MSCCEAU-Net and physics-informed constraint jointly-driven intelligent inversion for Ground Penetrating Radar

Ground Penetrating Radar (GPR) inversion is the process of inferring subsurface dielectric constant distributions from GPR data. Enhancing inversion accuracy plays a critical role in underground target detection, geological structure interpretation and engineering exploration. Traditional GPR inversion methods suffer from heavy reliance on the initial model and high sensitivity to both noise and model parameters, thus severely limiting the inversion accuracy under complex geological settings. While deep learning reduces computational costs, the lack of physics-informed constraints often leads to inversion results easily deviating from fundamental electromagnetic principles. To address these challenges, this paper proposes an intelligent GPR inversion method jointly driven by an Multi-scale and Efficient-channel Aggregation U-Net (MSCCEAU-Net) model and physics-informed constraints: (1) A Deep Convolutional Neural Network (DCNN) is constructed, integrating Multi-Scale Cascaded Convolutions (MSC) and Efficient Channel Attention Mechanisms (ECA). This architecture significantly enhances the model's discriminative ability to target features and signal-to-noise characteristics through multi-scale feature fusion and channel weight optimization; (2) A Maxwell-equation-based physics-informed residual loss function is formulated. This enables a dual data-physics driven intelligent inversion, simultaneously satisfying both observational data matching and electromagnetic field physical laws, markedly improves the well-posedness of the solution and the inversion accuracy. The experimental results demonstrate that the method jointly driven by the MSCCEAU-Net and physics-informed constraints outperforms other traditional and deep learning-based inversion approaches in terms of inversion accuracy, achieving a precision improvement of over 65%. This approach realizes a synergistic integration of physics-informed constraints and deep learning architecture, providing a novel solution for GPR data inversion that combines theoretical interpretability and practical applicability.