322.

[Cover] [Contents] [Index]

Page 89

2.5.2 Using ancillary multi-source data

A significant problem arises when one attempts to combine images and ancillary data, namely, the different nature of these data types. Data may be categorical and, even when continuous in nature, may violate the normality assumption required by most statistical classification techniques. Ancillary information can be continuous or categorical, as suggested by Strahler et al. (1978) and Strahler (1981). Examples of continuous ancillary data are elevation, slope and aspect derived from a digital elevation model. Examples of categorical data are land use, soil and geological types. The different approaches proposed for the combination of spectral and ancillary information include: the logical channel approach, stratification, the use of prior probabilities, and file- or object-based classification.

The ‘stacked vector’ or logical channel approach consists of adding each ancillary data set as an additional feature, so that the pixel vector is extended by the addition of this external information. This technique is called the logical channel approach by Strahler et al. (1978). The advantages and disadvantages of this approach are discussed in Chapter 7, Section 7.1.

Stratification involves the subdivision or segmentation of the study area into smaller areas or strata, based on rules derived from external knowledge, so that each stratum can be processed independently. This process is performed before classification with the purpose of increasing the homogeneity of the individual data set to be classified, or to separate different objects that are spectrally similar. Some advantages relating to this kind of approach might be considered: one advantage is the convenience of dealing with smaller data sets; another benefit is the reduction of variability within individual strata. Stratification is effective and easy to implement, but is deterministic in the sense that it does not handle uncertainty about the occurrence of certain classes or the gradation between strata. An incorrect stratification can invalidate the entire classification process.

In most classification processes, the probability of class membership of a given pixel is assumed equal for all classes. However, if the information about an area shows the preference of certain classes for particular locations in the terrain, then this information can be expressed in terms of prior probabilities of occurrence for each class and this information can be incorporated into the classification process. The major difficulty involved in this approach is to define a suitable function of prior probability relating to each class in terms of achieving optimal results. Most experiments rely on the analyst’s ad hoc decision.

A further application of spatial information is the use of boundary information in order to define objects prior to classification. Such a classification strategy is suitable for patchy areas (e.g. areas made up of agricultural fields), and might be less affected by boundary pixels, which generally result

[Cover] [Contents] [Index]


Classification Methods for Remotely Sensed Data
Classification Methods for Remotely Sensed Data, Second Edition
ISBN: 1420090720
EAN: 2147483647
Year: 2001
Pages: 354

flylib.com © 2008-2017.
If you may any questions please contact us: flylib@qtcs.net