Deep learning algorithm for tree defect detection

Abstract

This paper contains the final report for the final year project titled ‘Deep Learning Algorithm for Tree Defect Detection’, project number B3005-211. The arborist in Singapore uses visual as the first level of inspection; however, trees may look strong and durable on the outside but full of cavities and decays internally. Unnecessary lives were lost when potential tree fall went under the nose of the arborist and eventually gave out. The paper explores the possibility of using deep learning approach to learn and categorize whether a tree is healthy or abnormal. The ground penetration radar (GPR) was selected for producing the B-scan images which will be used as the dataset for the neural network. The dataset will be obtained using GprMax, an open-source electromagnetic (EM) simulation software. The tree models used for the simulation are randomly generated through a MATLAB software with arbitrary tree and cavity size. The simulated GPR will be positioned a short distance away from the tree surface and travel in a lateral direction. The scans obtained from the radar will be preprocessed before using as the data for the deep learning. Results have shown that the convolutional neural network is able to produce a validation accuracy of higher than 95% and testing accuracy of higher than 92%. In other words, use deep learning for detecting of tree defects is a feasible approach which can produce high accuracy outcomes.