Study Guide
Source Materials
Ma_RS9-SupervisedClassification2026.ppt
Ma_RS9-SupervisedClassification2026.ppt
Download original
Re-extract
16 slides.
Left = our PNG render of the original slide. Right = our extracted text (edit to correct).
slide-01
Introduction to Remote Sensing Supervised classification Supervised classification requires a priori knowledge about the image data. • which types of land-use exist in the study area •geographical (spatial) locations of reliable samples for each land-use type 1
slide-02
Introduction to Remote Sensing Procedure of supervised classification Selection of training samples Generation and evaluation of statistical signatures Class assignment using a decision rule 2
slide-03
Introduction to Remote Sensing Training samples Samples of homogeneous areas with known class (land-use) types Selection of training samples?? fieldwork aerial photography maps personal experience Spectral characteristics of training samples are used to generate signatures. 3
slide-04
Introduction to Remote Sensing Parametric signatures Multivariate statistical parameters • maximum, minimum, mean, standard deviation, variance, covariance Each signature corresponds to a class, and is used with a decision rule to assign the pixels to a class. Signature can be evaluated, deleted, renamed and merged. 4
slide-05
Introduction to Remote Sensing Mean, standard deviation, and variance Mean1 of Band1 = (24+26+28)/3 = 26 Mean2 of Band2 = (3+5+10)/3 = 6 (24,3) Standard deviation1 of Band1= 2 (26,5) Standard deviation2 of Band2 = 3.6 (28,10) k (Q Q ) 2 water signature i SQ i 1 with 3 pixels in k1 two bands where: i = a particular Variance is the squared standard deviation. pixel Variance of Band1 = 4 k = the number of pixels Variance of band2 = 12.96 5
slide-06
Introduction to Remote Sensing Covariance Measures the tendencies of data file values in a signature, but in different bands, to vary with each other, in relation to the means of their respective bands. Covariance of band1 and band2 = 7 (24,3) k (26,5) (Q i 1 i Q )( R i R ) (28,10) CQR k1 water signature with 3 pixels in two bands The covariance matrix is an n*n where: matrix that contains all of the variance i = a particular and covariances within n bands of data. pixel k = the number of pixels 6
slide-07
Introduction to Remote Sensing Decision rules A mathematical algorithm that performs the actual sorting of pixels into distinct classes. Examples of decision rules Parallelepiped Minimum Distance Maximum Likelihood/Bayesian Mahalanobis Distance 7
slide-08
Introduction to Remote Sensing Parallelepiped (1) Have upper and lower limits for every signature in every band. If pixels are found to lie in such a parallelepiped, then they are assigned to that signature’s class. Parallelepiped classification using +standard deviations as limits (Source: ERDAS Field Guide, 2002, p.229) 8
slide-09
Introduction to Remote Sensing Parallelepiped (2) The limits for every signature can be any values that the user specifies • e.g. either the minimum and maximum data file values of each band in the signature, • based on the user’s knowledge of the data and signatures. Advantages • a very simple, and fast supervised classifier. • useful for a first-class, broad classification. Disadvantages • pixels in gap regions will not be classified. • pixels in “corners” may be classified improperly. 9
slide-10
Introduction to Remote Sensing Minimum Distance (1) Calculates the spectral distance between the measurement vector for the pixel and the mean vector for each signature. n 2 D ist ( class _ m ) k m ,k ( B k 1 V ) A pixel belongs to a class w , if the spectral distance to that class is the closest (lowest). Minimum spectral distance (Source: ERDAS Field Guide, 2002, p.232) 10
slide-11
Introduction to Remote Sensing Minimum Distance (2) Advantages – no unclassified pixels. – a fast decision rule. Disadvantages – pixels which should be unclassified, because they are not spectrally close to the mean of any sample, will become classified. – it does not consider class variability. 11
slide-12
Introduction to Remote Sensing Maximum Likelihood (1) Based on the probability that a pixel belongs to a particular class. p c pi where i = 1,2,3, …, m possible classes The pixel at x belongs to a particular class c , if the likelihood that the correct class Maximum likelihood decision rule is c, is the largest. (Source: ERDAS IMAGINE V8.3 Professional Training Reference Manual 1997 p.29) 12
slide-13
Maximum likelihood classification Assumption: data for each class is normally distributed Let C=(C1, C2, …, Cnc) denote a set of nc classes. For a given pixel with gray level vector x, the probability that x belongs class Ci is P(Ci|x) , i=1, 2, …, nc. x will belong to class Ci, if P(Ci|x) P(Cj|x) for all ji. 13 http://nature.berkeley.edu/~gong/textbook/
slide-14
Introduction to Remote Sensing Maximum Likelihood (2) The most common decision rule for supervised classification. Assumptions probabilities are equal for all classes. the input bands have normal distributions (the training data statistics for each class in each band are normally distributed). 14
slide-15
Introduction to Remote Sensing Maximum Likelihood/ Bayesian (3) Bayesian decision rule If the user has a prior knowledge that the probabilities are not equal for all classes, the user can specify weight factors for particular classes. This variation of the maximum likelihood decision rule is known as Bayesian decision rule (Hord 1982). Characteristics – the most accurate classifier – it takes most variables into account – the parallelepiped or minimum distance decision rule may generate better results while the histograms of the bands of data do not have normal distributions. 15
slide-16
Introduction to Remote Sensing Maximum Likelihood/Bayesian (4) The pixel is assigned to the class, for which D is the lowest. 1 1 D log e (ac ) log e (| Vc |) ( X M c )T Vc 1 ( X M c ) 2 2 Mahalanobis where distance X= the measurement vector of the candidate pixel Mc = the mean vector of the data in class c Vc = the covariance matrix of the data in class c |Vc| = the determinant of the covariance matrix Vc Vc-1 = the inverse of Vc T = the transposition function ac = the probability that class c occurs in the image (equal for all classes, or is entered from a priori knowledge); 18