Flare Image Feature Extraction: An AI-Powered Approach to Independent Flare Surveillance and Reporting

Self-reporting of gas flare volume by operators to regulators has become a standard industry practice. It is often the easiest way of enforcing compliance with environmental standards and regulatory measures. However, the method is predisposed to high inaccuracies and errors, intentionally or unintentionally introduced into the report by the operators for several reasons, ranging from the company's reputation to reduced payable fines and technical limitations of available flow devices. Reconstruction of flare volume from satellite footage is currently the only independent means of assessing flare information, but the method is contingent on satellite spectral bandwidth, seasonal climate, heat signature, and flare size. This research, therefore, seeks to improve flare data auditability and accuracy by leveraging AI techniques for independent assessment, reporting, and evaluation of gas flare volume from flare images in pictures and videos. It sought to replace the need for sophisticated (and expensive) hardware monitoring devices with cameras and simple digital storage devices for real-time estimation of flare rate from the flare image(s) in pictures/videos. 23,000 high-quality images of actual flares with their associated flow rate were used to develop the deep learning detection-guided flare surveillance system. The system combines a flare detection model, a detection-guided multimodal image feature extraction algorithm, and a deep learning regression model for flare volume estimation. Real-life flare images and associated flow rates taken and recorded from the Agbada 2 Flow Station operated by Heirs Energies Limited, Nigeria, were used for the model training, analysis, and development. The Flare image detection model, trained to guide the multimodal image feature extraction algorithm, achieved an excellent metric score across various model criteria with 99.42% precision and 97% recall, demonstrating its ability to detect flare images in different picture quality, size, and orientation. The comprehensive flare surveillance system, evaluated on six metric scales—mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), mean absolute percentage error (MAPE), standard deviation (SD), and R-squared (R2)—demonstrated exceptional performance with values at 0.085, 0.046, 0.214, 6.146%, 0.209, and 0.995, respectively. This indicates that the flare surveillance system is suitable for field deployment and real-time estimation of flare volume from flare images in pictures and videos. The developed solution offers a novel, simple, independent, and transparent method of assessing gas flare volume from flare images. It eliminates the need for operators to self-report flare volume to regulators, thereby enshrining flare data transparency and unbiased imposition of flare penalties.