While the aerial imagery used to generate orthoimages is chosen to minimize shadows, shadowing is still present. As would be expected, the severity of this problem increases with the ruggedness of the terrain. These shadows need to be removed and replaced by simulated shadows resulting from a different illumination direction if we want to use the imagery to texture a terrain scene for a date/time different from when the source image was taken. Given accurate information about the direction of incident direct illumination and high resolution elevation data, expected shadow locations could be easily computed. In practice, we seldom have either and so another approach is needed.
The maximum likelihood classifier does a good job of identifying shadow areas and can even categorize different surface covers within the shadowed regions. This can be used to remove the photometric effects of shadowing, even when the direction of illumination is not known. For purposes of visual rendering, it is enough to renormalize shadowed portions of the orthoimage to have a brightness distribution statistically similar to unshadowed regions of the same surface type. In practice, it appears to be enough to standardize the mean and standard deviations of the shadowed regions. To improve the visual qualities of our talus texture, we apply image processing. The real talus exhibits shadowing and highlights from individual stones. Most of this is lost in the blurring of the photographics process. To add some shading variation, we apply two steps. First, we apply mean-variance normalization. Because the dynamic range in the shadows is so low, the variance normalization effectively translates quantization noise into uncorrelated additive noise. We then apply a gaussian blur, thus adding correlation to the noise for an appearance that more resembles the mid-size rocks of talus.
Often, shadow boundaries in orthoimages exhibit a penumbra-like effect, though at a scale much larger than the shadow penumbra that would be generated by a light source the angular extent of the sun. The causes of this phenomenon are not clear, but are likely due to a combination of interreflection, variations in sky brightness, and photographic dodging done as a pre-processing step in orthophoto preparation. Whatever the cause, these shadow fringes are visually distracting and can generate ground type misclassifications. Fortunately, it is an easy matter to replace dark pixels near classified shadow regions with lighter pixels slightly farther away, largely eliminating the problem.
Figure 6 shows the results of deshading and shadow removal applied to Figure 4 and followed by coloring based on the classification shown in Figure 5.
