How much cleaner is the air in east Asia now than fifteen years ago?

This past week, a scientific paper I led was published (after an especially long peer-review process) in the journal Atmospheric Environment (read here for free). We use satellite data to make daily high resolution maps of fine particulate matter pollution (PM_2.5) from 2011 through 2022. PM_2.5 is the leading environmental cause of death; fine particles can penetrate deep into the lungs, the blood, and even the brain and cause health problems in just about every system in the body. Over the study period, PM_2.5 in the region has improved dramatically thanks to powerful regulations, improving the lives and health of over one billion people. However, it is hard to tell how effective regulations have been because lots of PM_2.5 data is missing. Some places don’t start measuring PM_2.5 until after rules have changed, making it hard to evaluate impacts. Even today large areas of cities and especially rural areas don’t have measurements at all. One potential solution: satellites. Satellites cover areas where surface sensors are missing. More importantly, the satellite I use was launched before there were big efforts to measure air pollution at the surface in the region. In this study, we create a big dataset that uses satellites to fill in the gaps of surface measurements of PM_2.5, and we released it for free for anyone to download and use.

The idea for this study came from some colleagues at Yonsei University in South Korea. Back in 2010, the Korean Space Agency launched a satellite called GOCI to study algae blooms in the ocean near the Korean peninsula. It is geostationary, meaning it sits over the same place on Earth, and therefore can take multiple observations each day. Jhoon Kim, a professor at Yonsei, and his group wanted to use that satellite to measure the atmosphere instead. They did some clever tricks to take advantage of the different wavelengths of light measured by GOCI and developed an algorithm to measure Aerosol Optical Depth (AOD). AOD measures how light is blocked by particles in the atmosphere. AOD correlates with surface PM_2.5 (the more pollution, the more light is blocked), but the relationship is complicated. AOD is measured from space and can’t tell if pollution is at the surface, where it affects human health, or in a high plume far above the cities.

I worked with the Yonsei team and other collaborators to develop a machine learning algorithm linking AOD to surface PM_2.5 measurements. The figure below summarizes the workflow of the project. AOD measurements are in the top row, while surface measurements of PM_2.5 are in the middle row. Our machine learning algorithm is trained on the relationship between these two rows, accounting for lots of other information like meteorology, and infers the dataset on the bottom row at daily resolution. This new dataset fills in all the dots that are missing from the PM_2.5 maps on the middle row.

To give you a sense of what this dataset adds, here is a map of Seoul during an extreme pollution event in May 2016. Many areas outside the city center have no nearby observations (shown in dots), but this tool can help fill the gaps.

Along the way, we also developed tools to estimate AOD when the satellite fails (it doesn’t work on cloudy days or on snow-covered surfaces), and to extrapolate our predictions to periods with far less PM_2.5 data or where data is missing entirely — without biasing our results. One neat thing we tried: South Korea only began measuring PM_2.5 in 2015. This is tricky with machine learning, because missing data can lead to big biases. However, the environmental ministry did measure other pollutants (including coarser PM₁₀) which correlate closely with PM_2.5. We trained a separate machine learning algorithm to create a synthetic training set, covering this pre-2015 period so that we can create our full 2011-2022 maps including South Korea. Our synthetic PM_2.5 compares very well to a small network of observations in Seoul, which we don't use in training, so we’re reasonably confident this method doesn’t introduce bias.

The key takeaway: the cleanup of air quality in east Asia is one of the great environmental victories of our time. We find air quality has improved everywhere, with no population left behind. More work is necessary to achieve ambitious air quality standards and health goals, but the region is on the right track. Some much needed good news for your inbox today.