@article{Zhao2026, 
author = {Ying'ao Zhao and Guoqiu Fan and Qiaoling Han and Fengjun Chen and Yili Zheng and Jichen Xu and Hao Liang},
title = {Inversion and recognition algorithm for multi-root systems in plain plantation forests based on multi-scale generative adversarial network},
year = {2026},
journal = {Transactions of the Chinese Society of Agricultural Engineering},
volume = {42},
number = {13},
pages = {277-286},
keywords = {ground penetrating radar, root, generative adversarial network, deep learning, multi-target inversion},
url = {https://www.sciopen.com/article/10.11975/j.issn.1002-6819.202511180},
doi = {10.11975/j.issn.1002-6819.202511180},
abstract = {Root systems can play a crucial role in plant individuals. Their parameters are of great significance for the root functions and belowground ecology, including root size, distribution range, and three-dimensional structure. Fortunately, ground penetrating radar (GPR) has been widely used in the non-destructive, high-efficiency geophysical detection of root systems. High positioning accuracy is often required for root systems in underground soil, particularly with complex morphology and significant size differences. Nevertheless, the influencing factors have been unexplored so far, such as the complex structure of tree root systems, the heterogeneity of underground media, direct waves from ground reflection, and the interpretation of GPR data. In this study, a root system inversion was proposed to solve the GPR image interpretation and target detection of tree root systems using a multi-scale convolutional generative adversarial network (MCGAN). The GPR images of root systems were inverted and interpreted to reconstruct the spatial position distribution of root systems in real forest environments. The specific procedures were as follows: 1) A dynamic-weighted multi-scale convolutional layer (DWMC) was introduced into the generator of the MCGAN. The weights of each convolutional layer were dynamically assigned to efficiently extract complex features from GPR B-scan images at different spatial scales. Meanwhile, the adversarial training mechanism between the generator and the discriminator was adopted to generate high-precision structural maps of multi-target root systems. 2) Forward simulation of radar detection was conducted on tree root systems using the gprMax3.0 software. The 4 000 datasets of root system models and their GPR data were established, according to the real structure of tree root systems. A field forest experiment was conducted in Qingping Forest Farm, Gaotang County, Shandong Province, China. Populus tomentosa, a typical plant in the forest area, was selected as the research object. GPR scanning was performed on the root systems using the grid method. Profile excavation was conducted along the detection lines, with 132 sets of real field data. 3) Preprocessing and data augmentation were performed on the field data, including direct wave removal, depth gain adjustment, and median filtering. Noise was eliminated to enhance the reflection intensity of root systems. The indicators were adopted, including the structural similarity index measure (SSIM), mean square error (MSE), and mean absolute error (MAE). The feasibility and accuracy of the root inversion network were evaluated after optimization. The research results show that: 1) The MCGAN achieved the accurate inversion of multi-target root systems, with the SSIM of root inversion reaching 0.874 6, and the MSE and MAE of 137.442 1 and 1.217 9, respectively. 2) Compared with conventional inversion, the MCGAN network more accurately captured the complex features of the small targets at different scales via the multi-scale feature extraction and dynamic weight allocation mechanism, making it more adaptable to the subtle distribution of tree root systems. 3) The DWMC module was used to adaptively assign the feature weights of each convolutional branch, according to the actual feature distribution of the input image. Accurate identification of small-target root systems was realized for the high susceptibility to noise caused by fixed weights. This finding can provide a feasible and effective approach for the inversion of underground multi-target root systems. A great contribution was also gained for the non-destructive detection of the three-dimensional architecture of tree root systems in forest areas.}
}