Flare Smoke Identification and Classification 


Flare stacks play a critical role in the controlled combustion of hydrocarbons at oil and gas facilities for safety and emergency depressurization and to help reduce environmental effects of atmospheric emissions. A variety of factors including ambient conditions, variation in gas composition, flow rates, exit velocities, flare tip design and oxygen content can lead to the generation of smoke as a combustion byproduct. While smoking does not necessarily suggest inadequate hydrocarbon destruction, it may be indicative of increased particulate matter emissions, and smoking flares can have an effect on visual amenity. Traditional methods for identification and classification of smoke have involved employing certified human observers. Reliance on the human eye and judgement to detect and grade smoke opacity is inherently qualitative in nature and presents additional cost and logistical constraints compared to automated continuous monitoring. Here, we propose a method that leverages on-site cameras to identify flare smoke and classify the opacity following the standards of Ringelmann number estimation as per US EPA Method 9. Specifically, we introduce a novel deep learning solution that is able to overcome some of the challenges involved in detecting and classifying smoke including its amorphous shape and ever-changing textures and shades by using a deep learning architecture to identify temporal and spatial features of the object prior to Ringelmann number classification. This solution has the potential to assist in optimizing flare performance, and to help manage environmental risk outcomes for a wide range of smoky atmospheric emission types.


Our objectives is to develop an AI system that can predict the Ringelmann number of smoking flares from CCTV videos.

Project Plan:

  1. We collected a unique dataset of flare activity for training and testing a model.
  2. We generated a range of synthetic scenes which we use to augment our training data, increasing the robustness of our model to uncontrollable factors such as weather conditions.
  3. We developed a novel deep learning model with few parameters for real-time Ringelmann number estimation.


The project started in July 2022 and is expected to wrap up by December 2023.


We developed a novel, lightweight model for automating the Ringelmann number estimation at oil and gas sites with an accuracy of 95.27% on unseen test set. 

Funding Provided by:

Chevron Corporation


CSU Computer Science Department