An evaluation of spatial autocorrelation feature selection
Spatial autocorrelation analysis of ratioed image bands is compared to other methods of selecting features including the grouping of correlated bands, and the spacing of features equal wavelengths apart. In addition, features based on the band-passes of the SPOT XS, Multi-Spectral Scanner (MSS) and Thematic Mapper (TM) sensors are also synthesized. Three different types of spatial autocorrelation-based features are identified: (1) Narrow band features are selected by ranking the spatial autocorrelation of ratios of all possible combinations of bands. (2) Broad band features are produced by allowing the previously chosen best bands to merge with spectrally adjacent bands if this increases the overall spatial autocorrelation of the feature set. (3) Nonadjacent multiple band features are produced by removing the constraint that the merges are limited to neighbouring bands. The spatial autocorrelation ratio analysis suggests the majority of the most useful bands in an Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) scene of shrub-steppe vegetation are in the visible and near-infrared (VNIR) wavelengths less than 1100nm. The accuracy of classifications produced from the feature sets tends to increase with the number of bands, but saturates at approximately 10 features, irrespective of the band selection method. However, the spatial autocorrelation of the resulting classifications continues to improve as the number of features is increased. Spatial autocorrelation is highest for features comprising broad and nonadjacent multiple bands. Features based on groups of correlated bands showed variable results, but generally under-performed the broad and multiple band features chosen based on spatial autocorrelation. The classification of the synthesized SPOT, MSS and TM sensors suggests that the band passes of these sensors are indeed well placed. However, the quantization of the data appears to be as significant as the number of bands.