This paper proposes a new method for automatic co-registration of digital elevation models (DEMs) based on centroids of subwatersheds. Subwatersheds are stable physical features, making their centroids more reliable and accurate as control points (CPs) than the other features. In the present method, subwatersheds are delineated from DEMs using hydrological analysis procedures. Modified invariant moments are employed to measure the similarity of subwatersheds for determining the correspondences between the reference and input. Centroids of matched subwatersheds are then derived as CP candidates, where the root-mean-square error is applied to eliminate mismatches using a global consistency check method. The established CP pairs are used to estimate parameters of a 3-D conformal transformation model, which is employed to rectify the input DEM. The accuracy of CP detection was assessed using two DEM subsets of different terrains. The results of six tests showed that the maximum shift and rotation errors were about 2 m (1/45 pixel) and 0.006°. The developed approach was used to co-register the Advanced Spaceborne Thermal Emission and Reflection Radiometer Global DEM to the Shuttle Radar Topography Mission DEM at three locations. The results revealed that both the mean absolute error and the standard deviation of the elevation differences were reduced for all the tests after co-registration, showing the good performance of the proposed method. Comparisons have also been made against previous works, which suggested that our results were consistent with the previous studies.