Synthetic ground-penetrating radar image generation using denoising diffusion probabilistic models

This paper addresses the issue of data scarcity in ground-penetrating radar (GPR) imagery by leveraging advancements in generative artificial intelligence (AI) to create realistic synthetic images. The study employs Denoising Diffusion Probabilistic Models (DDPMs) to generate synthetic GPR images, with an initial step involving data augmentation to enhance the training dataset. Specifically, small 128 × 128 patches are randomly cropped from the original large, rectangular GPR images (1480 × 400 pixels), addressing three critical challenges: 1) the limited number of available images, 2) the non-square aspect ratio, and 3) the computational cost associated with large image sizes. Through this augmentation process, the dataset size is expanded from 127 to 12,700 images. DDPM training utilizes attention layers integrated within the convolutional layers of a U-Net architecture, capturing critical features at lower layers where finer details are processed and preserving their spatial relationships at higher layers. After 600 epochs, the model converges with a near-zero training loss and is then employed to generate high-quality synthetic GPR patches (128 × 128). These generated images bear a strong resemblance to the original training patches to the extent that they are visually indistinguishable by human observers. The Fréchet inception distance (FID) score of the generated images is 41.8, indicating satisfactory quality synthesis. The findings of this study demonstrate the effectiveness of the proposed approach in addressing data limitations and facilitating the use of synthetic GPR images in the development of advanced machine-learning models for GPR applications.