The detailed model architecture was developed and tested on the PASCAL Visual Object Classes (VOC) dataset. \citet{Ren_2017} included existing image sets, development kit functions, classification functions, detection functions, segmentation functions and layout functions. The Faster R-CNN model will be used as the foundation for training the model for detecting the plumes, and weights pre-trained on the VOC 2007 dataset will be used to initialize the model. The dataset used was converted into the PASCAL VOC data annotation format for easier compatibility with the existing code base.
Plume detection census
An important piece of information to extract from instances of plumes is the building that the plume originated from. This is a non-trivial task because it requires reconstructing 3D information from using only a 2D image. Photogrammetry is a field of research that is used to undertake this, and an existing mapping from image pixel to the 3-dimensional location was used to geolocate the plume with respect to its building of origin.