Digital Elevation Models (DEMs) are extensively utilized for terrain analysis, representation, and visualization. Various application scenarios require DEMs at different Level of Details (LODs). Although traditional multi-scale landform expression methods primarily target DEMs at small scales, the simplifying Digital Surface Models (DSMs) for urban modeling at large scales remains a significant challenge. By integrating map generalization theory with computer vision techniques, we developed a novel method for urban building simplification in DSMs, termed Building Simplification in Digital Surface Models (BS-DSM). First, the buildings extracted from the DSMs are subjected to morphological analysis and aligned to a specific orientation. Next, the DSMs are divided into rectangular pixel blocks through energy-driven sampling, followed by horizontal simplification of building shapes in a 2D projection plane according to the geometric characteristics of these pixel blocks. Finally, to preserve the average height and total volume of the simplified 3D buildings in the vertical direction, the building heights across different pixel blocks are adjusted and interpolated based on the skeleton lines of building roofs and calculations of adjacent height values. The proposed BS-DSM method was evaluated using the publicly available Vaihingen DSM dataset. The result shows that the BS-DSM method performs better in simplifying building shape and height while meeting basic multi-scale expression constraints compared with traditional filtering methods.