SatNet: A Low-Cost, Neural-Network based Algorithm Utilizing Satellite Images for Disease Hotspot Detection

Abstract

The rapid spread of infectious diseases poses a significant global health challenge, requiring timely and accurate detection for effective intervention. Traditional disease detection services, such as the Centers for Disease Control and Prevention (CDC) and the World Health Organization (WHO), play a crucial role in monitoring and responding to outbreaks. However, these services are largely inaccessible to people worldwide due to their high costs and resource-intensive processes, as they often rely on expensive data sources. Fortunately, satellite images are a great alternative data source. Modern satellites can provide detailed images that display a region’s financial status and pollution levels, two critical metrics in potential disease outbreaks. Therefore, this study aimed to develop a more affordable algorithm (SatNet) that utilizes publicly available satellite imagery to perform disease hotspot detection. The algorithm works by retrieving zoomed in satellite images of the city inputted by the user and feeding these images into a novel, hybrid, recursive convolutional neural network. This model, designed to classify regions within the images as low-income, high-income, or industrial areas, was trained and tested on a custom data set of 7,448 images and achieved a 94.872 training accuracy and 84.183 testing accuracy. The output of this model is then used to create a detailed heat map for the city, which indicates the regions most in danger of disease outbreaks. Overall, the affordability and accessibility of SatNet will allow governments/organizations worldwide to provide their people with the healthcare they need and significantly reduce the spread of diseases in an increasingly interconnected world.