SAGStree: A High-Performance and Highly Realistic 3-D Tree Componentization Method Based on 3DGS

Efficient and realistic 3D tree modeling is an important part of low-altitude remote sensing. Trees have high geometric complexity, and the reconstruction effect of traditional modeling methods is not satisfactory. The 3D Gaussian Splatting (3DGS) method is expected to achieve good results in 3D reconstruction of trees at a low cost. A new method based on 3DGS, SAGStree, is proposed for 3D tree componentization model, which mainly realizes the segmentation and reconstruction of tree trunk, branch, and leaves. Specifically, to suppress high-frequency artifacts and enhance the expression of foreground features, a semantic-guided smoothing filter is introduced in the 3DGS training process. Meanwhile, to improve the segmentation accuracy of each tree components, the Multi-scale Fourier Attention Aggregation Network (MFAANet) is constructed, which includes a Multiscale Adaptive Fourier module, a Pyramid Attention Aggregation module, and a Dual-Path Decoder module. In addition, a multi-view mask voting mechanism is designed to achieve accurate reconstruction of complex tree componentization structures through staged recognition and priority label allocation. This method performs well in multiple task computation evaluations, effectively avoids the problems of detail loss and semantic confusion in traditional modeling, and provides technical support for efficient 3D tree modeling and ecological applications.