What is texture? The GLCM Texture Calculations More Information and References
Some practical considerations
Textures may be calculated using only one channel of data at a time.
Many different texture measures can be calculated for each image channel
- Channel information can be consolidated using an index, principal components or other method before running a texture measure.
- The same texture measures can be calculated for more than one channel
- Three texture measures can be viewed simultaneously using rgb projection.
- It is possible to take the principal components of many texture images, but the result is very difficult to interpret.
If attempting it, it is useful to scale all output texture values to the same range of GLs before calculating the eigenvectors, so that one measure will not dominate the first components simply because of having a greater dynamic range. One way to do this is to transform each texture measure to its z score before entering in the principal component calculation. This is called "standardised principal components." Some software will allow you to choose to perform the principal components transformation using the correlation matrix instead of the covariance matrix: this accomplishes the same thing.
Texture images are image channels with a value for each pixel.
- Texture images can be used alone or with other data to define signatures for supervised or unsupervised classification.
- Texture images (channels) can be included in classifications.
- Texture output should be put into a 32R channel, at least at first. It may then be scaled into an 8-bit channel if desired.
- The actual numerical value of a texture measure is unimportant for most purposes. More important is whether it is relatively high or relatively low, compared to the same texture measure values elsewhere on the image. Despite this, when textures of different images are to be compared, the texture measure should be calculated using the same software for both images and the same scaling method if applicable.
Textures are calculated within a window, a small region of the image.
The test image used in this tutorial considers the entire image to be the area contributing to the texture. For larger images, a window is chosen to define this area. This window is, in practice, square and with odd numbered side lengths. In theory, a window can be any dimension, but again, practical calculation problems occur for even sizes and non square shapes of windows.
- The relative size of the window and the objects in the image will determine the usefulness of the texture measure for classification.
- It is expected that different objects will have different characteristic texture measures. To capture this, the window must be smaller than the object (h-resolution), but big enough to include the characteristic variability of the object
- Windows, like those used with filters, always have an "edge effect" where the window overlaps the border between distinct objects on the image. More information about window edges.
Example: The texture of a closed-canopy forest is determined by light and shadow among tree crowns. A window covering one tree will not measure the texture of the forest. A window covering the entire forest and the fields next to it will not measure the texture of the forest either.
Image Examples:
To see the edge effect of window size, compare the images linked below. They are all entropy calculations, and use the
same parameters, except to vary window size.
