Objective To propose an improved segmentation model that integrates frequency domain information, extract root phenotypic parameters, and address difficulty of extracting roots of Betula luminifera due to their complex morphology, fine structure, and strong background interference.

Method First, using TransUNet as the baseline model, a frequency domain information aggregation (FDIA) module was designed and used for downsampling at each layer. This module included frequency domain information and the small-scale feature guidance (SSFG) mechanism to improve the preservation ability of root contour boundaries and detailed features. Second, a lightweight cascaded attention mechanism was added to the Transformer layer to enhance the recognition of complex root topology. Finally, an improved dynamic upsampling method was introduced in the decoder to improve the reconstruction accuracy of fine-root regions.

Result On the root image dataset of tissue culture seedlings of B. luminifera, the proposed FED-Net model achieved segmentation metrics of 79.68% and 89.94% for root category intersection over union (IoU) and F₁-score, respectively, representing improvements of 3.74 and 3.87 percentage points compared with the TransUNet baseline model, while reducing the number of parameters to 21.664 × 10⁶. Further analysis of the segmentation results extracted four phenotypic parameters: Total root length, total root projected area, contour perimeter, and average root diameter. Their RMSE values were 18.276 mm, 9.229 mm², 24.612 mm, and 0.047 mm, respectively, and all R² values exceeded 0.8, meeting the accuracy requirements for root phenotype studies.

Conclusion The proposed model has good segmentation accuracy and phenotypic extraction capabilities, providing a reliable technical support for the analysis of the merits and demerits of B. luminifera families.