Abstract: Abstract: Non-destructive testing of tree roots can contribute to explore the interaction between plant roots and soil. Currently, ground penetrating radar (GPR) can make it possible to non-destructively measure tree roots. However, the complex images from GPR are difficult to parse, due to the low automation level and accuracy. In this study, an automatic estimation of root and diameter was proposed using YOLOv3 to realize the intelligent recognition and extraction of root reflection hyperbola region of interest (ROI). Image processing such as histogram equalization was applied to improve the difference between the hyperbola and background, thereby to optimize the dataset of edge detection. Random Hough transform (RHT) with strong anti-clutter interference was used to realize the precise positioning of the apex of the hyperbola. A Savizky-Golay filter was introduced to smooth the discrete sequence of GPR one-dimensional data (A-scan), thereby to improve the prediction accuracy of root diameter. A diameter estimation was addressed for the distribution of two positions of hyperbola vertex in A-scan data. A root embedding field test was conducted to evaluate the effectiveness of models, further to determine the influence of root diameter, buried depth, and orientation in three dimensions on the parameter prediction of root system. The total average relative error of root position and diameter was within 10.57% in all experiments. All embedded roots were accurately identified in the test of root diameter, where the average relative errors of diameter, X- and Y-position were 16.04%, 2.22%, and 4.81%, respectively. The embedded roots at different depths were also accurately identified in the test of root buried depth. The apex of the hyperbola was still accurately located in the test, although there was a little interference from clutter. The vertical positioning was corrected through multiple sets of GPR two-dimensional data (B-scan), when the root was tilted along the vertical direction in the root orientation test. The recognition precision and recall rate reached 96.62% and 86.94%, respectively. The average detection time was 40ms for a single image. A better performance was achieved to identify the thick roots in the field test, compared with the threshold segmentation. In addition, two influencing factors were determined to improve the data accuracy, including the soil and root moisture content. Three influencing factors were determined for the predicted accuracy, including the training degree of network model, the image processing, and the RHT threshold. Finally, the improvement strategy was given in the future research. The experimental and analytical results demonstrated that the proposed method can automatically extract the ROI of root reflection hyperbola, further to locate accurately the hyperbola vertex, and finally to replace manual calibration and interpretation. The finding can make a great contribution to the practical exploration for the root location and diameter prediction in the non-destructive testing of tree roots.