The idea is to find correlations between anthropic and natural features in common between all the states of the world, in order to find the main factors that could impact on the COVID-19 spread. Moreover, our project provides predictions on how changing the main factors could impact on the pandemic: thus, we integrate various EO-derived features with available and/or derived socio-economic data in various ways in order to discover and enhance our understanding of COVID-19 impacts.
Our country ended the lockdown recently, but we noticed that governments did not know which factors could be determinant to lower the infection cases. Joining multisource and multi-temporal data, we developed a solution to help countries, still fighting against COVID-19 spread, in predicting which actions can have a significant impact in defeating the pandemic and face future diseases outbreaks.
Taking advantage of our experience in the space/environment/data science field, we accurately searched and screened the most significant satellite data and the most updated socio-economic ones that could be correlated with the COVID spread. We then selected both natural and anthropogenic variables in order to create a dataset that, combined with socio-economic data, could feed an AI-model, able to extract COVID patterns exploiting several all those data. The neural network is able to generalize changing the value of the most important features, showing new scenarios based on these variations. All this information have been collected and visualized on a map in an innovative Web-App, where you can experiment some parameter combination, e,g population density, number of physicians, pollution etc. to see which effect they could have.
Data used are: Meteorological data (Temperature, Precipitation, Wind, Humidity), Land Use (e.g Permanent Crops, Land Area, Forest Coverage, Urban Area) and Pollution data (NO2, OZONE, PM2.5, CO2) from NASA satellite as OMI, Population data (e.g Global population density, life expectancy, physicians rates, the annual rate of population increase), Socio-economic data (e.g GDP per capita, Balance Import Export, Index of industrial production, Employment rate, Light night intensity rate) from many sources as JRC GHSL layer, United Nation and World Bank Datasets.
Together they generated more that 130 features to train our very precise model!
We used NASA satellite data to add fundamental features to our dataset, which are necessary to reach high prediction accuracy in the AI model. In fact, they provide worldwide coverage needed to obtain a complete overview of the situation.
We faced some problems related to the collection, standardizazion and fusing of data coming from multiple EO and non-EO sources. We also dealt with the problem of choosing the most appropriate variables that could be correlated with the epidemic impact.
Since for some nations we had some null values, we have implemented an algorithm to fill those empty values with high accuracy, reconstructing their value from nations with similar characteristics.
The model having with target the spread of the virus on a logarithmic scale reaches an accuracy of about 1.2 on a root mean square error metric.
The network is composed of 2 hidden layers, composed of 200 and 100 neurons, to better extract all the requested patterns and information.
Project Presentation:
https://lab.egeos-services.it/gitlab/gbottini/spaceapp-nasa/blob/master/BANDERSNATCH-19.pptx
Project App Link:
https://spaceapp.max-ics.lu/bandersnatch/
Youtube Video Link:
https://www.youtube.com/watch?v=jyBxENl5mRY&feature=youtu.be
For socio economic data:
For space data: