Air quality from space
In this blog post we will examine atmospheric data captured from space using satellites! In the previous blog post we used ground sensors to predict air quality, where now we will replace these sensors with satellite data. Lastly, we will examine how well the satellite data predicts air quality at the surface, potentially making the ground sensors redundant.
The map below shows an image captured by the Sentinel-5 Precursor (S5P) satellite as part of the European Space Agency. This satellite passes over the same location every day almost everywhere on Earth. It registers several air quality components, including nitrogen dioxide (NO2).
The black circles on the map are air quality ground sensors. Interesting to note is that the ground sensors are geographically sparse whereas the satellite data fills in all empty space. It means that the satellite has a higher geographic coverage. This is a big advantage over ground sensors as now we do not have to guess the content. The satellite data is also quite detailed, as you can see by the individual pixels varying from blue (low NO2) to red (high NO2). The smaller the pixels, the larger the geographic resolution. There is a tradeoff however, as seen in the graph below. You can see that over time the raw measurements of the satellite data (dark gray dots) are fewer and further between. This indicates that its resolution over time is lower as compared to the ground sensor measurements (light gray dots).
In the previous blog posts we talked about preprocessing the data, where smoothing was one aspect. You can see this by the lines running through the raw measurements. Using this smoothing we can also get a clearer idea of the behaviour and trends of the data. There are some clear differences, where the most obvious one is the lower volatility in the satellite data. One explanation for this is that the satellite sensor measures the whole troposphere over a large region (~7.5 x 3.5 km), therefore averaging out the NO2 concentrations. It also measures usually once a day, ignoring hourly fluctuations. Compare this to ground sensors which measure a single point close to the surface on an hourly basis, therefore being more subjected to local fluctuations. The measurement techniques also differ, where the satellite remote sensors are electromagnetic and the ground sensors usually chemical. Still, there seems to be somewhat of a similar trend between the data, making this satellite a good candidate to help with predicting air quality near the surface.
Just the Sentinel-5 data is not enough though, as we also need weather data to predict air quality near the surface. Currently, the weather data comes from weather stations on the ground. But just like the air quality ground sensors, they are sparsely located. Luckily there are assimilated datasets that combine ground measurements, satellite imagery and forecasting models to generate a widely covering grid of pixels with weather data. These can be seen in the map at the beginning of this blog post. You can open the layer menu at the top right and select several layers of weather data. In the graph below you can see their time series (dark blue), and how they relate to weather stations (light blue).
A clear resemblance between the ground and satellite / forecasting data is apparent when overlapping them in the time series. This is statistically supported because most of them have an r of 0.75 or higher. Now that we can replace the weather stations, let us experiment how well the satellite and forecasting data is able to predict air quality near the surface. The graph below shows the original predictions using purely the ground sensors (dark line) and the predictions where the weather station data is replaced by satellite / forecasting data (dashed line).
Air quality predictions ground versus space
Prediction performance statistics
Using weather stations: r=0.940, r²=0.883, mse=0.079, rmse=0.281, rpd=2.590
Using satellites / forecasting: r=0.904, r²=0.818, mse=0.116, rmse=0.341, rpd=2.133
As you can see the prediction performance is very similar! This means that we can replace the weather stations with satellite / forecasting data if we so desire. You can think of many cases where this is useful. For example, you could predict air quality at locations far from any weather station. Or you can train the prediction model using the distance and direction to existing air quality sensors, where with this information you could generate 'virtual air quality sensors' at any location in the Netherlands.
Covid-19 and air-quality
The predictions of course need to be tested with scientific rigour on more locations at different time intervals. But before we get there, a contemporary and practical question still remains unanswered. This concerns the effects of the new coronavirus (SARS-CoV-2) on our air quality. Not the virus itself of course, but the measures enforced by the government to prevent its spread. It has little to do with predicting air quality, but it is interesting nevertheless. Personally I did notice the air being cleaner in Amsterdam while cycling around. Is this coincidence or actually supported by the numbers and trends in the ground and satellite data? We will examine this in the next blog post.