1 Introduction
Smoke plumes contribute significantly to the air pollution posing serious health risks including heart disease, lung cancer and asthma. The city recently estimated that up to 2,700 premature deaths a year could be attributed to fine particulate matter and ozone in the air. In 2005, it was estimated that approximately 10,000 buildings in the city burned number 4 and 6 heating oils, which emit more air polluting PM 2.5, sulphur dioxide(SO2) and nickel than alternative fuels. In 2007, NYC launched a sustainability program, titled PlaNYC, which aims to bring significant emission reductions, with a goal of 30%, by 2030 [1]. According to [2], NYC buildings account for 80% of all of greenhouse gas emissions, meaning that in order to enact the necessary change, building energy usage needs to be addressed and understood. The wider implications of this study could impact many different city agencies and departments such as Energy, Environment, Health, Transport, Buildings and Housing.
The traditional methods for detection of plumes rely on point sources or more recently, sensor networks [16] where there is significant back and forth transfer of information making the process cumbersome and costly. Our objective is to make the spatio-temporal tagging of the plumes a real-time and viable process.
Many studies seek to detect the presence of smoke in an image as a means to detect fire [5-13]. These are designed under the motivation of replacing traditional smoke detectors due to their numerous advantages such as early fire detection, fast response, non-contact, absence of spatial limits, ability to provide live video that conveys fire progress information, and capability to provide forensic evidence for fire investigations [3]. Studies could not be found with the same problem statement as this paper, however the methods for detecting smoke in the context of house fires overlaps with that of detecting plumes.
Çelik et al [5] take a statistical approach, using color models to detect both regions with smoke and those with fire which are constructed using hand-engineered color features. Yuan [6] attempts to improve the false alarm rate of video-based smoke detection algorithms by incorporating the orientation of the smoke’s motion, helping remove the disturbance of other moving objects. Gubbi et al [7], Tung and Kim [11] use visual codebook style representations to detect the presence of smoke, employing support vector machines (SVMs) and random forest classifiers, respectively. Kurtek et al [12] attempt to detect smoke from a single image which is a non-trivial task. The image is assumed to be a linear combination of smoke and non-smoke components which are derived using atmospheric scattering models. CNNs have emerged as the state-of-the-art image classification algorithm due to its efficient architecture that takes advantage of the stationarity and locality of patterns found in images. They have also performed well on video classification tasks [14]. Frizzi et al [13] recognized that the existing literature was primarily rule-based models and hand-engineered features. Based on the CNNs previous success with image classification, a CNN was trained to predict fire and smoke from still images. Faster R-CNN is a specific form of CNN that includes a region proposal network which hypothesizes object locations via bounding boxes and performs better than its predecessors, R-CNN and Fast R-CNN [15].
This project aims to create a method for detecting and recording plumes of pollution in NYC using images gathered from UO cityscape images and to identify various statistics such as the building origin, the count of plumes, and the spectral characteristics. This will be performed by constructing a training dataset which will be used to train Faster R-CNN for plume location. 3D photogrammetry will then be used to identify the source building using a 3D model of the city.