Underground Diagnosis Based on GPR and Learning in the Model Space

Area of Science:

• Geophysics
• Signal Processing
• Machine Learning

Background:

• Ground Penetrating Radar (GPR) is crucial for underground detection.
• Automated identification of underground structures from GPR data is challenging due to unknown data characteristics.

Purpose of the Study:

• To propose a novel GPR B-scan image diagnosis method for automated underground anomaly identification.
• To enhance the accuracy and efficiency of pipeline detection and underground diagnosis.

Main Methods:

• A 2-Direction Echo State Network (2D-ESN) was developed to learn from GPR image segments.
• The 2D-ESN captures dynamic GPR image characteristics by considering horizontal and vertical connections.
• Semi-supervised and supervised learning were applied to the 2D-ESN models for diagnosis.

Main Results:

• The proposed 2D-ESN method effectively fits GPR image segments.
• The model demonstrated strong performance in capturing dynamic GPR image characteristics.
• Experiments on real-world datasets confirmed the effectiveness of the 2D-ESN for underground diagnosis.

Conclusions:

• The learning-in-the-model-space approach using 2D-ESN offers a stable and parsimonious representation for GPR data.
• The proposed method significantly improves automated underground structure identification from GPR B-scan images.
• This technique enhances the reliability of GPR in practical pipeline detection and underground diagnosis applications.