Rapid estimation of bulked root mass is possible with GPR. These results show correlation strength up to 79% using these methods. Further, we have demonstrated that increased VWC can improve the detection of bulked roots, as long as the dielectric is homogeneous across the study. The bulk dielectric of soil is driven primarily by water, and the interfaces of dielectric change cause the reflection of GPR energy. With a sufficiently high signal frequency, soil structure has the potential to introduce noise in GPR data through the minute reflection and scattering of EM energy, driven by soil features such as compaction layers, aggregates, and pores. By increasing the VWC, some soil pores fill with water, and the dielectric heterogeneities are reduced, leading to less noisy GPR data. Additionally, as the dielectric of the soil increases, the signal velocity reduces, effectively increasing the sampling resolution of the system, as the sampling is a function of time.
These two factors may explain why increased VWC improved the strength of correlation. Indeed, we hypothesize that as VWC variance decreases, the strength of the correlation should increase, possibly maximizing the predictive potential near field capacity. Unfortunately, soils near field capacity are easily compacted, and are difficult to work in. Therefore, some compromise must be found to maximize predictive power of the GPR and minimize the impact and difficulty of field work. This optimal level of soil moisture is most likely dependent on soil texture, and could be expressed as a fraction of field capacity. Further studies in multiple soil types should lead to standardized recommendations of optimum water content for major texture groups.
This study, like others before it, presents a supervised correlation – that is, the depth and mass of roots is known, so it becomes less difficult to determine the optimum depth of radar information to analyze. The window of analysis is relatively narrow – only five rows, or approximately 2 cm of soil depth – and selecting the correct depth without previous knowledge of the root depth is difficult at this time. As research continues, it may become possible to distinguish the zone of highest information density, and researchers are already working towards that goal [21]. In this study, however, the noise was too great to establish the root zone from only GPR data. For this application, noise may be considered as all recorded energy which is not reflected by plant roots. As discussed earlier, GPR systems record all intercepted energy in the antenna range, regardless of the energy source. Noise may also be generated within the GPR system itself, and there has been some indication that the prototype system used here is not immune to this type of noise. This can be reduced by careful engineering, and through data filtering, if the inherent noise has been characterized. Other sources of noise include reflections and scattering caused by variations in soil structure, stones, clay clods, surface roughness, and above-ground biomass. Because we placed roots in the soil, rather than growing them, aboveground biomass was not an issue in this study, but has been in other data which are not yet published.
As noted, the soil type and water content have a large effect on GPR data. This variation makes it difficult to build a unified correlation between studies, fields, or even dates. As such, GPR remains a relative measure of root mass, suitable for ranking within a single field and date, otherwise requiring a specific calibration at each use. It remains possible that a correction factor could change this. Inclusion of multiple blank plots in the study may provide that correction factor, such that data can be normalized to the feature values of the blank plot, accounting for the soil type and moisture content. Further studies are planned to investigate this possibility. Without locational correction, GPR data may still be used to rank plots for genotype, and rankings may be compared across locations and/or time.
These results demonstrate the effect of soil moisture not just on the ability to pick out roots, but also the effect on the method. As mentioned in the results, the depth of best correlation was deeper for the dry treatment than the irrigated treatments, which was unexpected. GPR energy is reflected at the interface of dielectric contrast. When the object causing the reflection, such as a root, has sufficient diameter, the reflection may happen at both interfaces on the signal vector – that is, it can reflect from both the top and bottom of the root. In other uses of GPR, the thickness of large objects can be estimated by measuring the distance between the top and bottom reflection. In this study, however, discreet returns were not observed; rather, the total reflected energy for a volume was measured. It is possible the reflected energy in the dry treatment was more intense at the bottom of the root, whereas the irrigated treatments reflected primarily from the upper interface. In all treatments, a nearly continuous range of window depths showed significant correlation to root mass, indicating that information about the root mass was present across a depth corresponding approximately to the root diameters. This also suggests the possibility of some distinct characteristic for that region, such that it may be possible to find that region using machine learning techniques, so that supervised correlation is no longer required, and furthering the usefulness of GPR as a predictive tool.
