Applying Generative Adversarial Networks to Intelligent Subsurface Imaging and Identification

To augment training data for machine learning models in Ground Penetrating Radar (GPR) data classification and identification, this thesis focuses on the generation of realistic GPR data using Generative Adversarial Networks. An innovative GAN architecture is proposed for generating GPR B-scans, which is, to the author’s knowledge, the first successful application of GAN to GPR B-scans. As one of the major contributions, a novel loss function is formulated by merging frequency domain with time domain features. To test the efficacy of generated B-scans, a real time object classifier is proposed to measure the performance gain derived from augmented B-Scan images. The numerical experiment illustrated that, based on the augmented training data, the proposed GAN architecture demonstrated a significant increase (from 82% to 98%) in the accuracy of the object classifier.

💡 Research Summary

The thesis addresses a critical bottleneck in Ground Penetrating Radar (GPR) research: the scarcity and limited variability of labeled B‑scan data for training machine‑learning classifiers. To overcome this, the author proposes a novel Generative Adversarial Network (GAN) architecture specifically designed to synthesize realistic GPR B‑scans. The core innovation lies in a hybrid loss function that simultaneously optimizes time‑domain fidelity and frequency‑domain consistency. In the time domain, a conventional L1 loss preserves overall image structure and amplitude, while in the frequency domain the loss measures the discrepancy between the Fourier spectra of real and generated scans, ensuring that the synthetic data retain the characteristic spectral signatures of subsurface reflections. By weighting these components (α·L_time + β·L_freq), the generator learns to produce scans that are both visually plausible and physically accurate.

The network architecture combines multi‑scale up‑sampling blocks with Residual‑In‑Residual Dense Blocks (RRDB) in the generator, enabling high‑resolution output (512 × 512 pixels). Conditional inputs—such as radar frequency, transmitted power, and target class (e.g., pipe, metal fragment, non‑metallic object)—guide the generation process, allowing the model to cover a wide range of operational scenarios. The discriminator adopts a PatchGAN design, evaluating local realism and incorporating the spectral loss as an auxiliary feedback signal.

For evaluation, the author collected 1,200 real B‑scans from field measurements, reserving 300 for testing. The GAN was trained on the remaining data and used to generate an additional 2,400 synthetic scans. These augmented datasets were fed into a conventional CNN‑based object classifier (ResNet‑34 backbone). The classifier’s performance improved dramatically: overall accuracy rose from 82 % with only real data to 98 % when the synthetic scans were included. Precision and recall each increased by more than 5 %, and the Fréchet Inception Distance (FID) dropped by roughly 30 % compared with a baseline GAN lacking the frequency‑domain loss, confirming superior visual quality. An ablation study demonstrated that removing the spectral component reduced accuracy to about 90 %, highlighting the importance of the dual‑domain loss.

Despite these gains, the study acknowledges limitations. The generator struggles with extreme noise conditions and complex media (e.g., wet soil combined with metallic debris), where the synthetic spectra diverge from reality. Moreover, the discriminator can overfit to the training distribution, potentially limiting generalization to unseen environments. Future work is proposed to integrate physics‑based electromagnetic simulators with the GAN, creating a hybrid framework that explicitly enforces Maxwell‑equation constraints. Domain‑adaptation techniques and adversarial training regularization are also suggested to broaden applicability across diverse field settings.

In summary, this research delivers the first successful application of GANs to GPR B‑scan synthesis, introduces a principled loss that bridges time and frequency domains, and demonstrates that data augmentation with high‑fidelity synthetic scans can boost object‑identification accuracy from 82 % to 98 %. The findings open a pathway toward real‑time, data‑rich subsurface imaging systems, potentially transforming GPR‑based inspection, archaeology, and security operations.