Go to Paper
Return to GeoComputation 99 Index
Texture Information and Supervised Classification of Hyperspectral Imagery by Means of Neural Networks
BOSCH, Edward H. (firstname.lastname@example.org), U.S. Army Topographic Engineering Center, 7701 Telegraph Road, Alexandria, VA 22315-3864
Key Words: neural networks, hyperspectral imagery, supervised classification, texture measures, spatial information
In numerous studies, including those of the U.S. Army Topographic Engineering Center (TEC), researchers have used the spectral dimension of hyperspectral imagery for the purpose of classifying such data sets. In some cases, hyperspectral data sets are made up of hundreds of narrow spectral bands. The totality of the useful spectral bands are treated as a high-dimensional vector, which may be modeled by means of different mathematical and statistical methods. In the case of a supervised classifier, a very simple and often used method consists in obtaining the spectral mean for each of the training classes, and then computing the angle between the spectra in question and each of the corresponding spectral means. Since the spectra is being treated as a vector, the above classifier literally corresponds to the inverse cosine of the inner product of two vectors. The sample being classified is assigned to the class associated with the smallest angle. Each spectral sample (vector) also is normalized.
At TEC, we have developed and extensively employed a version of the back propagation algorithm to obtain neuro-spectral models, which are used to classify hyperspectral imagery. In this study, we investigate the potential of neural networks to discriminate texture information using hyperspectral data. The data associated with texture will be obtained from one of the bands of the hyperspectral data set and will be used in both the training and classification phases. Since texture is related to spatial information, one of the questions that needs to be addressed is the pixel size of the search window (kernel).
Clearly, if the size of the window is just one pixel in length and width, the amount of data in this window is not sufficient to be considered texture information. Nevertheless, this case is not unreasonable if we use the remaining spectral bands to perform the analysis; however, this corresponds to the case of classifying hyperspectral imagery with spectral data as opposed to spatial data. Conversely, if the size of the window is too large, the network may not be able to correctly classify the texture data when there is more than one type of texture within the window. Although this situation is inevitable, a way to circumvent this problem may be provided by generating a texture class, which is a combination of at least two textures.
Since contiguous hyperspectral bands are so highly correlated, we will determine how well a neural network, obtained from one band, classifies such contiguous bands. This will help us further analyze surfaces whose textures are similar but whose mean amplitudes are different.