Spatial discrimination and classification maps

From MaRDI portal





A method is given serving to classify each of the mn observations Z(x) - for example, satellite observations on a regular \(m\times n\) lattice consisting of the energy spectra - into one of several categories and thus to construct a map of the terrain. Following \textit{P. Switzer} [Extension of linear discrimination analysis for statistical classification of remotely served satellite imagery. J. Int. Assoc. Math. Geol. 12, 367-376 (1980)] the method utilizes besides the observations themselves also their neighbours, and thus the error of misclassification may be reduced. Local spatial continuity of the neighbourhood of each observation is assumed, i.e. if \(Z_ 0\) belongs to \(\Pi_ i\) then so does every \(Z_ j=Z_ 1,...,Z_ s\) of the neighbour points \(Q_ j\). It is also assumed that the observations Z(x) and Z(y) can be factorized as \(cov\{Z(x),Z(y)\}=\{\rho | x-y| \}\Lambda\) where \(\rho\) (\(\cdot)\) is the isotropic correlation function and \(\Lambda\) is the common covariance matrix of the k populations \(\Pi_ i\sim N(\mu_ i,\Lambda),\) \(i=1,...,k.\) For the two-dimensional problem three cases are considered: the Four- Adjacent-Point Window with \(Z_ 1,Z_ 2,Z_ 3,Z_ 4\), \(s=4\) (the Switzer's case), the Four-Corner-Points Window with \(Z_ 5,Z_ 6,Z_ 7,Z_ 8\), \(s=4\), and the All-Square-Points Window, \(s=8\). The coefficients of the allocation vectors for these three windows and the respective efficiency parameters are given. Some generalizations of this problem are discussed (in particular the choice of window size) in a more general form. Sample estimation of parameters when training data are used is also considered, and some numerical examples are given.











This page was built for publication: Spatial discrimination and classification maps

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3727156)