Suppressing ground penetrating radar clutter to predict root parameters using deep neural networks

Non-destructive testing of tree roots has the vast application potential to the root system evaluation of fruit and ancient trees for the better management of plant health. Among them, ground-penetrating radar (GPR) can be expected to non-destructively detect the tree roots, due to the portable and low-cost. However, the interpretation of GPR data is dependent mainly on the manual analysis, leading to the reduced automation and accuracy. Moreover, the clutters and noise that caused by soil heterogeneity can severely affect the accuracy of root detection and identification. In this study, a detection method was proposed for the clutter suppression and root parameter prediction of ground penetrating radar. The data collection was divided into two parts: the dataset and field test. The datasets consisted of simulated and synthesized real data. The roots with the radius ranging from 5 to 30 mm were randomly distributed within 300 mm underground. The field experiment was conducted in three sand pits. The soil heterogeneity was simulated to create the different moisture levels, when adding water into different parts in two of the dry sand pits. The rest sandpit was a condition of higher water content after heavy rainfall. At the same time, the roots were excavated to verify the model under real soil conditions. Four test pits contained the roots of different tree species, the different soil environments, as well as the different radii and depths. The attention mechanism was integrated into the U-net model, in order to enhance the capability in the clutter suppression and root target reflection restoration. A clutter suppression network was constructed to test the effectiveness of the model. The encoding-decoding structure of U-net was separated the target root hyperbolic reflection from the overall background. The attention mechanism enabled the network to better adaptively focus on the hyperbolic reflection of the target root system without identifying the clutter as the target reflection. The clutter suppression was realized for the original B-scan image to remove the adverse effects, due to the soil heterogeneity and radar antenna coupling. Furthermore, the network of root parameter prediction was constructed using residual blocks and inception. The multi-scale receptive field of inception was used to extract the global and local features, and simultaneously to predict the root radius and depth. The predicted results were obtained by parallel inputting the clutter suppressed image and the original B-scan image into the network. The effectiveness of the model was evaluated using datasets and field experiments. The test results of clutter suppression were evaluated using peak signal to noise ratio (PSNR) and structural similarity (SSIM), which were 39.42 dB and 0.991, respectively, better than before. The test results on datasets showed the mean absolute error (MAE) of 1.7 mm, and the coefficient of determination R² value of 0.914 for the root radius prediction, while the mean absolute error (MAE) of 6.3 mm and the R² value of 0.989 for the root depth prediction. The clutter suppression was also achieved in the better performance than the rest on the field data. The maximum prediction errors for the root radius and depth were 1.85, and 13.6 mm, respectively, where the total average relative error was 6.55%. Influencing factors were determined for the prediction effectiveness of ground-penetrating radar. The future research directions were pointed out as well. The findings can be applied to the actual prediction of root depth and radius of tree roots, particularly for the health administration of fruit trees and the protection of ancient trees.protection of ancient trees.protection of ancient trees.