Deep Convolutional Neural Network for Plume Rise Measurements in Industrial Environments

The estimation of plume cloud height is essential for air-quality transport models, local environmental assessment cases, and global climate models. When pollutants are released by a smokestack, plume rise is the constant height at which the plume cloud is carried downwind as its momentum dissipates and the temperatures of the plume cloud and the ambient equalize. Although different parameterizations and equations are used in most air quality models to predict plume rise, verification of these parameterizations has been limited in the past three decades. Beyond validation, there is also value in real-time measurement of plume rise to improve the accuracy of air quality forecasting. In this paper, we propose a low-cost measurement technology that can monitor smokestack plumes and make long-term, real-time measurements of plume rise, improving predictability. To do this, a two-stage method is developed based on deep convolutional neural networks. In the first stage, an improved Mask R-CNN is applied to detect the plume cloud borders and distinguish the plume from its background and other objects. This proposed model is called Deep Plume Rise Net (DPRNet). In the second stage, a geometric transformation phase is applied through the wind direction information from a nearby monitoring station to obtain real-life measurements of different parameters. Finally, the plume cloud boundaries are obtained to calculate the plume rise. Various images with different atmospheric conditions, including day, night, cloudy, and foggy, have been selected for DPRNet training algorithm. Obtained results show the proposed method outperforms widely-used networks in plume cloud border detection and recognition.