Detecting natural gas storage microleakage based on K-means clustering under constraint of Jeffries-Matusita distance criterion using mobile LiDAR data.

Natural gas leaks alter both the spectral reflectance and the structure of surface vegetation, which can be used to indirectly monitor microleakages in gas storage facilities. However, existing methods predominantly focus on the spectral rather than structural response of stressed vegetation, and it is not clear whether structure characteristic can be used to identify natural gas stressed vegetation. In this study, the utility of mobile LiDAR in detecting vegetation structure changes due to natural gas stress was demonstrated by analyzing LiDAR data from a field experiment with bean and grass plants in their growing phase. A method utilizing the Jeffries-Matusita distance criterion constrained K-means clustering (JCKC) algorithm was proposed, which comprises three main steps: First, response of vegetation structure characteristic to natural gas stress was quantitatively analyzed at plot and pixel scales using LiDAR data. Second, the optimal set of structure characteristic parameters indicating natural gas stressed vegetation was determined using hierarchical clustering algorithm. Third, the reduced LiDAR data was clustered using K-means algorithm, and the clusters were classified under constraint of Jeffries-Matusita distance criterion to identify stressed vegetation. The results indicated natural gas stress significantly changes vegetation structure (p = 0.05), decreasing parameters like height, projected leaf area, canopy relief ratio, coefficient of variation of vegetation height, and entropy, while increasing homogeneity, contrast, and dissimilarity. The set of structure characteristic parameters based on height, homogeneity, and contrast can stably indicate natural gas stress, with Jeffries-Matusita distance values for comparing healthy and stressed vegetation samples exceeding 1.8. The proposed model achieved pixel-level identification accuracies of 98.95% for bean and 96.22% for grass, with average localization accuracies of 0.15 m and 0.12 m, respectively. This study demonstrates the potential of vegetation's structure characteristic in reflecting response to natural gas stress and monitoring natural gas storage microleakage in vegetated areas.