Project: Tracking Air Pollutants and Reconstructing 3D Scalar Fields from 2D Satellite Images via Machine Learning

Team members: Marco Giometto, Civil Engineering and Engineering Mechanics, Columbia Engineering; Pierre Gentine and Mostafa Momen, Earth and Environmental Engineering, Columbia Engineering; Carl Vondrick, Department of Computer Science, Columbia Engineering.

This team is developing machine-learning models and improved satellite-imaging techniques that will help environmental officials locate and characterize hazardous pollutants in the lower atmosphere, allowing them to design strategies to mitigate pollution.

The models will use machine learning to track how pollution plumes are transported by atmospheric turbulence, which controls the dispersion of contaminants in the lower atmosphere.

Satellite images are inherently two-dimensional and do not provide information about the three-dimensional structure of pollution plumes. The team will use data-science techniques to reconstruct three-dimensional pollutant plumes from corresponding 2-D images. No one to date has automatically tracked the dispersion of pollutants in the atmosphere from 2-D images. And the idea of creating 3-D structures of pollutants from 2-D satellite images hasn’t been tried previously. To do this, the team will use high-fidelity numerical simulations coupled with deep neural networks. The simulations will generate a database of pollutant plumes, which will be used to train the deep neural networks to reconstruct 3-D plumes from the 2-D images.

There have been previous studies on reconstructing 3-D pollution plumes from 2-D images, but they have relied on subjective physical models, says Marco Giometto, an Assistant Professor of Civil Engineering and Engineering Mechanics and a member of the Data Science Institute.

“Objectivity is a necessary condition for the reliability of any coherent structure-detection method, yet no objective method is now available,” Giometto says. “So the combination of machine learning tools, imaging techniques and turbulent flow research has great potential to yield significant insight into pollution.”

The team hopes to provide improved images and models of pollutant concentrations, which will allow environmental scientists to better monitor the flow of pollutants in the lower atmosphere. Such monitoring will allow them to understand the levels of pollution in different regions and the dangers those concentrations represent to the people who live and breathe in the air.

“I’ve heard it said that air pollution is the new tobacco, and that the simple act of breathing is killing seven million people a year and harming billions more,” says Giometto. “I hope our research leads to a better understanding of air pollution and its hazards and what can be done to reduce those hazards so all of us can breathe cleaner and healthier air.”

— Robert Florida