Model predicts 63% of within-city dengue variance
With ever-evolving technologies, there is growing amount of data to be interpreted. Of this data, satellite images are already being used to predict climate change activities and are now being utilized in the field of Public Health.
Marcia Castro, PhD, Professor of Demography with the Harvard TH Chan School of Public Health, began utilizing satellite imaging while working in Fortaleza, Ceará, Brazil where she had access to geographic data indicating instances of Dengue and Chikungunya. Dr. Castro decided to investigate if the urban landscape contributed to the disease transmission. Like many cities with vector control teams, the city already had several high-risk transmission areas identified including places with discarded tires and scrap yards but identifying a comprehensive list of high-risk areas is time-consuming work and likely improbable.
Dr. Castro’s team acquired a series of high-resolution images of the area through a grant from the DigitalGlobe Foundation. Their goal was to find a way to identify neighborhoods that are more susceptible to breeding habitats using characteristics found in these images. The images would be used to identify informal settlements to identify higher-risk areas to show where the efforts should be focused to combat active disease transmission.
Jonathan Jay, DrPH JD, a Postdoctoral Fellow with the Harvard TH Chan School of Public Health in Castro’s lab, liked the idea of predicting potential risk areas from satellite imaging and began working with Dr. Castro. They worked to expand the project to use the dataset of Dengue infections to see if the elements in the city corresponded with the high transmission areas.
Of the satellite images, Dr. Jay says, the images are at a high enough resolution where one can see what is going on at ground-level and “you can tell if you are looking at a well-organized, higher income neighborhood or a lower income neighborhood with more informal settlements or a commercial downtown area.”
Due to the vast number of satellite images of the Fortaleza city area, the images are put into a pre-trained machine learning model called a Convolutional Neural Network that is also used for facial recognition software. The powerful algorithms within this model are trained on over a million images that were labeled by hand.
The algorithm can identify different shapes, colors and textures within an image to recognize features of each photo. Dr. Jay’s team assigned a numerical value to various characteristics for the model to sort including whether or not an image contained hard roofing edges.
In Fortaleza, the spacing between homes varies drastically between strategically-planned neighborhoods and those that are more informal settlements where it is hard to discern where one roof begins and another ends. As seen in the images below, if an algorithm noted sharp roofing edges or homes that are spaced further apart, that would indicate strategic urban planning which would ultimately correlate to the area being a lower risk (risk quintile 1-2) for Dengue transmission.
Once the data is gathered, they are then put into a predictive model to study the probability for an active Dengue transmission based on the variety of risk factors. During a presentation at the 2018 American Public Health Association (APHA) Annual Conference, Dr. Jay shared that when the results of the trained model were paired against actual transmission data using cases in 2011-12 to see how well it could predict 2013 cases, it did quite well, explaining 63% of the variances in cases.
Dr. Jay notes that because the images from Fortaleza are such high resolution, they can help train the algorithms to detect risk factors that can then be detected in images that may not have the same level of resolution.
Satellite Imaging to Predict Gun Violence and Overdoses
While mosquito-borne transmission prediction has the most obvious link to environmental risk factors with mosquitoes clearly preferring certain breeding habitats to others, this technology can be used to create risk maps for a variety of urban health risks. Beyond the work with infectious disease and vector borne illnesses, Dr. Jay also used his prediction algorithms for other public health concerns including potential for gun violence and risk of overdoses. According to Dr. Jay, there is a growing body of evidence around how different neighborhood-level physical risk factors contribute to gun violence risk and how it varies from block to block.
According to studies on gun violence risk, there has been a correlation between unkempt vacant lots contribute to gun violence risk and how simply cleaning them up can reduce that risk. In his work within Philadelphia, the research seems to show that young people who spend more time among trees in Philadelphia are at a lower risk for gun violence. Because of this, there is good reason to believe one could train an algorithm to look for something like tree cover and vacant lots when determining risk for an area.
For Dr. Jay’s work with overdose prediction in Portland, Oregon, the risk factors have more to do with understanding whether or not people living in the area are overdosing in public places versus in their private homes and what patterns may be associated with that. The satellite images are most utilized in locating the liminal spaces between urban settings where overdoses have been recorded as satellite imaging is very accurate is determining urban connectedness. Due to the strong commitment of the local fire department to Public Health, a pilot version of the satellite-based risk estimates map interface is being utilized by the Portland Fire Department.
According to Dr. Jay and Dr. Castro, urban imagery is likely to emerge as a go-to public health data source. For public health officials, building the capacity to interpret satellite imaging data could help target preventions, surveillance and treatment efforts where they are potentially needed most.