Comparison of the Certainties in a GIS
D.Arnarsaikhan, M.Ganzorig
Informatics & RS Centre, Mongolian Academy of Sciences
av.Enkhtaivan-54B, Ulaanbaatar-51, Mongolia
Abstract
At present there are many techniques for land cover mapping using . RS data. The results of various parametric and non-parametric , classifications with different certainties (qualities) are put to . a GIS and used in decision making. The most commonly used . algorithms are the parametric and non-parametric (eg, k-nearest . neighbor maximum likelihood classifiers against which other . classification algorithms are compared. Various authors have used the Dempster-Shafer theory of evidence for the land cover , discrimination and argued that the result is better than the result of the standard maximum likelihood classification. The aim of the study is to compare these methods. For the comparison, two different natural regions have been selected. 1
1. Introduction
The linkage between RS and GIS can be made automatically by the use of some pattern recognition techniques and manually by a visual interpretation. Over the years, researchers have developed various techniques to improve the classification results. The results of the existing classification techniques containing various certainties can be put into a GIS or directly used in some decision making. The processing steps of RS data to be used in a decision making process is shown in Figure 1. When the result of an image classification is used to update a layer of a GIS or in a decision making process the requirements proposed by Mulder et al. (1990) and Arnarsaikhan et al. (1992) should be met. Traditionally, the parametric maximum likelihood classification incorporating a priori probabilities or ancillary data is considered as one of the most efficient methods for the spectral classification. However, it has many limitations; for instance, in practice it is rare that the reflectance values have a multivariate normal distribution (Ince 1987, Mulder et al. 1990) .A number of authors (eg, Srinivasan and Richards 1990, Bronsveld et al. 1992) used the theory of evidence in the classification process and judged that the result is better than the result of the standard maximum likelihood classification. However, assumptions made in the Dempster theory are weaker than those made in the maximum likelihood classification, because it uses some assumptions (eg, AnB) used in fuzzy logic schemes. These techniques have been tested in two different natural regions. The results showed that by the use of the same technique, different results can be expected according to the context of the area.
Figure 1 Processing Steps of RS Data to Update a GIS
2. The Test Sites and Materials
2.1. The first target area selected was in the central part of
Mongolia, near Ulaanbaatar city and has various natural diversities. In this area, the boundaries between the object classes are hardly distiguishable, and it is not easy to choose the satisfactory training samples. In this area soil, vegetation and water classes were selected.
2.1.1. For this area, a Landsat TM image of 1990 has been selected. In the maximum likelihood classification bands 2, 3,4 and in the classification which uses the theory of evidence bands 3 and 4 were used, respectively. A landuse map was produced using a knowledge of a specialist who worked for a long time in the study area.
2.2. The second test area was selected in the south eastern part of the country. The area is related to the steppe zone and during the summer time the land in this region is mainly covered by grass. For our study, we selected a part where vegetation, soil and solonchak are clearly separable from each other. In this area soil, vegetation and solonchak classes were selected.
2.2.1. For the image classification, 3 bands (green, red and NIR} of MSK-4 data converted to a digital format were used. The data were resampled to a spatial resolution of 30 meters. The band selection and a landuse map production in the area have the same procedures as mentioned above (see section 2.1.1}.
In both cases, the training samples (total of 180 to 240 pixels for each class} were selected avoiding multimodel and asymmetric distributions.
