A Review of Research on Dense Matching Algorithms in Digital Surface Model

Digital Surface Models (DSM), critical for 3D surface representation, rely on dense matching algorithms for accuracy and efficiency. This review examines two decades of advancements in feature-based, region-based, and deep learning-driven methods. Feature-based methods such as SIFT and ORB can achieve sub-pixel accuracy in high - texture scenes. However, they have a mismatch rate of approximately 20% in low - texture areas and are suitable for small - scale photogrammetry. Region - based methods like Semi - Global Matching (SGM) can achieve a Root Mean Square Error (RMSE) of ≤0.5 meters in homogeneous terrains. But in complex urban scenes, they may have errors of about 1.2 meters. These methods are used for large - scale DSM generation and have a computational complexity of O(n2). Deep learning-driven methods such as GC-Net can reduce the mismatch rate by 30–50% in low-texture regions, with F1 - scores greater than 0.9. However, they require 20–50 times more GPU memory and are applied to high - precision DSM in complex environments. Currently, the challenges include the trade-off between accuracy and efficiency and the interpretability of deep learning. Future directions include AI-driven interdisciplinary integration, enhanced data augmentation, and addressing complex scene challenges.