A Deep Learning-Based Methodology for Rapidly Detecting the Defects inside Tree Trunks via GPR

This paper proposes a deep learning-based approach for rapidly detecting the defects inside tree trunks via ground penetrating radar (GPR) technology. In this approach, GPR measurements are performed centimeters-away from the surface of tree trunk on a straight trajectory. The n the B-scans obtained from GPR measurements are processed via a deep learning algorithm to detect the defects inside the tree trunks, classify their types, and estimate their sizes/severities. An open-source finite-difference time-domain (FDTD) simulator is used to produce a large set of B-scans from random realizations of realistic 2D tree trunk cross-sections without and with different size of defects (cavities, decays, and cracks). The data set is then used to train and test a six-layer convolutional neural network (CNN) with drop-out layers and weight regularization to avoid overfitting. Our preliminary results show that the testing accuracy of the CNN algorithm is more than 90%. The testing results demonstrate that the current methodology al lows accurately detecting the types and sizes of defects inside tree trunks to monitor the health condition of trees.