Attention-Module-Guided Time-Lapse Leakage Plume Imaging Driven by LeakInv-CUNet GPR Inversion Framework

Accurately characterization of pipeline leakage by capturing time-lapse features and their pathway orientations using ground penetrating radar (GPR) is crucial for optimizing the operational efficiency of water supply systems and mitigating water resource losses. However, leakage imaging concerning temporal-spatial details remains a significant challenge for current deterministic or probabilistic inversions due to the complexity of signals, environmental interference, and computational burden involved in inversion. This paper develops a novel attention-guided GPR inversion framework, LeakInv-CUNet, to implement refined leakage imaging of time-lapse features and plume pathways. To enhance the training of the LeakInv-CUNet network, extensive GPR datasets are created by augmenting data with simulations and experimental detections to account for variations in injection orientation, plume evolution, and the properties of the subsurface media. By leveraging the dual virtues of Convolutional Block Attention Module (CBAM) and UNet architecture, the developed LeakInv-CUNet framework is capable of extracting critical and even slight response features triggered by leakage anomalies, allowing for refined imaging of leakage plumes and their orientations. Specifically, the training is fed by the inversion of GPR datasets corresponding to different permittivity distributions of leakage, which is subsequently feedbacked through functional mapping of the Topp equation nexus to water content distribution imaging. In particular, we concentrate on the leakage features by taking into account the effects of water injection orientation and time-lapse characteristics. Simulation and on-site field experiments are conducted to assess its performance in terms of anti-noise robustness and practicality compared to the conventional UNet and Enc-Dec networks. With a mean absolute percentage error (MAPE) of attention-guided leakage imaging less than 0.707%, and a structural similarity (SSIM) and peak signal-to-noise ratio (PSNR) greater than 0.924 and 45.416, respectively, the attention-guided leakage imaging exhibits excellent efficiency and accuracy for potential applications in early leakage warning and pipeline time-lapse monitoring.