This study proposes a three-stage monocular Simultaneous Localization and Mapping (SLAM) method that aims at enhancing the fine-grained mapping performance under dynamic environments. In the uncertainty prediction stage, the proposed method concatenates semantic features extracted from CLIP-DINOiser with 3D-aware DINOv2 features from WildGS-SLAM, a shallow MLP is employed to learn semantic-dynamic correlations, which generates a semantic-aware uncertainty map and weights the uncertainty map using the processed per-pixel dynamic probability map. In the tracking stage, the adjusted uncertainty map is employed as weights to carry out dense bundle adjustment, realizing the examination of dynamic distractors. In the mapping stage, a semantic consistency loss function is established to improve 3D Gaussian Splatting reconstruction of static regions and suppress the generation and expansion of Gaussian points in dynamic regions. Simulation results demonstrate that the proposed method achieves superior mapping results compared to state-of-the-art dynamic SLAM algorithm under challenging dynamic scenes.