Algorithm for Air Pollution Hot Spots

In 2020, a team of researchers from Duke University, North Carolina (United States) published a novel method to measure localized air pollution levels using micro-satellite imagery and machine learning. In this study, the team focused on measuring PM 2.5: particulate matter produced through combustion that is finer than 2.5 micrometres and contributes to serious health risks such as respiratory illnesses. Until now, it has been challenging to measure PM 2.5 levels precisely and accurately due to the limited coverage of air quality sensors in urban areas, the large scale of available satellite imagery, and the limited accuracy of existing satellite imagery analysis. In their article, Zheng et al. outline a method that uses images from Planet’s microsatellites at a scale of 3m per pixel and applies a machine learning algorithm to detect the level of pollution. The model combines a random forest algorithm (made up of many decision trees based on local weather data), and convolutional neural networks that determine the level of blurriness in an image. Combined, this information enables the model to generate a measurement of PM 2.5 at a scale of 40,000 square metres (0.04 square kilometres) at a high level of accuracy. This method has the potential to enable more consistent and accurate local air pollution monitoring, allowing for more effective mitigation and prevention of air pollution impacts on public health.

Zheng, T., Bergin, M. H., Hu, S., Miller, J., & Carlson, D. E. (2020). Estimating ground-level PM2.5 using micro-satellite images by a convolutional neural network and random forest approach.


Artificial Intelligence, Climate Change, Data, Lifestyle, Monitoring, Pollution