Towards Interactive Landscape Visualization
Malte Clasen
Dissertation, TU Berlin, 2011
Abstract
In this thesis we present the building blocks of an interactive landscape visualization system focused on terrain and vegetation. First, we describe the data sources in a typical landscape visualization scenario, where terrain elevation and aerial images are exported from a geographic information system (GIS) and enriched with distributions of third-party plant models. This part sets the constraints for the following methods, presented in order of application:
Level of Detail (LoD) generation for the plants has to be done only once for each model, independent of the visualization project at hand. We present a method based on lines and ellipsoids. We leverage the Expectation Maximization algorithm with a Gaussian Mixture Model to create a hierarchy of high-quality leaf clusterings, while the branches are simplified using agglomerative bottom-up clustering to preserve the connectivity. The simplification runs in a preprocessing step and requires no human interaction. For a fly by over and through a scene of 10k trees, the resulting LoD can be rendered on average at 40 ms/frame, up to 6 times faster than billboard clouds with comparable artifacts.
Next we describe how to load and organize the spatial data for a landscape visualization project. We describe a conceptually simple pipeline that handles on-the-fly data decompression and synthesis in a unified process. Possible terrain data sources range from static satellite imagery over per-texel-processing such as image blending routines to light-weight simulations and synthesizers such as noise and filter based texture generators. Point data such as plant instances and polygonal shapes such as building outlines can be handled in the same data structures. The sources are evaluated in parallel based on dependency chains.
Given the resulting terrain textures, we describe a terrain rendering algorithm for spherical terrains based on clipmaps. It leverages the high geometry throughput of current GPU to render large static triangle sets. The vertices are displaced by a height map texture. Our main contribution is mapping of texture coordinates to calculate the height map sample position based on the static vertex offset and the variable view position.
On top of the terrain, we render the preprocessed plant model LoD by raycasting the line and ellipsoid primitives. We extend the ellipsoids by noise textures for alpha-test opacity and normal mapping. This yields a more realistic image, while still avoiding the aliasing artifacts of subpixel-sized primitives. We further show how physically based shading improves the perceived depth complexity.
As a last step rendering step, we postprocess the rendered image to apply deferred shading including shadows and atmospheric scattering.
Following the methods, we describe the resulting landscape visualization system from a user’s perspective. After an overview over the features, we present an evaluation in a case study of cliff erosion for climate change research.
Download