BACKGROUND
The prevalence of obesity has reached alarming levels globally, necessitating innovative approaches to understand and address this complex health issue. Spatial technologies such as Geographic Information Systems (GIS), remote sensing, spatial machine learning, and spatial analysis have emerged as powerful tools in obesity research, offering novel insights into the environmental and social determinants of this epidemic. This study leverages these technologies and deep learning to predict the obesity rate for every census tract in Missouri.
OBJECTIVE
Our research aims to utilize deep convolutional neural networks (DCNNs) to examine medium-resolution satellite imagery for predicting obesity rates. Focusing on 1,052 census tracts in Missouri, USA, the project seeks to provide a scalable method for predicting obesity prevalence using deep neural visual features (DNVFs) and geospatial modeling. This approach is intended to enhance the accuracy of public health initiatives and inform policy decisions.
METHODS
Our study employed a three-step analysis. First, Sentinel-2 satellite images were processed to extract features of the built environment using ResNet-50. These images were cut into 63,592 chips of 224x224 pixels. Second, 1,052 polygons of census tracts from the Tiger Line shapefiles were merged with obesity rates data obtained from the CDC. Third, a spatial lag model was used to predict obesity rates and assess the association between the DNVFs and obesity prevalence. The study used spatial autocorrelation measures to identify clusters of high and low obesity rates and spatial outliers.
RESULTS
The analysis reveals significant spatial clustering of obesity rates across Missouri, with a positive Moran’s I value of 0.68, indicating that census tracts with a similar obesity rate are located near each other. The spatial lag model demonstrates strong predictive performance, with an R² value of 0.93 and a spatial pseudo-R² of 0.92, indicating that the model explained 93% of the variability in obesity rates. The LISA analysis identifies specific regions with high-high and low-low clusters of obesity rates, which are visualized using choropleth maps. These findings highlight the potential of integrating satellite imagery and advanced machine learning techniques to capture environmental features correlated with obesity.
CONCLUSIONS
This study underscores the effectiveness of using DCNNs and spatial modeling to analyze and predict obesity prevalence based on environmental features extracted from satellite imagery. The integration of deep neural visual features with spatial statistics provides a robust framework for understanding the spatial dynamics of obesity. The high predictive accuracy of the spatial lag model indicates its potential for informing targeted public health interventions and policies. Future research should expand the geographical scope and incorporate additional socio-economic and health-related data to further refine the model and enhance its applicability. This multi-modal approach offers a promising avenue for advancing obesity research and addressing this pressing public health challenge.