Abstract
Maximum likelihood method and a neural network approach are the most common supervised classification method used with remote sensing multispectral image data such as Landsat TM data. In these method, training samples from each desired set of classes on the original data are used to estimate the parameters of the particular classifier algorithm. Consequently, these parameters depend on observed season and latitude of the observed area.
In this paper, a season and latitude independent analysis method is developed. Information of the original data are separated into a parameter which depends on season and latitude, and parameters which are independent of these conditions by a self-consistent data correction and a normalization. The condition independent parameters are expanded by three principal terms obtained from typical spectral patterns of water, vegetation and soil.
The pattern components are available to analysis and to classify remote sensing multispectral image data under free from the observed conditions and also available to compare directly with data observed on the ground using multispectral radiometer.
