Awards & Nominations
PAGROS has received the following awards and nominations. Way to go!
Where There’s a Link, There’s a Way
COVIDCompare
Summary
COVIDCompare is heatmap comparator to find correlations between COVID-19 cases and other environmental data sources
How We Addressed This Challenge
There are two key problems poised in this challenge:
No Central COVID-19 Environmental Links Hub
While there is a lot of data on individual environmental factors and their relation to the spread of COVID-19, there is currently no dedicated hub to consolidate these possible correlations. At the end of the day, there is no cross-collaboration between research of different environmental links, slowing down the chances of discovering causal links. This is where COVIDCompare steps in.
Our project makes it simple to visualise correlations between environmental factors with COVID-19 cases through an easy to use interface. With COVIDCompare, anyone can upload data sources, get them automatically converted into heatmaps and easily see possible causal relationships between human health and the environment. By plugging in multiple data sources, a more comprehensive picture of the environmental factors that contribute to the spread of COVID-19 can be seen.
How is this useful?
COVIDCompare serves as a useful resource for the Earth science and health science communities alike. Acting as a one stop hub for all environmental links, more collaboration between research can be achieved. More importantly, our project lends a hand towards the predicting the risk of new COVID-19 clusters as its built to harness the combined datasets of NASA, ESA, JAXA, CNES and CSA. With enough data factors and correlations, dynamic neural network models could be created to map and predict future clusters before they occur.
How We Developed This Project
Motivation
In recent times, government responses to COVID-19 have been largely reactionary in nature. While it is understandable given the suddenness of the virus, our team felt that to really mitigate and control its spread, more proactive measures have to be created. The lack of work on predictive models for the virus prompted our team to start on COVIDCompare, paving the way to more international collaboration. This will open up the possibility of reaching the eventual goal - being able to predict future COVID-19 clusters before they occur.
Our Approach
The process for uploading and normalising environmental data sources into heatmaps for comparison is a difficult and time consuming task.
When we discuss heatmaps, we are handling spatial information. Often, the information given to us is in the form of a list e.g. GDP by county, healthcare ratings by state. This can be hard to convert to visual information. While looking for a tool to project these lists into heatmap, we thought: why not build it?
Map software requires us tedious encoding of country boundaries, polygon filling, and hotspot generation. But once we got that out of the way, we had a set of tools that could convert all these lists into a spatial representation. We did all this in the first day.
While playing with the NASA data visualisation website, we were inspired by the idea of heatmaps. Epidemiologist John Snow famously studied cholera outbreaks with heatmaps. Why not we do the same for COVID?
Of course there are different things to consider. Heatmaps are essentially images and so we are trying to figure out how similar these images are. We did a literature review where such approaches are considered. For image generative adversarial networks (GANs), images are often compared for structural similarity. This is quite close to what we are doing when we compare two heatmaps based on the types of slopes and features. However, most GANs are often used to check if two images look similar given a surrounding context. In heatmap comparison, it would benefit if we are able to have quantities that account for how absolutely similar two images are.
We turned to another area in the field of super resolution and found many promising approaches. We decided to use two of them - structural similarity and peak signal noise ratio. These two measures quantifies the closeness of two normalised images (on a 'pixel'' to 'pixel' basis) in a single value. We then visualised this closeness on yet another heatmap.
Additionally, we noticed earlier on that the raw values used for making heatmaps results in very high contrast which is bad for comparison. Hence, we also turned our focus to preprocessing the heatmaps before sending through the correlation function. We played with a variety of methods to filter out heatmaps. Ideally to make meaningful comparison, images should be preprocessed so that the measures work as they intended. There are many factors and methods to accomplish this. Some of the methods we considered where histogram flattening, applying a log map to the values and various other filters, such as Gaussian blur.
Ultimately, our final toolset will come with these preprocessing filters and measures that would make heatmap comparison a brisk task for anyone who wishes to dive in an initial analysis of correlations between different datasets.
Achievements
We managed to create a working tool that speeds up the time it takes to compare datasets to COVID-19 cases (or other data sets) by a large factor. What would take days of work can now be completed in less than an hour, enabling researchers to find non-obvious correlations between data quicker.
Project Demo
Visit our website to see our work!
https://covidcompare.wixsite.com/home
Project Code
Data & Resources
Software:
Python for HD5/ CSV/ JSON file conversion to standardized TIFF format
QGIS for visualising TIFF files and creating demo heatmap images in website
PowerBI for interactive maps used in demo video
Data Sources:
GDAM for shapefiles used to create US county/state/country borders
NASA Earth Observations (https://neo.sci.gsfc.nasa.gov/) for the following datasets:
Aerosol Particle Radius - Apr 2020
Temperature (day) - Apr 2020
Temperature (night) - Apr 2020
Vegetation Index (NDVI) - Apr 2020
Cloud Fraction (NDVI) Apr 2020
Night Light - 2016 (https://earthobservatory.nasa.gov/features/NightLights)
John Hopkins COVID Dashboard (https://coronavirus.jhu.edu/us-map) for COVID data by state.
County Health Rankings (https://www.countyhealthrankings.org/) for the following dataset:
2020 County Health Rankings Data - v1 containing a wealth of socio-economic & health data.
United States Census Bureau (https://www2.census.gov/programs-surveys/popest/datasets/2010-2019/counties/totals/) for the following dataset:
US Census Data 2019
*The full list of data used is presented in a table in our Github page with relevant description of the variables and their sources.