In order to effectively improve the accuracy and precision of 3D real scene model reflecting various types of geographic features to fully support the 3D real scene construction in China, this paper systematically classified the model defects existing in the current 3D models reconstructed by UAV tilt photography into four types and analyzed their causes. Subsequently, three key technologies were proposed: First, interactive suspension removal based on connectivity and Euclidean distance clustering, using graph cut theory and random consistency sampling algorithm to identify and remove suspensions; Second, water surface semi-automatic leveling based on graph cut theory, using Grabcut algorithm to extract water surface area while giving it a uniform elevation to achieve water surface leveling; Third, water texture inpainting based on deep network model, using convolutional neural network and deep adversarial training to obtain the characteristics and structure of water texture to achieve repair of missing water texture. Finally, practical production data from Yichang City, Hubei Province was used for experiments. The experimental results showed that the key technologies proposed in this article could effectively repair 3D model defects, thereby enabling 3D real scene model to have more refined geographic entities and geographic landscape restoration.