Pattern Recognition and Neural Networks

A major topic in pattern recognition is feature extraction. An excellent discussion of feature extraction and selection problem in remote sensing with multispectral and hyperspectral images is given by Landgrebe (5). In remote sensing, features are usually taken from the measurements of spectral bands, which this means 6 to 8 features in multispectral data, but a feature vector dimension of several hundred in hyperspectral image data.

With a limited number of training samples, increasing the feature dimension in hyperspectral images may actually degrade the classification performance, which is referred to as the Hughes phenomenon. Reference (5) presents procedures to reduce such phenomena. Neural networks have found many uses in remote sensing, especially with pattern classification. The back-propagation trained network, the radial basis function network, and the support vector machine are the three best-performing neural networks for classification.

A good discussion on statistical and neural network methods in remote sensing classification is contained in Ref. 11 as well as many other articles that appear in Refs. 3, 4, and 12. A major advantage of neural networks is that learning is from the training data only, and no assumption of the data model such as probability density is required. Also, it has been found that combining two neural network classifiers such as combining SOM, the self-organizing map, with a radial basis function network can achieve better classification than either one used alone (13).

One problem that is fairly unique and significant to remote sensing image recognition is the use of contextual information in pattern recognition. In remote sensing image data, there is a large amount of contextual information that must be used to improve the classification.

The usual procedure for contextual pattern recognition is to work with image models that exploit the contextual dependence. Markov random field models are the most popular, and with only a slightly increased amount of computation, classification performance can be improved with the use of such models.

Another pattern recognition topic is the change detection. The chapters by Serpico and Bruzzone (14) and Moser et al.(15) are recommended reading.

Data Fusion and Knowledge-Based Systems. In remote sensing, there are often data from several sensors or sources. There is no optimum or well-accepted approach to the problem. Approaches can range from more theoretic, like consensus theory (16), to fuzzy logic, neural networks, multistrategy learning to fairly ad hoc techniques in some knowledge-based system to combine or merge information from different sources.

In some cases, fusion at decision level can be more effective than fusion at data or feature level, but the other way can be true in other cases. Readers are referred to chapters by Solaiman (17), Benediktsson and Kanellopoulos (18), and Binaghi et al. (19) for detailed discussion.