Abstract: The 3D terrain plays an important role in constructing various virtual outdoor scenes, but it is difficult for the existing terrain synthesis algorithms to automatically create realistic and predictable terrains. In this paper, we present a novel algorithm for conveniently and intuitively synthesizing 3D digital terrains. The algorithm first extracts a set of mountain outlines from a given image, and evaluates the depths of the outlines with the deep convolutional neural network model and creates a depth image. The outlines are then mapped to 3D space via the inverse process of planar imaging of 3D objects. Subsequently, based on the mapped outlines in 3D space, our method constructs a feature sketch to define global features of the terrain to be created. Taking a group of terrain blocks with local features, sampled from real terrains, as basic constructing elements, our method finally synthesizes the terrain by matching and blending the blocks guided by the feature sketch. Experiment results showed that, our algorithm can automatically synthesize realistic 3D terrains with various features, which meet user’s expectations well under the constraint of real mountain outlines.