ABSTRACT
This book introduces some basic statistical concepts for multi-spectral and synthetic aperture radar images viewed as random mechanisms. In an optical/infrared or a synthetic aperture radar image, a given pixel value g(i, j), derived from the measured radiation field at a satellite sensor, is never exactly reproducible. It is the outcome of a complex measurement influenced by instrument noise, atmospheric conditions, changing illumination and so forth. It may be assumed, however, that there is an underlying random mechanism with an associated probability distribution which restricts the possible outcomes in some way. A random variable can be used to represent a quantity, in the present context an image gray-scale value, which changes in an unpredictable way each time it is observed. The evaluation of a land cover classification model involves repeated trials with a finite number of independent test observations, keeping track of the number of times the model fails to predict the correct land cover category.
