GPR-CUNet: Spatio-Temporal Feature Fusion-Based GPR Forward and Inversion Cycle Network for Root Scene Survey

作者：Xiaowei Zhang, Xuan Zhao, Shuang Li, Shenghua Lv, Lin Chen, Jian Wen · 发表于：IEEE Sensors Journal · 年份：2025 · DOI：10.1109/jsen.2024.3522888 · 被引用次数：8 · 研究领域：Geophysical Methods and Applications、Microwave Imaging and Scattering Analysis、Soil Moisture and Remote Sensing

Ground Penetrating Radar (GPR) forward and inversion methods are key techniques for studying radar imaging mechanisms and investigating subsurface scenes. Efficiently interpreting radar wave data will facilitate the development of subsurface structure detection applications, especially in the intricate plant root distribution. Existing forward and inversion models are constrained by the highly computational and time-consuming forward process, making it difficult to be applied to complex real-world subsurface scenarios. Inspired by the spatio-temporal properties during radar wave imaging, a spatial and temporal fusion cycle U-shaped model named GPR-CUNet was proposed. The model is more adapted to the transformation between permittivity distribution and GPR B-Scan data in complex environment. Firstly, to extract the spatial and temporal features from the permittivity distribution and radar data, a Spatio-Temporal Feature Fusion Module (STFM) based on CNN and BiLSTM was designed. Then, for the translation between the permittivity distribution and the radar wave data, two identical U-shaped networks with the STFM module were constructed. Finally, guided by predictive consistency and cyclic consistency, a hybrid loss function based on multi-scale structural similarity (MS-SSIM) and L1 norm was configured to boost the performance of both the forward and inversion networks. The numerical simulation experiments revealed that the proposed model imparted exceptional performance and eff...