I am not sure if you mean that including texture feature with known information such as spectral signatures will train the algorithm better to separate very high spatial resolution image into homogeneous classes. if this is the case you must consider many issues such as the method whether it is object oriented or pixel based method, whether the image is noisy or not, what is intended to extract as information...etc. Sometime it may be useful other times it may not depending how you answer the questions before.

Textures add to the amount of information available in analysing the imagery. Other measures, based on first-order statistics (e.g. mean, X-order moments, etc,) will look at very local variations. Textures are computed over larger windows, and can tap into the richness of data variations. Using second-order statistics, they can detect structures invisible to the human eye, because of physiological constraints (there were very good articles showing that in practice, in 1970s classics like the book "Psychopictorics" or in articles by pioneers like Haralick and others). Mryka Hall-Beyer already answers the question in considerable detail, and points at GLCMs. Which parameters to use, out of the 20 or so available in the literature, will depend on the applications. I started with GLCMs using microwave imagery from Venus, and carried on with sonar imagery from the seabed: for each type, I use different parameters.

The pitfalls to avoid are the size of the areas over which textures are computed. Too small, and one looks at high-frequency noise in the images (speckle). Too large, and one looks at general variations, but not at structures of interest. The other parameters, when using GLCMs, are the displacement within the computation window and the direction in which to look. I would personally advocate averaging over all directions (if only to avoid directionality effects when acquiring images from different angles) and check different displacements (generally, the optimal ones are related to the scale of the structures of interest). Used with caution, textural signatures are a great asset to classification, improving the accuracies by looking at subtle changes in the overall signatures. And using multispectral data, this should be even better (sigh ... we do not often have multispectral data in sonar imagery :-)

Thanks Phillippe. As mentioned by many, the efforts involved in using texture for classification ( though it adds some more information ) is significantly higher than using general classification techniques w.r.t. remote sensing imagery. Most of the times it is the scene and the nature of features which user should be aware and take a decision to go for inclusion of texture parameters.

When i try to automate the procedure ( with minimal user parameters ) its becoming highly involved. Is there any software / library in open source that has implemented these for production mode environment for large scale ( high resolution ) multispectral imagery ?

Most of the texture-calculation equations are open-source, but specific software like Matlab also have their own versions (e.g. to calculate entropy), based on the same articles. I am sure other software do have them now as well.

Will depend on the features of interest and your classification system, but the use of texture to improve classification accuracy is well established in the RS lit.