Classification accuracy improvement and delineation of mixed pixels using hierarchical image classification

Among the supervised parametric classification methods, the maximum likelihood (MLH) classifier has become popular in remote sensing. Reliable prior probabilities (PPs) are not always freely available, and it is a common practice to perform the MLH classification with equal PPs. When equal PPs are used, the advantages of the MLH classification may not be attained. This study explores a hierarchical image classification (HIC) method to estimate PPs for the spectral classes using Landsat Thematic Mapper (TM) data and spectral class signatures. The TM pixels are visualized in spectral space relative to the spectral class probability surfaces. Prior probabilities are estimated from the pixels which fall within spectral class probability regions. The pixels likely to be misclassified are classified with the MLH classification with the estimated PPs. Besides the classified image, the HIC delineates mixed pixels and their land use/land cover class components at the specified significance level. The classified image resulting from the HIC shows increased accuracy over three classification methods. Delineated mixed pixels and their class components show visual agreement to the false color composites and aerial photographs