Methodology

The approach taken for detection of polluting plumes in New York City was as follows:
  1. Background subtraction - a differencing technique is applied to the images to remove stationary features which are common between images, and highlight differences between static and moving objects, including people, vehicles, clouds, shadows, and of course pollution plumes . 
  2. Image labelling for compilation of training set - using visual inspection of differenced images paired with statistical heuristics to identify images containing plumes along with other sources of noise and perturbance which a model may pick up as signal.
  3. Model development - training and testing of Faster-RCNN model to develop a plume detection algorithm for use across the larger set of images (those not previously included in the training or testing set).
  4. Plume detection census - run the model across the complete set of images to compile a census of all plumes along with statistics such as source, frequency, and type of plume emitted by buildings in the images.

Results

The Faster R-CNN model was trained on a down-sampled set of all tagged annotations (Table \ref{547161}).