Advancements in Using Deep Learning Methods for GPR Detection of Tree Roots

In recent years, the effects of emerging diseases have caused significant worries among environmentalists and communities, requiring putting efforts into the monitoring and management of natural resources. In this regard, tree roots are one of the most vital and fragile organs of the tree, as well as one of the most complex to investigate. In this way, non-destructive testing (NDT) methods have become one of the most popular techniques for assessing and monitoring tree roots, as opposed to conventional destructive techniques. In this context, ground penetrating radar (GPR) applications have proved to be precise and effective for investigating and mapping tree roots. The inhomogeneity of the soil, however, is a significant obstacle towards the GPR identification of tree roots, and a deep learning (DL)-based method has been recently proposed to tackle this issue. This research, therefore, aims to improve upon the above-mentioned approach, by customising two convolutional neural networks (CNN) methods for the analysis of GPR spectrograms. In this study, the GPR signal is first processed in both the temporal and frequency domains to filter out noise-related information, and subsequently spectrograms are generated. Afterwards, two specifically modified CNN classifiers are implemented and then compared to other DL methods, already validated for tree roots detection. The findings of this study further support the viability of the suggested methodology and open the way for the application of new approaches for evaluating tree root systems.