Research On Data-Driven Nature Landscape Generation

The modeling and visualization of natural landscape is the research spot in the field of virtual simulation,which can lead to a higher level of immersion and interaction for the user via accurate and delicate modeling for various elements in natural landscape.The simplification and abstraction of objects in natural landscape are faced with arduous challenges,which are specifically embodied in: the changing weather environment,perennial plate movements,and complex ecosystems during the evolution of natural landscapes.The coupling of above factors results in diverse heterogeneous topologies and texture features for natural landscapes.However,current semi automatic methods are difficult to accurately satisfy user's intention while allowing users to perform detailed modeling for large-scale natural landscapes.With the rapid development of big data and artificial intelligence technologies,and widely available open-source natural landscape data,data-driven methods provide a new solution for natural landscape generation.This paper designs a series of data-driven generation algorithms for a variety of typical natural landscape elements based on cutting-edge procedural modeling and deep learning techniques.The main work of this paper includes:(1)We propose a data-driven multi-style terrain authoring method that allows users to sketch terrains with specific styles.In order to maximize the style differences of generated terrains,this method adopts a generative adversarial network composed of multiple discriminators,and the generator can handle multiple terrain styles(such as mountains,plains)under the guidance of the multi-style discriminators.In addition,the method extracts global-to-local terrain features on multi spatial scales,to mix terrain styles at different spatial scales and support fine-grained editing of terrain style.(2)We propose a terrain texture modeling algorithm based on deep learning to map terrain height map and terrain texture weight map.To supervise the generator,this method adopts differentiable rendering to render terrain scenes with height maps and textures,and brings terrain satellite images to guide the rendered terrain images.The method renders the terrain with random illumination,and introduces an adaptive terrain wrapping algorithm to optimize domain matching between the rendered terrains and the satellite images.(3)We propose an example-based procedural modeling method for river scene,which can control the shape of river scenes with proposed compact parameter model.Meanwhile,we introduce the guidance of meaningful features from exemplar to guarantee the synthesized scene is visually consistent with the input example.In order to make the generated scene as similar as possible to the exemplar,the method designs a measurement metric for evaluation of river scenes,and adopts simulated annealing algorithm to minimize the difference between the generated scene and the exemplar scene.(4)We propose an example-based method to generate the distribution of vegetation on the target terrain,while the distribution is consistent with the style of the exemplar scene.The proposed method extracts the local features of terrain and vegetation distribution with a pre-trained neural network.Then the method mergers the extracted features with patch matching algorithm.Finally,a convolutional neural network is trained to predict the target vegetation distribution.Besides,in order to train the model,a procedural generation algorithm for vegetation distribution is proposed to generate training data.These methods have studied multi-style terrain,terrain texture,river,and vegetation distribution.Users are able to create realistic natural landscapes with little interaction based on above algorithms.A large number of experimental results verify the effectiveness and practicability of the proposed methods.