RSTGS: Relightable Single-Tree Modeling with 3D Gaussian Splatting from Multi-View Images

High-fidelity relighting models of trees are crucial for real-scene 3D GIS and digital-twin geographic environments, where the realism requirements are substantially higher than those of conventional 3D modeling. However, the intricate structural complexity of trees often leads to high computational costs due to geometric reconstruction, and background interference in multi-view images can decrease reconstruction accuracy and illumination modeling. To address these issues, a novel approach, RSTGS: Relightable Single-Tree Modeling with 3D Gaussian Splatting from Multi-View Images, is proposed. RSTGS incorporates a parent–child Gaussian field supervision regularization mechanism to improve training stability while promoting multi-scale structural consistency through cross-level geometric and radiometric constraints. Additionally, a Gaussian density adaptive control mechanism driven by dual-domain gradient consistency is also proposed, which can simultaneously constrain the spatial and normal domains to preserve geometric details while reducing redundant points. For illumination modeling, a geometry-aware incident illumination estimation method is employed, which can infer dominant light directions and construct normal-aligned incident light distributions to achieve physically consistent lighting and realistic shadow effects. The results of experiments on three types of trees with distinct characteristics indicated that RSTGS has significant advantages compared to SOTA approaches in novel view synthesis and relighting tasks for models of real trees, achieving good results in PSNR, LPIPS, SSIM, and visualization quality.