Early Detection of Pipeline Natural Gas Leakage from Hyperspectral Imaging by Vegetation Indicators and Deep Neural Networks.

The timely detection of underground natural gas (NG) leaks in pipeline transmission systems presents a promising opportunity for reducing the potential greenhouse gas (GHG) emission. However, existing techniques face notable limitations for prompt detection. This study explores the utility of Vegetation Indicators (VIs) to reflect vegetation health deterioration, thereby representing leak-induced stress. Despite the acknowledged potential of VIs, their sensitivity and separability remain understudied. In this study, we employed ground vegetation as biosensors for detecting methane emissions from underground pipelines. Hyperspectral imaging from vegetation was collected weekly at both plant and leaf scales over two months to facilitate stress detection using VIs and Deep Neural Networks (DNNs). Our findings revealed that plant pigment-related VIs, modified chlorophyll absorption reflectance index (MCARI), exhibit commendable sensitivity but limited separability in discerning stressed grasses. A NG-specialized VI, the optimized soil-adjusted vegetation index (OSAVI), demonstrates higher sensitivity and separability in early detection of methane leaks. Notably, the OSAVI proved capable of discriminating vegetation stress 21 days after methane exposure initiation. DNNs identified the methane leaks following a 3-week methane treatment with an accuracy of 98.2%. DNN results indicated an increase in visible (VIS) and a decrease in near-infrared (NIR) in spectra due to methane exposure.