Abstract: Objectives: Interferometric Synthetic Aperture Radar (InSAR) and optical photogrammetry are two primary techniques for generating Digital Surface Models (DSMs). However, both techniques have inherent limitations that affect the quality of DSMs. The fusion of multi-source DSMs can effectively leverage the complementary advantages of the data to enhance the quality of DSMs derived from a single data source. Traditional weighted average fusion methods rely on prior information for weight allocation and do not adequately consider the impact of terrain features on elevation errors. There is a need to develop a new fusion method. Methods: A stepwise regression fusion method for InSAR and optical DSMs that takes into account topographic features is proposed. Using InSAR and optical DSMs as research subjects, and considering the commonalities and differences in error characteristics between InSAR and optical DSMs, terrain geometric features, land cover features, and elevation difference features are extracted as training samples in the training area. Subsequently, the stepwise regression method is employed to construct a regression model between elevation errors and terrain features by incrementally adding or removing variables. Elevation errors are then predicted to adaptively determine the fusion weights. Finally, fusion experiments were conducted using TanDEM-X DSM and AW3D DSM data from mountainous areas in the United States, and the results were tested in two different experimental zones.. Results: A stepwise regression fusion method that takes terrain features into account is proposed. Using InSAR and optical DSMs as research subjects, and considering the commonalities and differences in error characteristics between InSAR and optical DSMs, terrain geometric features, land cover features, and elevation difference features are extracted as training samples in the training area. Subsequently, the stepwise regression method is employed to construct a regression model between elevation errors and terrain features by incrementally adding or removing variables. Elevation errors are then predicted to adaptively determine the fusion weights. Finally, fusion experiments are conducted in complex mountainous areas for validation. Conclusions: The method effectively utilizes terrain features to enhance the quality of DSMs and performs well in complex terrain. It provides a solution for multi-source DSM fusion that balances accuracy and interpretability.