The saddle point is one of the important terrain feature points that reflects the fluctuation of the surface morphology. Accurate extraction of saddle points is beneficial to the analysis of spatial relationship and structural feature of terrain. Existing saddle point extraction methods are usually directly based on regular grid DEMs data which not have topological relation, unable to take into account the spatial topological relationship between saddle point and surrounding terrain and the influence of complex terrain on them. Many false feature points which are not in the saddle region will be extracted, and saddle points in some key areas also will be ignored. Thus, this paper designs an algorithm for extracting saddle points based on contour data according to the morphological characteristic of saddle points. This algorithm makes use of the feature of contour closure, converts contour lines into contour polygons data according to certain rules, and then uses the adjacent topological relations between contour polygons to recursively search and automatically extract the saddle-points. The experimental results show: (1) The number and position of saddle points are significantly related to the size of the contour interval. Within a certain range of scales, the smaller the contour interval is, the more saddle points are extracted, and the position accuracy is gradually improved. (2) Compared with the method based on regular grid DEMs data extraction, this method can more effectively filter many false feature points which are not in the saddle region, improve the accuracy of saddle point extraction, and also reduce the complexity of the saddle point extraction algorithm. In addition, DEM scale is selective for saddle point extraction. Compared with incremental contour method to extract saddle points, the idea is to find the intersection point which is the saddle point by extending the adjacent pseudo-saddle point contour line to the junction, this way has the advantage of more accurate position. However, this method requires DEM data with high precision, and when applied to 5-meter spatial resolution DEM data, most saddle points can't be extracted. The method in this paper is to use the midpoint of the nearest point in the two peaks as the saddle point, whose point position is not completely accurate, resulting in a certain error in spatial position accuracy. However, it can identify the majority of saddle points effectively, which has an advantage over the method of extracting saddle points mentioned above. Thus, the method in this paper can effectively improve the integrity of saddle point extraction and is more suitable for the DEM scale data (5 m DEM) in this paper.