Leveraging Deep Learning models: To identify healthcare facilities and predict the location of new healthcare infrastructure from satellite imagery.

Prof. Manish Gangwar, Prof. Madhu Viswanath, Dr. Shruti Mantri, Gokul Kumar

 

Satellite Imagery can be used to extract information and provide solutions in diverse areas such as remote sensing for observing human rights movements, tracking human trafficking, providing early signs of famine, observing population growth in cities, spotting undeclared nuclear plants, retrieving intelligence by the military on enemies, etc.2

 

Today, fast-paced technological innovation, economic uncertainty, and geopolitical volatility establish the world order. In such an environment, harnessing the potential of location data can help governments make better policy decisions that lead to accelerated economic growth. Geospatial data and remote sensing technology recently have played a critical role in aiding governments with pandemic preparedness and response. Some areas where geospatial technologies helped authorities realize pandemic response was personnel tracking, grid management, spatial Big Data analysis, rapid visualization, prediction of regional spreads, and even panic elimination.3

 

With the onset of the Pandemic, the Maharashtra government had to grapple with multiple problems- insufficient existing healthcare facilities and identifying newer structures suitable to be converted to makeshift healthcare centers. The government also had to coordinate the urban and rural health services to ease social contradictions in healthcare access.

   

In the past, in Maharashtra, public hospitals have offered minimal services. Left with little choice, people have had to go to private hospitals to avail more comprehensive medical facilities. Private hospitals and corporatized hospital chains have filled the void left by the fund-starved public health care system.

 

It was critical that the government select the right site for the makeshift hospital during the Pandemic. Appropriate hospital site selection would help optimize the allocation of medical resources. It would also offer improved access to healthcare and reduce the time to reach the hospital during a medical emergency by the citizens. All this would help curtail the spread of disease and enhance the quality of life of the people. Optimum hospital site selection would also be cost-saving on capital strategy.

 

Realizing the need of the hour, the researchers in this study designed and developed Deep Learning models to identify healthcare facilities from satellite images. The model also helped identify sites and existing infrastructure suitable for makeshift hospitals and healthcare facilities. 

 

Researchers collected data from hospital locations from Municipal Corporation websites and the latitude and longitude coordinates from Google. They adopted Computer Vision Annotation Tool to annotate images to train the Deep Learning Models. The researchers obtained satellite images of 60 hospitals from 43 cities in Maharashtra from Mapbox and obtained data in tiles at a resolution of 0.5m and 512*512 pixels for detection. The study considered Medical Colleges, Government hospitals, and Specialty & Multi-Specialty hospitals from rural towns and cities.

 

The study faced its own set of challenges. The researchers found it difficult to identify unique vector parameters such as external roads, internal roads, open space in and around the hospital premises, adjacency, and incidence matrix of the road infrastructure from the satellite images. There was also data imbalance in the aerial imageries and non-uniformity in the hospital structure in the images. Due to unclear images, the researchers found it difficult to draw bounding boxes and label objects of interest. 

 

The researchers observed that the building infrastructure of healthcare facilities in rural Maharashtra was different from that in urban Maharashtra; hence there was a need to develop two separate models. They classified images as Urban and Rural based on the location and geographic demographics. The study applied the UNET Model architecture to train and predict hospital images, sites, and facilities for makeshift healthcare infrastructure during the Pandemic. The Deep Learning model developed by the researchers has been incorporated in Geo-Qi Mapper (developed by Geospoc), identifying rural and urban hospitals from satellite images. 

 

The model that was developed facilitated both the prediction of existing healthcare buildings from satellite images and the selection of new sites for makeshift healthcare facilities. The government adopted the model outcomes that expedited its decision-making in constructing a new makeshift healthcare facility during the Pandemic. The citizens also were able to identify the nearest hospital/healthcare facility in their vicinity. 

 

The algorithm generated had a few over predictions due to open-source data sets. Adopting spatial images of higher resolution images will drastically reduce over forecasts. 

 

Future Collaborations

Population Surveillance, Case Identification, Contact Tracing. Public Health in the future will get increasingly digitized. Nations will need to get aligned for future preparedness for infectious diseases.

 

IIDS aims to partner with non-government and government organizations to utilize our data to establish hospitals and healthcare facilities for potential sites. We would also like to collaborate with research institutes to identify other features remotely. Adding more layers to the existing study to perform a detailed analysis of infrastructures available on the ground will help.

 

Acknowledgment

The current study is a part of the UNICEF Innovation Fund. Geospoc sponsored the problem statement to the ISB Institute of Data Science, Indian School of Business. We are thankful to Geospoc for allowing us to solve a complex business problem during pandemic times.

 

Next Set of Goals

Our objective at The ISB Institute of Data Science is to find innovative solutions to complex problems. In the last year, we have worked on a challenging problem of hospital detection through open-source satellite imagery. With this, we aim to formulate innovative solutions to cutting-edge problems in the future. We will continuously seek to collaborate with Government organizations, industry, and academia, and intend to extend the project to other states and countries. 

 

 

References:

1: Gangwar M., Viswanath M., Dr. Mantri S., Kumar G. ‘Leveraging Deep Learning models: To identify healthcare facilities and predict the location of new healthcare infrastructure from satellite imagery,’ 2021

2: https://www.geospatialworld.net/blogs/15-mind-blowing-works-satellite-imagery-can-used/7/

3: https://www.geospatialworld.net/article/how-geospatial-data-and-technologies-can-help-in-disease-prevention-and-control/