Under the growing demand for 3D generation in VR, 3D displays, and stereoscopic content creation, sparse-view reconstruction remains challenging due to weak geometric constraints. In such cases, 3D Gaussian Splatting (3DGS) often suffers from floating Gaussians, scale drift, and detail degradation. We propose SPDGS, a progressive reconstruction framework with three key designs. (1) Spatial Consistency–Guided Pruning trains two parallel Gaussian estimators and cross-validates them in 3D space to prune unstable points, reducing floating noise. (2) Volume-Aware Scale Regularization penalizes the product of per-axis scales to suppress volume inflation while preserving anisotropy and sharp boundaries. (3) Patch–Global Structural Consistency Supervision leverages DepthAnything priors to align patch structures and stabilize global scale via normalized targets, without requiring absolute depth. Under sparse-view settings on LLFF and Mip-NeRF360, both quantitative and qualitative results, along with ablation studies, demonstrate that the three modules are complementary and achieve significant synergy. Overall, our method yields more compact Gaussian fields, cleaner geometry, sharper edges, and fewer artifacts, while maintaining structural stability and detail fidelity even with very limited views.