A convolutional block multi-attentive fusion network for underground natural gas micro-leakage detection of hyperspectral and thermal data

Underground natural gas micro-leakage can be indirectly detected by stressed vegetation on hyperspectral and thermal images. This paper conducted field simulation experiments to acquire images of grass, soybean, corn, and wheat. Considering the complementarity between hyperspectral and thermal imaging, a Convolutional Block Multi-Attentive Fusion Network (CBMAFNet) was proposed for feature-level data fusion to identify stressed vegetation under natural gas micro-leakage. The CBMAFNet model incorporated a channel attention mechanism, a spatial attention mechanism, and a three-dimensional convolutional neural network (3D CNN), which the spectral features, spatial features, and canopy temperature features were effectively extracted and hierarchically integrated for classification. Finally, the gas leakage was successfully detected, and the suspected leakage point was accurately localized by fitting circular-like features of stressed vegetation. Experimental results demonstrated that the CBMAFNet achieved an average overall accuracy of 94.60 % and a kappa value of 0.93, indicating significantly better performance compared to methods using single-sensor data (OA: 86.20%–92.42 %, kappa: 0.83–0.90). In addition, the proposed model exhibited robust performance across four types of stressed vegetation, achieving accuracies of 97.12 % for grass, 93.46 % for soybean, 95.29 % for corn, and 92.51 % for wheat. This study provides a new method for natural gas micro-leakage detection with high potential for practical applications.