According to the World Health Organization, one in eight deaths worldwide are caused by air pollution and nine in ten people do not breathe safe air.[1] Fine particulate matter (PM 2.5) air pollution is small enough enter the lungs and diffuse into the blood stream, causing strokes, heart disease, lung diseases, and other health conditions.

The parts of the world with the most polluted air have the fewest quality monitors.[2] Air quality monitors can measure air pollution with great precision and accuracy, but are expensive to purchase and maintain. Low-cost air quality sensors are much cheaper and are beginning to be deployed around the world, but are not as accurate as monitors.[3] Even with low-cost sensors, large air quality data gaps still exist, especially in many developing countries. Satellites can measure air quality where humans can not, covering the entire world to help fill in these gaps, but satellites measures aerosol optical depth (AOD), not PM2.5 directly;[4] this means satellite data has to be calibrated with on-the-ground air quality measurements, meteorological, and geographic data. We have developed a method that combines monitor, sensor, weather, and satellite data to improve air quality data coverage in the developing world. These PM2.5 estimates can then translated into human health consequences and negative economic impact to inspire action to reduce air pollution.

We propose a two-stage model to incorporate all of these data.We would like to identify a simple conversion factor that can directly translate satellite measurements into PM2.5 estimates globally.To identify this conversion factor, we propose combining satellite, meteorological, and geographic data and fitting it to available high-quality air quality monitor data with a linear model. We can use the machine learning technique of L1 regression (Lasso) to remove some of the highly-correlated parameters to ensure a better fit. This global model will allow the calculation of rough PM2.5 values anywhere in the world from satellite data.

To incorporate local sensor data, we propose rescaling our global model with available local sensor data to create more accurate, location-specific models.A scaling factor will be calculated by linear regression between our global model PM2.5 estimates and available site-specific low-cost air quality sensor data. This fit will help mitigate error that arises from poorer quality low-cost sensor data. Incorporating local sensor data by calculating a location-specific rescaling factor at the city level will give us more accurate air quality estimates for that location.

Our example map of air quality in Addis Ababa, Ethiopia, provides a rough estimate of PM2.5 from the satellite detected AOD. These data could be further calibrated with the available low-cost sensor data for Addis Ababa. These local PM2.5 estimates can be used to identify the health and economic impacts[5] to inspire action to improve air quality.

We have created a powerful conceptual framework to provide actionable information in locations that currently have little air quality data.

[1] WHO, 2019, https://www.who.int/airpollution/en/

[2] WHO, 2019, https://www.who.int/airpollution/en/

[3] EPA, 2018, https://www.epa.gov/sciencematters/epas-air-sensor-toolbox-offers-new-tools-community-led-air-monitoring

[4] NASA, 2016, https://www.nasa.gov/feature/langley/the-future-of-monitoring-air-quality-from-space/

[5]EPA,2019 https://www.epa.gov/benmap