NOx Detection in Video Data Based on Ensemble Machine Learning

Mining is crucial for advancing technology and science by providing essential resources and materials. Quarry blasting, a necessary process in mining, releases these resources and chemicals, driving research and engineering improvements. However, NOx (nitrogen oxides) is an undesired contaminant produced during quarry blasts, posing a significant negative impact on mining activities. Therefore, detecting and quantifying NOx emissions is essential for designing effective blasts, analyzing their dynamics, and reducing contaminant production. Image processing and machine learning, including deep learning methods, have been applied to solve smoke plume detection and segmentation tasks. Most research has focused on smoke from campfires or forest fires, which are important for early fire and smoke detection. However, there is limited research on quarry blast smoke using machine learning methods, which presents challenges for recognizing the smoke plume from complex background scenarios and objects, including collapsed rocks, scattered soil dust and dirt, and the dynamic movement of the scene due to blast vibrations. Detecting NOx within the smoke plumes is further challenging because NOx must be distinguished from the overall smoke plume, as they share many common spatial features. To resolve this issue, we propose an ensemble method for NOx detection that combines smoke plume segmentation with NOx fume detection. First, we apply a smoke plume segmentation model to identify the smoke plume regions. Then, we use these masked-out smoke plume regions to detect NOx based on its unique color features. The first stage of the ensemble method helps block out complicated confusers in the background, such as reddish and brownish dirt. Next, we specifically design a color classifier for NOx to further improve its sensitivity. We evaluate the overall video-wise performance using accuracy, false alarm, and miss rates based on sampled image frames from videos. Additionally, we assess the mean Intersection over Union (mIoU) for smoke plume segmentation and the mean Average Precision (mAP) for image-wise performance from an algorithmic perspective. In summary, this paper represents the first attempt to apply machine learning approaches to NOx detection in a memory-efficient and time-efficient manner.