A convolutional neural network for Creating Near Surface 2D Velocity Images from GPR Antenna Measurements

In this research, our focus lies in exploring the effectiveness of a frequency-velocity convolutional neural network (CNN) in the efficient and non-intrusive acquisition of 2D wave velocity visuals of near-surface geological substances, accomplished through the analysis of data from ground penetrating radar (GPR). In order to learn complex correlations between antenna readings and subsurface velocities, the proposed CNN model makes use of the spatial features present in the GPR data. By employing a network architecture capable of accurately detecting both local and global patterns within the data, it becomes feasible to efficiently extract valuable velocity information from ground penetrating radar (GPR) readings. The CNN model is trained and validated using a substantial dataset consisting of GPR readings along with corresponding ground truth velocity images. Diverse subsurface settings, encompassing different soil types and geological characteristics, are employed to gather the GPR measurements. In the supervised learning approach employed to train the CNN model, the GPR measurements serve as input, while the associated ground truth velocity images are utilized as target outputs. The model is trained using backpropagation and optimized using a suitable loss function to reduce the difference between the predicted velocity images and the actual images. The experimental results demonstrate the effectiveness of the proposed CNN method in accurately deriving 2D velocity images of near-surface materials from GPR antenna observations. Compared to traditional techniques, the CNN model exhibits superior velocity calculation precision and achieves high levels of accuracy. Moreover, when applied to unseen GPR data, the trained model exhibits promising generalization abilities, highlighting its potential for practical subsurface imaging applications.