Deep Learning-Based Ground-Penetrating Radar Inversion for Tree Roots in Heterogeneous Soil.

Area of Science:

• Geophysics
• Ecology
• Computer Science

Background:

• Tree roots are crucial for ecosystem health and resource management.
• Accurate detection of subsurface structures like tree roots is challenging.
• Ground-penetrating radar (GPR) is a non-destructive geophysical method for subsurface imaging.

Purpose of the Study:

• To develop a deep learning-based GPR inversion method for real-time imaging of tree roots and heterogeneous soil structures.
• To improve the accuracy and detail in subsurface material property (permittivity) inversion.
• To validate the method's effectiveness using both simulated and real-world GPR data.

Main Methods:

• A novel deep learning model, PyViTENet (pyramid convolutional network with vision transformer and edge inversion auxiliary task), was developed.
• PyViTENet combines pyramidal convolution and vision transformers for enhanced feature extraction.
• An edge inversion auxiliary task was incorporated to focus on structural details.

Main Results:

• PyViTENet outperformed other deep learning methods in accurately inverting permittivity and soil stratification from simulated GPR data.
• The model effectively captured the fine-grained heterogeneity of layered soils around tree roots.
• Transfer learning with PyViTENet on measured GPR data successfully reconstructed scatterer information (permittivity, shape, position).

Conclusions:

• The proposed PyViTENet demonstrates superior performance in GPR inversion for complex subsurface environments.
• The method offers high accuracy and generalization ability for non-destructive detection of underground structures and their surrounding media.
• This work provides a robust foundation for advanced GPR-based ecological and geological surveys.