333.

[Cover] [Contents] [Index]

Page 99

(2.30)

The distribution of the ratio is approximately Gaussian, and so it can be used as a test statistic for the null hypothesis that the observed accuracy differs from zero only as a result of sampling from a large population. Individual class values of kappa (‘conditional kappa’ values) can be derived as follows:

(2.31)

The kappa coefficient takes not just the principal diagonal entries, but also the off-diagonal entries, into consideration. The higher the value of kappa, the better the classification performance. If all information classes are correctly identified, kappa takes the value 1. As the values of the off-diagonal entries increase, so the value of kappa decreases. Table 6.2 lists four confusion matrices with their associated kappa values. It should be noted that the interpretation of the kappa statistic is based on the assumption of a multinormal sampling model. If the test data are not chosen properly, the above assessments become less reliable. Another consideration relates to sample size and sampling scheme; the consensus view appears to be that simple random sampling is required for the use of the kappa coefficient, and that a minimum sample size is needed in order to ensure a specific, predefined level of accuracy. Issues concerning sample size and sampling procedures are considered in Section 2.6. Recent surveys of accuracy measures are Congalton and Green (1998), Foody (2000a), Jansen and van der Wel (1994) and Stehman and Czaplewinski (1998).

None of the methods of assessing the accuracy of a classification derived from remotely sensed data considers the spatial distribution of error. Without considering any other factors, one might expect that erroneously classified pixels should be randomly distributed over the study area. Observations of the actual distribution of such erroneously classified pixels may show that this is not the case. For instance, in a per-pixel classification it might be seen that erroneously classified pixels are distributed around field boundaries, or along spatial features such as roads and railways that are not included in the classification. Vieira and Mather (1999) present some methods of visualising the spatial pattern of classification error, and of analysing these patterns using simple methods of spatial statistics. The same authors show how cartographic methods can be used to depict the reliability of classifications derived from remotely sensed data. The spatial distribution of errors in agricultural crop classifications was found by Vieira and Mather (1999) to be closely related to the positions of field

[Cover] [Contents] [Index]