Virtual Terrain Generation Based On Deep Learning

Virtual terrain is a kind of virtual geographic scene constructed based on limited information under the guidance of geospatial cognitive knowledge.It is widely used in geoscience analysis,game and movie scene production,virtual reality,and military battlefield simulation.When investigating the battlefield environment in some hard-to-reach areas,the rapid construction of accurate electronic terrain sandboxes based on limited information such as mountains and valleys in the field of view plays an extremely important auxiliary role for strategic deployment and command.Therefore,it is necessary to study the method of constructing virtual terrain according to limited information.This paper proposes a deep neural network based on Conditional Generative Adversarial Nets named Terrain-CGANs,which was trained on the dataset containing existing high resolution DEMs and terrain features extracted from them.The control effects of different terrain feature in the process of generating virtual terrain are analyzed,and a detailed virtual terrain quality evaluation method was developed.Finally,the practical application ability of Terrain-CGANs was discussed in three practical application cases.The main contents and conclusions of this paper are as follows:(1)A virtual terrain generation model Terrain-CGANs based on conditional generation confrontation network was proposed.Terrain-CGANs,based on the Conditional Generative Adversarial Nets,designed for generating virtual terrain is trained using the terrain feature dataset,which realizes the virtual terrain with accurate terrain features and rich surface details by inputting a small number of terrain features.(2)The influences of terrain feature on the results of virtual terrain generation were investigated.In this study,five key experiments,with Terrain-CGANs being trained on different combination of valley line,ridge line,positive topography areas and these topographic features,were designed to probe into the way that different topographic features affect the generated virtual terrain.The results show that the input topographic feature controls the characteristics of the skeleton of the virtual terrain and the characteristics of the regional shape.The comprehensive terrain features can be used to make the virtual terrain accurately express the comprehensive terrain features and surface details.(3)A virtual terrain quality evaluation method was proposed.According to the application scenarios of virtual terrain,this paper designs a set of methods that evaluate the quality of virtual terrain from three aspects:visual aesthetics,terrain feature preservation and terrain factor extraction.The results prove that the virtual terrain generated by Terrain-CGANs has a good visual simulation and basically maintains the input topographic features and can be used to extract terrain feature factors and perform simple terrain analysis.(4)The flexibility and stability of Terrain-CGANs was tested in typical application cases.The paper analyzes the performance of Terrain-CGANs in three typical virtual terrain application scenarios.By editing the topographic features,the local features of the virtual terrain and the adjustment of the global topography are realized.By designing the topographic features of different erosion stages,the geomorphic erosion process is simulated.The results prove that the Terrain-CGANs have certain flexibility and stability.The method generating virtual terrain with terrain features reveals the important role of terrain features in landforms from a new perspective.The generated virtual terrain has correct geospatial structure and rich detailed surface.This paper provides a method reference for virtual geographic scene construction and terrain surface interpolation,and provides a basis for studying the control effect of feature terrain in terrain.The virtual terrain generation method based on deep learning proposed in this paper expands the application of deep learning method in geosciences and enriches the method system of digital terrain analysis.