DEVELOPMENT OF A NEURAL NETWORK-BASED ALGORITHM FOR DETECTING HYPERBOLAS IN GROUND-PENETRATING RADAR DATA

This article discusses the development of a methodology for creating a semi-synthetic dataset to train a neural network for the detection and segmentation of hyperbolic diffractions in ground-penetrating radar (GPR) data. The proposed method offers a solution by generating annotated data where target synthetic objects are placed onto real background sections that are confirmed to be free of the target objects. This approach enables model fine-tuning without direct manual labeling for any specific survey type. The final algorithm was successfully validated on GPR data from the Austre Grønfjordbreen glacier. This solution will significantly simplify the analysis of the internal structure of polythermal glaciers and allow us to assess their hydrothermal structure.