Spatial autocorrelation analysis of hyperspectral imagery for feature selection

The spatial information in a single spectral image can be estimated from the spatial autocorrelation, which is a measure of how the local variation compares with the overall variance in a scene. In images of random noise, the local variation tends to be similar to the overall variance. In contrast, scenes in which large features can be discerned have clusters of pixels with similar values, which cause the local variation to be much smaller on average than the overall scene variance. A comparison of the autocorrelation of images formed by the ratios of two spectral bands is an excellent way to determine which combinations provide the best spectral representation of objects greater in size than the spatial resolution of the sensor. This is because an image formed from the ratios of two nonredundant bands will enhance spectral objects and thus tend to have greater spatial autocorrelation than the ratio of two bands that are very similar. Ratios are a particularly effective method of combining images because this operation tends to reduce the effect of illumination differences and to enhance spectral features. e selection is the process of finding a subset of the original bands that provides an optimal trade-off between probability of error and classification cost (Swain and Davis, 1978). Three feature selection problems are addressed in this paper: (1) narrow band feature selection, which is the selection of a subset of individual bands; (2) broad band feature selection, in which groups of adjacent bands are selected, and (3) nonadjacent multiple band feature selection, in which selection. of the groups of bands is not limited to adjacent bands. Spatial autocorrelation is useful in all three feature selection problems. Narrow band feature selection is carried out by ranking the spatial autocorrelation of all possible combinations of ratioed bands. Broad band feature selection can be carried out by iteratively grouping adjacent bands that are the most similar. If the grouping is started from the previously identified best bands, it is possible to develop a metric to check that the incorporation of each additional band to the group enhances the spatial autocorrelation of all the groups of bands together. Nonadjacent multiple band feature selection is simply an extension of the broad band case, except any of the original bands can potentially be grouped in any of the features. Tests with simulated data indicate that the spatial autocorrelation based methods consistently identify the best bands or groups of bands. Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) data of eastern Washington state are used to illustrate the technique on -real data. The results suggest that visible and near-infrared bands provide a large proportion of the spectral and spatial information in that scene. Adjacent bands in many cases provide similar information, but there are important exceptions such as on the red edge of the infra-red plateau.