AUTOMATIC HYPERBOLA DETECTION AND APEX EXTRACTION USING CONVOLUTIONAL NEURAL NETWORK ON GPR DATA

• Main Author: Dewantara, Daffa
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/61826
• Institution: Institut Teknologi Bandung

Ground Penetrating Radar (GPR) is a non-destructive geophysical method used for subsurface mapping studies. The cylindrical objects detected by GPR method have a hyperbolic signal pattern on the radargram. The shape of the hyperbola depends on the depth, the material of the buried object, and the material of the surrounding environment. In this study, a framework is proposed to minimize the time required for detecting hyperbolic patterns on a radargram. The framework that was developed is involving the use of convolutional neural network, and image processing techniques presented through 2 different modules. The first module is a hyperbola detection module which has a boundary box output containing a hyperbolic pattern from a radargram input in a raster format. The detection model in the first module is trained using synthetic radargram data simulated by gprMax. Synthetic radargram simulation is optimized using computational graphics processing units(GPU) to shorten simulation time. The hyperbola detection module was evaluated on 96 synthetic radargrams and resulted in a precision value of 93.98% and a recall of 73.4%. The second module is apex coordinate extraction which has a hyperbola vertex coordinate output. The apex coordinate extraction module contains a coordinate search algorithm on the image by searching for pixels with the maximum intensity from the designed search window. The apex coordinate extraction module has been successfully implemented on a hyperbola that is not interfered with by the subsurface boundary. Using the developed framework, the detection of buried cylindrical objects using GPR can be done automatically.