Vadim Sabetski

Ground penetrating radar (GPR) is a non-invasive geophysical imaging tool capable of locating coarse diameter tree roots (>0.5 cm) in soil to approximately 1 m in depth. However, most applications of GPR for tree root detection have relied on manual interpretation by a skilled operator; this approach is subjective, time consuming and requires extensive operator experience. An automated root detection (ARD) approach to processing GPR scans for tree root identification has been proposed for commercial use by Tree Radar Inc. This research evaluates the accuracy of Tree Radar’s ARD approach at two spatial thresholds by surveying the tree root distribution at two sites using 400 and 900 MHz GPR antennas, followed by excavation of the actual tree root network and comparing field-verified root locations (n=563) to those determined using the ARD. Optimal levels of ARD user input settings, as well as algorithm sensitivity, were assessed by processing GPR scans under different operator scenarios. Results indicate that under optimal settings ARD can detect between 58% and 88% of roots greater than 1 cm in diameter with data from either of the 400 MHz or 900 MHz antennas; a small portion of these detections (19% to 24%) are characterized by high spatial accuracy (i.e., within 5 cm horizontal and 10 cm vertical distance from a field verified root) whereas most (38% to 64%) generally approximate the location of tree roots (i.e., +/- 15 cm horizontal distance). Logistic regression was used to determine if root characteristics and soil properties were good predictors of ARD detection accuracy. Soil texture (proportion of sand) displayed significance in modeling 900 MHz data, where root density and root depth were significant for 400 MHz data. For both antennas, the modeling was limited in explaining variance of results of both antennas (8% for 900 MHz and 13% for 400 MHz data), suggesting the variables included do not influence ARD detection accuracy.