Machine learning for 3D reconstruction of tree roots

Abstract

Trees play an integral part in our ecosystem and the planet. In particular, tree roots are essential structural components in protecting tree health. Their key functions such as the absorption of water and preventing soil erosions prevent flooding and help preserve groundwater reserves. Thus, a better understanding of these root systems is necessary. Ground penetrating radar (GPR) is a useful non-invasive tool that allows scanning of the soil environment without causing harm to tree roots. In this project, a novel deep learning framework to extract the features and 3D reconstruct root architectures from GPR data is proposed. Techniques like domain adaptation are also implemented to aid in the process. The framework comprises of three main steps: (i) conducting research on various deep learning models, (ii) building and training a deep learning model for feature extraction and data reconstruction on simulated data in the source domain, and (iii) testing the deep learning model on real data in the target domain. gprMax software was used to create simulations before using data from real soil environments. The results obtained from the simulation and real data show a potential in the deep learning model developed.