Before rendering terrain, a corresponding volumetric terrain model is established from the elevation data and the aerial photographs or satellite images of a terrain, so that each voxel of the volume data set under the terrain surface has a certain color value. Levels of Detail (LODs) can be achieved by appropriately selecting pixels along each dimension of the terrain. For example, by selecting every other pixel, we generate a volume with less detail, which has a resolution of half the original in each dimension. More complex filtering methods can be used to meet higher requirements.

The usefulness of volume rendering for visualizing the large amounts of data generated in a variety of scientific disciplines has been widely demonstrated; however, the computational expense of this technique limits its routine and interactive use, especially in applications dealing with large volume data, such as digital terrain modeling and visual simulation. Therefore, parallel volume rendering algorithm is used on a 16-processor MIPS R10000(194MHZ) based Power Challenge IR graphics system, in order to achieve real-time fly simulation.

Our parallel algorithm is based on the serial ray casting method: Rays are cast from the viewing position through the volume data. The data is resampled at evenly spaced locations along each ray by trilinearly interpolating values of surrending voxels. Finally, ray samples are composited to produce an image. A key advantage of ray casting over other volume resampling techniques is that algorithmic optimizations have been developed, such as early ray termination and spatially adaptive image sampling, which significantly reduce its image generation time. Furthermore, different LODs are selected during rendering stage according to several considerations, including the distance to the viewing point, and the position in the image(screen). The smoothness between continuous frames and avoidance of cracks between patches with different LODs in each frame are taking into account.

A MIMD machine like the Challenge is suitable to our parallel processing, since an intricate serial algorithm with many special cases, due to optimizations, does not map well onto SIMD architectures. Image based task partitioning is more efficient and can be more easily implemented in Challenge's shared memory. At present, only scanline base image partitioning is explored with 16 processors by revising the existing serial algorithm in VolVis. More complex and efficient partitioning methods will be studied, and a separate parallel renderer will be established to approach the real-time fly-through. Furthermore, different voxolized geometric objects like vehicles and airplanes will be added into the resultant scene with different LODs, according to their distance and moving speed.