We developed a modeling and rendering environment that meets the goals and considerations described above (Figure 3). Input consists of terrain elevation data and orthophoto data. These two sources of information are used as input by a pattern classification system to categorize each pixel in the image into one of a set of possible classes (e.g., rock, snow, tree, etc.). The classifier is also capable of identifying shadowed regions in the original image, together with unshadowed areas of the same class. This is sufficient to enable a deshadowing process in which shadowed areas of a given type are made to be visually similar to the corresponding unshadowed regions. The classification is also used to develop explicit geometric models for larger objects such as trees. Finally, a physically-based snow simulation is used to compute snow cover for a given date. All of these are combined to form a terrain skin covered by a colored texture map and a set of geospecific vegetation models.
The major obstacles that must be crossed to create such a system include:
