Complex forest structures, interspecies similarities, and intraspecies variations constrain the acquisition of species-specific tree phenotypes. This study develops a scalable framework for extracting species-specific structural parameters at the individual tree level. Leveraging ultrahigh-resolution UAV-based RGB and LiDAR data, we propose a novel self-attention-guided spectral–structural multimodal fusion transformer (SAMFormer). Key components include: (1) an adaptive feature enhancement module (AFEM) that employs spatial and channel attention to selectively highlight canopy features while suppressing background noise; (2) a cross-modal fusion module (CMFM) that captures intra- and inter-modal dependencies through the cross-attention mechanism, generating highly discriminative representations. SAMFormer achieves fine-grained tree identification in complex forest environments, relieving issues of blurred canopy segmentation and species misclassification. K-fold cross-validation demonstrates robust performance across diverse scenes, achieving 86.3 % F1-score and 88.0 % mAP@0.5, significantly outperforming single-modal inputs and mainstream instance segmentation models. We generate large-scale species-specific maps of tree structural parameters based on SAMFormer outputs, allometric equations, and a sliding window strategy. Subsequently, these parameters are utilized to map carbon stock. Ecological analysis reveals a coupling relationship between tree competition and structural parameters/carbon stock: competition intensity exhibits a significant negative correlation with both (p < 0.001). Trees adapt by adjusting growth strategies (e.g., reducing radial growth and limiting canopy expansion), ultimately lowering biomass accumulation and carbon stock. Additionally, species mixing enhances carbon stock, as mixed forests store more carbon than monocultures. This work provides a high-throughput, non-destructive pathway for forest phenotyping, supporting precision forestry and climate-adaptive management practices.

Accurate acquisition of forest spatial competition and tree 3D structural phenotype parameters is crucial for exploring tree-environment interactions. However, due to the occlusion between tree crowns, current UAV-based and ground-based LiDAR struggles to capture complete crown information in dense stands, making parameter extraction challenging such as maximum crown width height (HMCW). This study proposes a canopy spatial relationship-based method for constructing forest spatial structure units and employs five ensemble learning techniques to train 11 machine learning model combinations. By coupling spatial competition with phenotype parameters, the study identifies the optimal fitting model for HMCW of Chinese fir. The results demonstrate that the constructed spatial structure units align closely with existing research while addressing issues of incorrectly selected or omitted neighboring trees. Among the 10,191 trained HMCW models, the Bagging model integrating XGBoost, Random Forest (RF), Support Vector Regression (SVR), Gradient Boosting (GB), and Ridge exhibited the best performance. Compared to the best single model (RF), the Bagging model achieved improved accuracy (R² ​= ​0.8346, representing a 1.6 ​% improvement; RMSE ​= ​1.4042, reduced by 6.66 ​%; EVS ​= ​0.8389; MAE ​= ​0.9129; MAPE ​= ​0.0508; and MedAE ​= ​0.5076, with corresponding improvements of 1.63 ​%, 1.49 ​%, 0.1 ​%, and 7.06 ​%, respectively). This study provides a viable solution for modeling HMCW in all species with similar structural characteristics and offers a method for extracting other hard-to-measure parameters. The refined spatial structure units better link 3D structural phenotypes with environmental factors. This approach aids in canopy morphology simulation and forest management research.