6.

[Cover] [Contents] [Index]

Page 102

Chapter 3
Pattern recognition using artificial neural networks

The efficiency of the human eye/brain combination in solving pattern recognition problems led researchers in this field to consider whether computer systems based on a simplified model of the brain can be more effective than standard statistical classification methods. Such research led to the adoption of artificial neural networks (ANN), which have been increasingly used in remote sensing over the past 10 years, mainly for image classification. An advantage of neural networks lies in the high computation rate achieved by their massive parallelism, resulting from a dense arrangement of interconnections (weights) and simple processors (neurones), which permits real-time processing of very large data sets.

Artificial neural networks are generally described as non-parametric; that is, the use of a neural network does not require any assumptions about the statistical distribution of the data. The performance of a neural network depends to a significant extent on how well it has been trained, and not on the adequacy of assumptions concerning the statistical distribution of the data, as is the case with the maximum likelihood classifier. During the training phase, the neural network ‘learns’ about regularities present in the training data and, based on these regularities, constructs rules that can be extended to the unknown data. This is a special ability of neural networks. However, the user must determine the architecture of the network, and also define parameters such as the learning rate, which affect the training time, performance and the rate of convergence of a neural network. There are no clear rules to assist with the design of the network, and only rules of thumb (or heuristics) exist to guide users in their choice of network parameters.

Five kinds of fundamental neural network architecture, including the multilayer perceptron with back-error propagation, the self-organised feature map (SOM), counter-propagation networks, Hopfield networks, and ART systems, are introduced in this chapter. All of these different types of network have been, or can be, used for classifying remotely sensed images (see Bishop (1995), Garson (1998), Haykin (1999), Hewitson and Crane

[Cover] [Contents] [Index]