Accurate segmentation of fine roots in field rhizotron imagery is essential for high-throughput root system analysis but remains challenging due to limitations of traditional methods. Traditional methods for root quantification (e.g., soil coring, manual counting) are labor-intensive, subjective, and low-throughput. These limitations are exacerbated in in situ rhizotron imaging, where variable field conditions introduce noise and complex soil backgrounds. To address these challenges, this study develops an advanced deep learning framework for automated segmentation. We propose an improved U-shaped Convolutional Neural Network (U-Net) architecture optimized for segmenting larch (Larix olgensis) fine roots under heterogeneous field conditions, integrating both in situ rhizotron imagery and open-source multi-species minirhizotron datasets. Our approach integrates (1) a Convolutional Block Attention Module (CBAM) to enhance feature representation for fine-root detection; (2) an additive feature fusion strategy (UpAdd) during decoding to preserve morphological details, particularly in low-contrast regions; and (3) a transfer learning protocol to enable robust cross-species generalization. Our model achieves state-of-the-art performance with a mean intersection over union (mIoU) of 70.18%, mean Recall of 86.72%, and mean Precision of 75.89%—significantly outperforming PSPNet, SegNet, and DeepLabV3+ by 13.61%, 13.96%, and 13.27% in mIoU, respectively. Transfer learning further elevates root-specific metrics, yielding absolute gains of +0.47% IoU, +0.59% Precision, and +0.35% F1-score. The improved U-Net segmentation demonstrated strong agreement with the manual method for quantifying fine-root length, particularly for third-order roots, though optimization of lower-order root identification is required to enhance overall accuracy. This work provides a scalable approach for advancing automated root phenotyping and belowground ecological research.