Realism in Data-Based Terrain Synthesis
The landscapes on Earth are varied and complex, having been created by innumerous physical processes over millions of years. The creation of artificial terrain that replicates the realism of landscapes on Earth has been a major challenge for computer graphics. Many different approaches have been taken, including approximating the terrain with fractals and splines, simulating the terrain using models from the physical geography, and reconstructing terrain from elements of real-world data. A primary issue in the field of terrain synthesis is the lack of, and evaluation of, realism in synthesized terrain. This thesis identifies and discusses the flaws of existing data-based methods based on example-based texture synthesis methods. It provides improvements to an existing data-based method using algorithms from the field of geographic information science, and presents a novel algorithm, ``terrain-optimization'', based on the example-based texture synthesis technique of texture-optimization. Finally, it discusses a new approach to the experimental evaluation of terrain realism, with the largest experiment conducted to date. The results of this show that each of the tested methods is indistinguishable from reality in certain circumstances and that those circumstances differ for each method tested, and that subjects with a high level of expertise in physical geography are the most qualified for identifying real terrain from synthesized terrain. Overall, the thesis provides substantial analysis and evidence about the challenges of data-based terrain synthesis while also developing new approaches in the field that perform as well as existing state-of-the-art methods.