351.

[Cover] [Contents] [Index]

Page x

‘c’. In the case of classification of an agricultural area, a pixel labelled as ‘carrot’ is more likely to be surrounded by pixels of the same class rather than by other classes such as ‘water’ or ‘wheat’. The ability to model such contextual behaviour may reduce confusion in the classification process.

The decision to write this book was triggered by our experience in attempting to use new methods of describing and labelling pixels in a remotely sensed image. While a number of valuable but more general textbooks are available for undergraduate use, we know of no coherent source of advanced information and guidance in the area of pattern recognition for research scientists and postgraduate research students in remote sensing, together with students taking advanced remote sensing courses. We hope that this book will contribute to the increased understanding and adoption of recently developed techniques of pattern recognition, and that it will provide readers with a link between the remote sensing literature and that of statistics, artificial intelligence and computing. We do suggest, however, that attention be paid to experimental design and definition of the problem, for an advanced pattern recognition procedure is no substitute for thinking about the problem and defining an appropriate set of features.

Chapter 1 introduces the basic concepts of remote sensing in the optical and microwave region of the electromagnetic spectrum. This chapter is intended to introduce the field of remote sensing to readers with little or no background in this area, and it can be omitted by readers with an adequate background knowledge of remote sensing.

Chapter 2 introduces the principles of pattern recognition. Traditional decision rules, including supervised minimal distance classifier, Gaussian maximum likelihood and unsupervised clustering techniques, are described, together with other methods such as fuzzy-based procedures and decision trees. The chapter also contains brief accounts of dimension reduction methods, including orthogonal transforms, the assessment of classification accuracy, and the principles underlying the choice of training data.

Chapter 3 describes widely used neural network models and architectures including the multilayer perceptron (also called the feed-forward neural network), Kohonen’s self-organised feature map, counter-propagation, the Hopfield network, and networks based upon adaptive resonance theory (ART).

Chapter 4 deals with pattern recognition techniques based on fuzzy systems. The main topics of this chapter are the construction of fuzzy rules, fuzzy mapping functions and the corresponding decision processes.

Chapter 5 presents a survey of methods of quantifying image texture, including fractal—and multifractal-based theory, the multiplicative autoregressive random field model, the grey level co-occurrence matrix and frequency domain filtering.

Chapter 6 addresses the theory and the application of Markov random fields. The main application of Markov random fields is to model contextual relationships. Other related topics, including function formulation,

[Cover] [Contents] [Index]