4.

[Cover] [Contents] [Index]

Page 100

boundaries. The use of a buffering operation within a GIS improved classification accuracy without compromising the integrity of the results, and removed autocorrelated errors.

2.8 Epilogue

In this chapter, we have attempted to provide an overview of methods of pattern recognition, plus a summary of related issues such as sampling and accuracy assessment methods. The range of material covered may appear to the reader to be daunting; however, the following chapters provide considerably more detail and descriptions of the principal methods of pattern recognition that are summarised in this chapter. The consensus view, in recent years, has emphasised the superiority of artificial neural network methods over statistical methods, largely because of their non-parametric nature (that is, the fact that ANN do not assume any particular statistical frequency distribution of the data). More recently, the use of decision trees has assumed a higher ‘profile’; like ANN they are nonparametric, but—unlike ANN—they do not need extensive design and training. However, their use of hyperplane decision boundaries parallel to the feature axes may restrict their use to cases in which classes are clearly distinguishable.

The use of multiple classifiers may present a way out of the spiral of increasing complexity. Rather than develop increasingly refined decision rules, it may be sensible to make best use of what is available. Users of these methods should be aware of the need to ensure that the component classifications are independent.

Hyperspectral data present difficult problems due to their high dimensionality. It is impracticable to consider the collection of large volumes of test and training data when the intrinsic dimensionality of the hyperspectral data set is considerably less than the number of spectral bands. A proportion of the variance exhibited by hyperspectral data is noise, either random or coherent. Hence, it makes sense to consider the use of orthogonal transforms, particularly those that explicitly discriminate between signal and noise. The two methods described in this chapter—the MNF and MAF procedures—offer the possibility of reducing the size of the data set without compromising accuracy, while avoiding the problems associated with high dimensionality.

The need to estimate the accuracy of a thematic map, or validate the methodology, is an issue that requires considerable attention; in fact, no pattern recognition exercise is complete unless it includes validation and accuracy estimation. The simple ‘overall accuracy’ measure deriving from the confusion matrix gives only a very rough estimate of the true accuracy, and the use of kappa is statistically more acceptable, provided that the user is aware of the need to ensure a sample of test data of sufficient size that has been collected using a random sampling scheme. None of these meas

[Cover] [Contents] [Index]