ABSTRACT
Thus far we have thought of multispectral and SAR images as three-dimensional arrays of pixel intensities (columns × rows × bands) representing, more or less directly, measured radiances. In the present chapter we consider other, more abstract representations which are useful in image interpretation and analysis and which will play an important role in later chapters. The discrete Fourier and wavelet transforms that we treat in Sections 3.1
and 3.2 convert the pixel values in a given spectral band to linear combinations of orthogonal functions of spatial frequency and distance. They may therefore be classified as spatial transformations. The principal components, minimum noise fraction and maximum autocorrelation factor transformations (Sections 3.3 to 3.5), on the other hand, create at each pixel location new linear combinations of the pixel intensities from all of the spectral bands and can properly be called spectral transformations (Schowengerdt, 1997).
