Efficiently extracting discriminative local features and exploiting long-range spatial correlations remain critical challenges in point cloud semantic segmentation. Existing methods often struggle to balance complex surface topology resolution with computational efficiency, frequently ignoring critical spatial correlations during training. To address these limitations, this paper proposes the Multi-Scale Umbrella Feature Network with Spatial Self-Attention (MUSF-SSA) to efficiently encode irregular 3D surface geometry. MUSF-SSA employs a k-d tree search for rapid neighboring point identification, followed by multi-scale umbrella feature construction to integrate key information across various spatial scales. An encoder-decoder structure is introduced to further refine these features, while a spatial self-attention mechanism explicitly models long-range spatial correlations. Evaluations on the S3DIS dataset demonstrate that MUSF-SSA achieves a mean Intersection over Union (mIoU) of 74.8%, a mean Accuracy (mAcc) of 82.9%, and an Overall Accuracy (OA) of 91.2%, surpassing state-of-the-art (SOTA) methods with comparable parameter complexity.