Multispectral Remote Sensing Image Classification Using Wavelet Based Features

Multispectral remotely sensed images composed information over a large range of variation on frequencies (information) and these frequencies change over different regions (irregular or frequency variant behavior of the signal) which need to be estimated properly for an improved classification [1, 2, 3]. Multispectral remote sensing (RS) image data are basically complex in nature, which have both spectral features with correlated bands and spatial features correlated within the same band (also known as spatial correlation). An efficient method for utilization of these spectral and spatial (contextual) information can improve the classification performance significantly compared to the conventional non-contextual information based methods. In general, the spectral component of a remotely sensed image pixel may be a mixture of more than one spectral information that usually comes from the different regions of the study area. However, the conventional multispectral remotely sensed image classification systems detect object classes only according to the spectral information of the individual pixel/pattern in a particular frequency band, while a large amount of spatial and spectral information of neighboring pixels of different regions at other frequency bands are neglected. Hence the pixels are classified based on its spectral intensities of a specific band and does not give attention to its spatial and spectral dependencies and thus the spectral intensities of the neighbors at different frequency bands are assumed to be independent. Such approaches may be reasonable if spatial resolution is high or when the spectral intensities are well separated for different classes, which is rarely found in any real life data sets. For example, in the classification of urban areas, the densities of the spectral intensities are seldom well separated. Thus it is important to decide whether the arrangements of spatial data can be used as features directly in its original form or with a set of extracted features obtained through any feature-extraction method where, the information