Multi-scale weighted fusion network for hyperspectral and LiDAR data to identify stressed vegetation caused by natural gas storage microleakages

Microleakages in underground natural gas storage, which pose significant potential safety and environmental risks, can be monitored by observing the responses of surface vegetation to gas stress. However, previous methods primarily employed the spectral response of stressed vegetation as an indicator to detect microleakages and their accuracy was limited by spectral biases, while the structural response of stressed vegetation was often overlooked. Therefore, this study proposed a multi-scale weighted fusion network (MsWFNet) integrating spectral and structural response from hyperspectral and LiDAR data to identify natural gas-stressed bean and grass collected through field experiments. This network uses multi-scale hyperspectral-LiDAR data as input and performs weight learning on scale data during training to simultaneously extract high-quality features indicative of different vegetation stress states across multiple scales, thereby identifying natural gas-stressed vegetation and locating microleakage points. Accuracy assessments demonstrate pixel-level identification accuracies of 94.13% and 94.63% for bean and grass plots, respectively, with mean absolute location errors of 0.1183 m and 0.0787 m, outperforming existing methods. This work demonstrates a considerable improvement of detection accuracy and robustness by combining hyperspectral and LiDAR data to identify vegetation stress from natural gas microleakages, which is promising for further applications.