In 2019, Delgado et al. reported a similar study designed to test commercially available GPR models in bulked root imaging [18]. A C-Thrue GPR system (IDS Georadar) was mounted to a computer controlled gantry and passed over a climate controlled sandbox. Cassava roots of varying sizes were buried at orientations parallel, orthogonal, and 45° to the scan direction. A single antenna pair was passed in transects at 2.5 cm intervals over the sandbox, with signal pulses every 0.2 cm. The GPR data were interpolated to form 3D models of the buried roots and interpolated image dimensions were compared to physical dimensions. The study illuminates several important factors for the application of GPR to root measurement, namely, the superiority of vertical antenna polarization over horizontal, and the effect of root orientation on measurement accuracy. However, the study differs significantly from the current – the focus was on 3D imaging rather than mass estimation, a single antenna pair was used in a high-density grid rather than an antenna array, the antenna was ground coupled rather than air-launched, and the soil medium was air dry such that no effect of soil water content was studied. Finally, the data collection method was not appropriate for high volume phenotyping.
The application of GPR for the quantification of roots is still in its infancy, and significant research is required before it can be used as a predictive tool. We have shown here that GPR data contain information about root mass, but it is also clear that other factors influence the data, and noise is a problem. Radar data is highly sensitive to processing parameters, such that adding or removing a step readily effects correlation. The presented methods utilized a multi-channel radar to rapidly collect 3D information. To produce the 3D information, the channels were interpolated using simple linear interpolation. Similar to other 3D data, such as LiDAR, care must be taken to align the interpolated entities. In the case of GPR, the primary point of alignment is usually the soil surface, because it is discrete and constant. In this study, the field was level, and the antenna was facing straight down at nadir, resulting in well-aligned channels with consistent positioning of the surface between channels. This is not always possible. In many cases, the antenna array cannot pass directly over the center of the root mass because plants are still present, so the antenna may be angled to point towards the plant center. Additionally, errors in channel calibration can produce small offsets that change the apparent height of the antenna relative to the ground surface. Finally, uneven ground surface can cause differences between channels. In these cases, the ground surface must be identified in each channel so they can be aligned before interpolation, as described by Dobreva et al [21]. Automated methods of identifying the ground surface would greatly reduce the time required for channel alignment.
Though each application will have its unique problems, there remain several constant considerations which we suggest become standard practice when using GPR to measure roots. Foremost among these is to understand your radar system. Unlike visual tools such as LiDAR, GPR emissions are not highly focused and are generally shaped like an ellipsoid bubble, meaning the energy extends in front of, behind, and to the sides of the antenna. This is why at least 1 m was allowed between the cart and the first study plot, so that initial readings would be outside the plot. This also means care must be taken for transitory reflectors, such as workers, to not enter the volume of sensitivity while collecting the data.
GPR data are highly dependent on the dielectric of the soil; therefore, it is strongly recommended that dielectric measurements always be made at the time of scanning. This can be done in many ways, such as measuring dielectric directly with a probe, measuring water content and converting using the Topp equation, or by burying a reflector at a known depth, which allows a velocity estimation by dividing the known depth below the surface by the difference in signal time from the surface to the reflector [24, 29]. Knowing the dielectric, or the signal velocity (see Eq. 1), allows the conversion of data from the time domain to the space domain, enabling estimation of depth. Further, some GPR processing techniques require these parameters. Many studies will be interested in the root mass at certain depths, as is currently measured with destructive techniques. This is only possible if the signal velocity is known.
Published methods to date have relied on measuring reflected energy, whether by amplitude threshold and pixel counting, or summations of other features. These techniques are inherently tied to the volume of soil analyzed, meaning that plot size will auto-correlate with feature count. Plot size must therefore be carefully controlled. In this study, plot size was controlled in the field through careful measurement and marking. Other studies have controlled plot size by cropping the data, and others have controlled by conversion to either feature density, root density, or both. We recommend the former whenever possible, as it protects the integrity of the data. However, current root phenotyping methods frequently use root density as a measurement and is acceptable to many researchers [30].
Whichever way the plot length is controlled, the data must be related to the field. Some GPR systems are capable of integrating GPS data into the scan data, while others can utilize digital markers. Some have neither capability, thus plot positions must be derived another way, possibly by placing reflectors at plot ends. Experience dictates caution in the latter method – the reflector must be easily identifiable in the radargram, and reflectors placed on the soil surface are easily lost in the surface reflection. In such cases, an aerial reflector is recommended.
Finally, based on the results of this study, care must be taken to ensure homogeneous dielectric environment at the time of scanning. Depending on the hydraulic conductivity of the soil, several days may be required after an irrigation event.