In the master's thesis of research assistant Orhan Apaydın at the Department of Geophysical Engineering, under the supervision of Prof. Dr. Turgay İşseven, deep learning-based methods were investigated for the interpretation of data from ground-penetrating radar (GPR), a geophysical method used in near-surface investigations. GPR technology is a widely used method to investigate the buried objects within the near-surface. These objects typically appear as hyperbolic structures in GPR images, with the shape of the hyperbola varying based on the geometric properties of the objects. This study investigates the effectiveness of deep learning-based object detection methods in identifying and classifying these hyperbolic structures in GPR images. Specifically, the study explores the use of deep learning algorithms such as Faster R-CNN, YOLOv5, and SSD (Single-Shot Detector) for detecting and classifying objects according to their geometric shapes. In the experiments, objects with rectangular and cylindrical shapes were used to evaluate the detection and classification capabilities of these models. The deep learning models successfully identified the objects with high accuracy, classifying each object based on its geometric shape. The results demonstrate the potential of deep learning methods in improving object detection and classification in GPR images. This study highlights the growing role of deep learning in enhancing the accuracy of subsurface investigations. As these technologies continue to evolve, they will enable more efficient and precise fieldwork, leading to better mapping and understanding of subsurface structures. The integration of deep learning in geosciences will likely play a crucial role in advancing exploration and excavation techniques in the coming years.

Figure. Network architecture of Faster R-CNN. There are two learnable stages: region proposal network and classification network (Apaydın & İşseven, 2024).

Figure. Some object detection and classification test results in GPR images of (a) Faster R-CNN, (b) YOLOv5, and (c) SSD models. Models output detected and classified objects with a bounding box and confidence score in percentage (Apaydın & İşseven, 2024).

Reference:

Apaydın, O., & İşseven, T. (2024). Detection of objects with diverse geometric shapes in GPR images using deep-learning methods. Open Geosciences, 16(1), 20220685.