Abstract: Currently,large-scale terrain generation methods still rely on traditional mathematical algorithms,which lack user’s control and are difficult to be realized.Although deep learning techniques have been used in terrain generation,the related training datasets are not publicly available and the inherent defects of classical generative networks have not been improved.To obtain more realistic terrain data,a terrain dataset consisting of feature sketches,grayscale images,elevation segmentation images,and elevation models is created.Also,DEM-cGAN framework constrained by feature lines and elevation range is proposed,and a dual-scale parallel network ParallelGen is designed.Users can obtain elevation data with DEM-cGAN by sketching the terrain feature lines and giving the elevation range.Multiple experiments are conducted on the generated results at the visual,numerical,and geographic levels.The results show that,DEM-cGAN framework can correctly generate elevation data with a maximum raster size of 512×512.And the results reproduce the undulation trend in the feature sketch and conform to the realistic geographic laws.