The size and depth of fruit tree roots can reflect the growth and health of fruit trees and affect the profits of the orchardist. However, the roots are more difficult to observe and sample than the subaerial parts of fruit trees, such as the tree trunk, branches, and crown. Ground penetrating radar (GPR), as an emerging non-destructive testing technology, has the advantages of simple operation and convenient carrying. However, using GPR to quantify the radius of the roots is still a challenging task. To that extent, a prediction method for tree root radius and depth was proposed based on GPR and convolutional neural networks. Firstly, the simulated one-dimensional data of ground penetrating radar (A-Scan) was used as the data set to train the model. Secondly, the attention mechanism allocated more weights to essential features, highlighting key features and speeding up convergence. Finally, the feature information was extracted through the convolutional layer. The local features learned by the previous convolutional layer were integrated into the global features of the A-Scan data through the fully connected layer to predict the root radius and depth accurately. The model was tested on simulation data and real data. In the simulation experiment, the maximum error of root radius prediction was 2.9mm, the coefficient of determination value was 0.990, the root mean square error was 0.00068m, the maximum error of root depth prediction was 11.2mm, the coefficient of determination value was 0.999, and the root mean square error was 0.0020m. In the field experiment, the maximum error of sample roots radius prediction was 1.56mm. The maximum error of sample roots depth prediction was 9.90mm. The total average relative error was 5.83%, indicating the proposed method’s efficacy for estimating the radius and depth of roots.