A plume is a cloud of smoke or vapor emitted from buildings during the production of energy, through the burning of heating oils. They are typically made up of pollutants such as PM 2.5, SO2, and CO2.
NYC has ~10,000 buildings that are still burning heating oils.
75% of greenhouse gas emissions are from buildings, according to the City of New York.
If we want to be able to reduce emissions, we need to understand the emission patterns of buildings.
We trained Faster R-CNN, an image classification and localization model, to detect plumes automatically using RGB images taken from a building-top camera. It provides a cost-efficient way of measuring the frequency of building emissions without the need of deploying an expansive in-situ sensor network to detect plumes using local air quality measurements.
Yes! We gathered a dataset of approximately 14,000 plumes total. We trained the model on around 6,000 examples of plumes using background subtracted images which highlight areas of the image that are changing in color or intensity and remove areas that remain constant. By doing so, we achieved a mAP of 62%. You can see some examples of detections here.
You could probably tell that the labels weren't always consistently showing up in every frame. This is because each detection is performed without any knowledge of the detections on the images surrounding it. Our next step would be to train the model to use the temporal context to improve predictions and generate tracking information automatically.
We also plan to run the model over a long period of time to create a census of all plumes emitted in the observed view. The plumes can then be mapped to their physical geographic coordinates in 3D space using a technique called photogrammetry, allowing us to know where, precisely, in the city the plume occurred.
We would like to thank Dr. Federica Bianco and Dr. Greg Dobler for being fantastic advisors and project sponsors, along with the Urban Observatory for providing us with the images, the computing resources, and their support. We’d also like to thank Anupama Santhosh for her work in the first half of the project and are sorry that she was unable to complete the project with us. And of course, Mohit Sharma for his tireless work in all things CUSP.
