A Deep Learning Approach for Tree Root Detection using GPR Spectrogram Imagery

Monitoring and conservation of natural resources have become increasingly necessary, as the impact of new diseases has created major concerns for environmentalists and communities in recent years. Within this context, tree roots are one of the plants' most important and vulnerable organs as well as one of the most challenging ones to inspect. To that effect, the non-destructive testing (NDT) methods have become one of the preferred methods of tree roots assessment and monitoring, as opposed to other conventional and destructive techniques. Within this context, applications of ground penetrating radar (GPR) have proven to be accurate and efficient for tree roots' investigation and mapping. However, a major challenge for GPR detection of tree roots is the soil inhomogeneity. This study aims to mitigate the uncertainty in root detection by proposing a deep learning method based on the analysis of GPR spectrograms (i.e., a graphic representation of a signal's frequency spectrum with respect to time). In this study, the GPR signal is first processed in both the time and frequency domains to filter the existing noise-related information and hence, to produce spectrograms. Subsequently, an image-based deep learning framework is implemented, and the effectiveness in detecting tree roots is analysed in comparison with conventional feature-based machine learning classifiers. The preliminary results of this research demonstrate the potential of the proposed approach and pave the way for the implementation of new methodologies in assessing tree root systems.