3D scene semantic understanding is a fundamental challenge in computer vision, aiming to accurately recognize and segment three-dimensional spatial structures. With the advancement of applications such as autonomous driving and robotic navigation, this task faces increasingly stringent requirements.The recent introduction of 3D Gaussian Splatting has significantly improved reconstruction efficiency by several orders of magnitude while maintaining comparable rendering quality to baseline methods. However, semantic decoupling under the 3DGS framework remains insufficiently explored.Given the high complexity and limited availability of annotated 3D datasets, directly training networks for 3D semantic understanding is challenging. To address this, we propose a method that leverages 2D semantic priors to encode low-dimensional semantic fields, extracting prior knowledge via a pre-trained 2D segmentation network and training with differentiable volume rendering.A dynamic thresholding strategy enables coarse semantic segmentation, followed by statistical denoising and recalibration. Furthermore, a decoupled semantic binding approach is introduced for independent parameter control.Extensive experiments show that our method achieves baseline performance within seconds of optimization and can be seamlessly integrated into downstream tasks such as scene editing.