Mapping Tree Roots Using GPR and Machine Learning

Urban environments present unique challenges for long-term urban tree survival and vitality. Drought, temperature extremes, pollution and contaminants - such as pavement de-icing salts - soil compaction, and mechanical damage to trees, including their roots, are some of the primary factors that contribute to tree decline, morbidity, and mortality in urban areas. 

Urban development in the form of land clearance and construction, as well as insufficient soil volume for trees to grow adequate root systems, are leading causes of poor tree health and mortality (Bassuk et al., 2011; Day et al., 2010; Tobin et al., 2007). The roots of urban trees are vulnerable to injury through mechanical disturbance of their soils and occur due to adjacencies to redevelopment and construction activities. In a city like Toronto, construction disturbance is a leading cause of the decline of urban trees (City of Toronto, 2021). Therefore, proper management, protection, and enhancement of roots in urban locations can help mitigate these risks and promote healthy tree growth and flourishing.

Such management would be aided by better knowledge of a tree’s root system and its spatial configuration under the ground, and to better understand the locations where construction activity or soil disturbance is most at risk of damaging and injuring a root or collection of roots. In contemporary practice, the spatial extent of the tree root system has been estimated based on the diameter of tree trunk at breast height (DBH) metric, or the horizontal extent of the tree’s crown. For example, the City of Toronto establishes, in its Tree Protection Zone By-Law, work boundaries that constrain and prevent construction work within a radius from tree’s trunk, which estimates the minimum distance required to protect the tree’s root systems from harm. In practice, these distances are gross rules of thumb that do not consider the site-specific details about the individual’s roots that would enable a more accurate determination. 

Understanding the exact spatial extent and configuration of an urban tree's root architecture is difficult because most roots are not visible without removing the covering material using invasive “daylighting" techniques (such as fluid-spading or another method of manual excavation), that are potentially harmful to the tree. A non-destructive method for accurate mapping of tree roots would be valuable to practitioners whose work prioritizes protection of the tree.

Using GPR to detect tree roots

Ground Penetrating Radar (GPR) offers an alternative, non-invasive method of locating and recognizing tree roots without the requirement of any physical or manual excavation. Applications of GPR have been extensive in various geotechnical, structural, and archaeological contexts, where they minimize the costs of excavation-intensive surveying and provide a potentially more fulsome view of the subsurface geometry, spatial configuration, and material properties of the soil medium (Benedetto et al., 2016). The use of GPR is relatively limited in tree root system surveys, and potential practical applications are complicated in urban conditions, where the complexity of the soil environment may hamper any effort at acquiring high quality spatial data.

Most GPR devices emit a bi-directional antenna pulse of electromagnetic (EM) energy into the ground. The pulses – or waves - are then reflected, transmitted, and refracted by embedded targets – like roots - of varying geometries and EM properties (Bigman 2018; Conyers, 2004). After undergoing alterations to their intensity, some of the EM energy is received back by the GPR antenna and interpreted in the form of a sign wave called an A-scan (Conyers, 2004). Tree roots, because they contain water, often contrast in moisture content compared with the surrounding soil matrix. This creates a discontinuity in the subsurface environment that alters the EM pulse emitted by the GPR. The timing of each returning pulse is the basis of understanding the depth to reflecting root. It is by this principle that GPR can detect tree roots (Stokes et al., 2002). (Figure 1)

A major challenge with GPR data interpretation is that all data outputs tend to be noisy, and targets of interest can be difficult to recognize, if they are even detectable at all. Various filtering techniques may be used to clarify the GPR data. The GPR data can then be transformed and loaded into a program that builds a two-dimensional (2D) or three-dimensional (3D) representation of the clarified signal, which usually aids in user interpretation. Unsurprisingly, noisy data is correlated with more complex soil environments, often driven by varying water moisture in the soil footprint of the scan. Noisy images could be caused by the sheer morphological complexity of typical root systems, which tend to scatter and attenuate the energy of the returning pulse (Figure 2) 

Enhancing tree root detection

While using manual interpretation of GPR mapping can aid the field practitioner to estimate the locations of the tree roots, much could be gained both on the cost and confidence in the detections. Manual interpretation of the data can take time and be prone to error. What could look like a root system, or an individual root, could be something else, like a pipe with water flowing through it.

In this research, I aim to improve tree root detection by building an automated tree root detector using machine learning. The detector recognizes the dendritic structure of the root system network in the horizontal, top-down view. The detector first recomposes multiple scans into a suitable 3D representation of the data using interpolation. It then clarifies and denoises the data. For each horizontal view of the 3D data, it segments each view to simplify the classification task later. Finally, the segments are classified according to their fit to the expected morphology of a tree root system, gaining recognition of the linear, filament like forms that converge to a common emanation point – that is, the tree’s trunk – that are so commonly revealed in manual excavations.

The detection model that is built will be supported by numerous training data of roots in the soil, using samples derived from GPR scans using a soil box laboratory (Figure 3). 

At a minimum, the tree root detector should be able to recognize linear filaments in the horizontal view and taken at multiple depths. This would enable it to round out spatial areas of high confidence where there are tree roots (as opposed to soil with other objects). To increase model confidence, a greater scanning, and therefore, sample area must be obtained. The output of classification over a confidence gradient, especially if it could be delivered in a short time, would compliment the more intuitive judgements of a user who can also assess the data visually.