As an important terrain representation model, DEM usually needs to be generalized at multi-resolution adaptable to different level applications. DEM data directly represents the elevation of individual grid and indirectly depicts the hidden landform features such as valley, ridge, saddle, pit and peak. These landform features play significant roles in terrain generalization. Traditionally DEM generalization is just based on image processing by resample or low-pass frequency analysis methods without considering the distribution of landform features. This study develops a method to generalize DEM data based on the simplification of valley system. In this structured generalization approach, the valley information is firstly extracted and then based on the characteristics of valley catchment, the DEM generalization is conducted by two steps: geographic decision and geometric operation. According to the hydrological significance, unimportant valley branches are detected and the corresponding coverage is filled by raising the terrain to make the terrain surface smooth. In contrast to the conventional algorithms based on the image processing, this method is able to preserve the main geographical characteristics more effectively in terrain representation.