324.

[Cover] [Contents] [Index]

Page 90

in classification errors. In agricultural crop applications, the objects are fields, the boundaries of which can be derived either by digitising paper maps of an appropriate scale, or by applying an edge detection algorithm to the image. Each object (field) is characterised by global statistical parameters, and is represented by a unique vector in feature space (Mégier et al., 1984; Erickson and Lickens, 1984; Mason et al., 1988; Jansen et al., 1990).

In contrast to the mainly ad hoc procedures described above, two specific multisource data fusion mechanisms known as extension of Bayesian theory and Evidential Reasoning have their particularly practical value because both approaches have a robust theory basis for dealing with multisource data sets. Details of these two methods for managing multisource data are provided in Chapter 7.

2.6 Sampling scheme and sample size

A sampling scheme describes the way in which sample pixels are selected from the image in order to characterise the thematic classes of interest. Sample size is important in terms of the accuracy with which estimates of statistical parameters describing these classes are obtained. Both sampling scheme and sample size are also of importance in assessing the accuracy of the thematic map derived from remotely sensed data. Samples may be derived from field observation, from farm records (in the case of agricultural crops), or from maps or air photographs. If the target classes are of a temporally changing nature, care must also be taken to ensure that the sample data adequately represent the temporal state of the phenomena being observed. There are certain restrictions on sampling, including cost, availability of source information such as maps and air photographs, and accessibility. If the area of interest is large, then it may not be possible to conduct a thorough and statistically valid sampling procedure close to the time of the satellite overpass. When cloud is a problem, the logistics of sampling are made more difficult by the need to be ‘on call’ for a number of overpass times. Nevertheless, attention must be paid to the questions of sampling scheme and sample size, particularly if statistical methods of pattern recognition are employed.

A distinction is made between two kinds of sample data. Training data are used in supervised methods of pattern recognition to ‘teach’ a classifier the main characteristics of each class. Campbell (1987) points out that:

● the number of sample observations has a direct relationship with the confidence interval of the estimate of the accuracy of a classification, and on the estimates of statistical parameters used in the particular, chosen, classifier. For example, estimates of the mean vector and variance-covariance matrices of the individual classes are required by the maximum likelihood classifier, and

[Cover] [Contents] [Index]