Satellite stereo pipeline for automatic large-range digital surface model generation in water areas

Digital surface model (DSM) construction based on satellite images is one of the important contents of photogrammetry with many applications. The popular pipeline does not cover a large number of production problems, or it may not be suitable for advanced deep learning algorithms, especially in texture-less water areas. There is currently a lack of a complete automated satellite stereo process pipeline that incorporates advanced technology. In this paper, we re-examine the process of digital surface model construction based on satellite images, and reconstruct the existing process. On the one hand, we add the extraction of automatic control points into the conventional stereo vision pipeline and we provide a framework for deploying deep learning stereo matching models for large-scale images. On the other hand, we increase the optimization steps based on DOM, and make rational use of the current rapidly developing monocular vision model to solve the problem of poor matching quality of traditional models and deep learning-based models in large-scale textureless waters. Experiments on different areas demonstrate that our method is accuracy and robustness. Our system integrate advanced stereo matching algorithms and can handle complex real-world production problems , which helps promote better development in the fields of satellite photogrammetry and remote sensing.