surveys and it is both time-consuming and costly. Many studies show that the built environment affects the health of the population. At the same time, satellite images can serve as an indicator of the features of the built environment. In this study, the approach was proposed for assessing population health indicators based on satellite images. The approach consists of two stages: using convolutional neural networks to extract features of the built environment from satellite images, and then using extracted representations estimate population health indicators using regression models. The prevalence of physical inactivity of the population and the level of mental distress were used as indicators of the population's health. Satellite imagery was collected using the Google Maps Static API. Pretrained ResNet50, VGG16, VGG19, InceptionV3 models, and convolutional denoising autoencoder were used to extract features of the built environment. XGBoost and LightGBM models and Lasso regression were used to solve regression problems. It was concluded that the use of satellite imagery to estimate population health indicators using machine learning methods provides results comparable to traditional methods.