346.

[Cover] [Contents] [Index]

Page ix

nation can make decisions and judgements on complex problems in milliseconds, using experience as a guide. Automatic methods are more suitable to the routine processing of images that show predictable patterns. Such methods have been developed and applied in a number of disciplines, ranging from speech and handwriting recognition, industrial process management, and medical diagnosis, as well as in the collection of military intelligence. The main distinguishing characteristic of Earth observation data is its volume. Hence, methods that can be applied in other applications may not be suited to the analysis of remotely sensed data, because of the computational requirements. A further point to note is that there is often a discrepancy between the dimensionality of remotely sensed data sets and the volume of training data that is available. Where training data are sparse relative to the dimensionality of the data, it becomes difficult to estimate the characteristics of each training class, and so error may become significant. This phenomenon, that of increasing error with increasing data dimensionality, is sometimes known as the Hughes effect.

Advances in technology have led to rapid developments in methods of pattern recognition, leading to the formulation of new and more sophisticated decision rules. Some of these new methods have been introduced into the field of remote sensing, and have shown encouraging results. A further feature of remote sensing applications in recent years is the use of combinations of data derived from different sensors, or from different time-periods, plus terrain and other data extracted from GIS databases. In addition, the spectral information contained in remotely sensed images is often augmented by derived measures, such as values of texture and context. In dealing with multidata sources, a significant problem is the considerable increase in the computational cost. Other problems, such as data scale and data reliability, have also to be considered. There is an increasing interest in seeking methods for efficiently manipulating multisource data in order to increase classification accuracy. It should always be remembered, though, that sophisticated algorithms can not compensate for lack of training data, or an inadequate definition of the problem (in terms of the number and nature of the classes to be recognised relative to the scale of the study).

Texture is the tonal variation within an area. A simple example that illustrates the concept is the pattern on a carpet. If we treat each pattern as a whole, then the carpet can easily be described. If the carpet is seen as a set of small rectangular units then the problem of describing its properties is more difficult. In some cases, texture information seems to be more effective than tonal information to describe the objects, and one can develop texture features corresponding to different kinds of patterns to improve the performance of a classifier.

Contextual information describes how the object of interest may be affected by its neighbours. For instance, English words starting with the letter ‘q’ are more likely to be immediately followed by the letter ‘u’ than ‘z’ or

[Cover] [Contents] [Index]