List of Tables
7. DISCUSSION
7.3 CLASSIFICATION METHODS
7.3.2 Spectral mixture analysis
The limitations of hard classifiers found in this study mirror those reported by other researchers. The main reason for the difficulty in applying hard classification techniques to wetland mapping is that the spectral characteristics of these environments represent a continuum from open water with no emergent vegetation through to complete vegetation cover where no water is visible. These complex landscapes do not lend themselves to hard classification and are much better served by sub-pixel classification methods.
Table 7.3 compares the endmembers used by a number of other studies with those applied in this study. Each of the studies included a 'dark' endmember, often called shade but sometimes referred to as clear water, or just water. Some authors commented on the spectral similarities of shade and clear water and indicated that these endmembers were usually interchangeable (e.g. Mertes et al, 1995). Shade was increased by the presence of trees (Adams et al, 1995) but was considered less of an issue in the grassland and wetland environments of the Eastern Shores. Each of the reviewed studies also included at least one vegetation endmember, mostly representing dark green, leafy vegetation. In the current study, two vegetation endmembers were used, one called woody, the other non-woody. The inclusion of these vegetation endmembers was crucial for the success of this study as it allowed the complex mixing of water and vegetation to be described. These two endmembers did not represent specific vegetation species, but grouped together classes of vegetation types with similar spectral characteristics. Indeed, Adams et al. (1995) noted that it was not appropriate to interpret vegetation endmembers as actual vegetation species but rather as a grouping of vegetation types. The woody vegetation endmember represented all green, leafy vegetation that had not been excluded by the masking process, while the non-woody endmember was chosen to represent all grassland and wetland vegetation.
Table 7.3: Comparison of endmembers used in the current study and those used in studies reported in the literature.
Study Current study Mertes et al (1993) Adams et al. (1995) Mertes et al. (1995) Mertes (1997) Binaghie/a/. (1999) Elmore et al. (1999) Roberts et al. (1999) Small (2001)
Lu et al. (2004)
Theau & Duguay (2004)
Endmembers selected Water, woody, non-woody
Shade, soil, vegetation, various water/sediment mixtures Shade, green vegetation, non-photosynthetic vegetation, soil Shade/clear water, vegetation, sediment-laden water
Clear water, muddy water, vegetation Water, wetland
Shade, vegetation, light soil, dark soil
Shade, green vegetation, non-photosynthetic vegetation, soil High albedo, low albedo, vegetation
Shade, soil, green vegetation Shade, canopy, lichen
In keeping with the largely image-based techniques utilised in this study, endmembers were selected from the images themselves and no reliance was placed on laboratory
measurements. This was similar to the techniques described by Mertes et al. (1995), Binaghi et al. (1999), Elmore et al. (1999), Small (2001) and Lu et al. (2004), among others, who all used image-derived endmembers in their research. The actual method of image endmember selection varied from study to study, with some identifying endmembers from scattergrams and others using ground truth data. Bateson & Curtiss (1996) described a method for selecting endmembers from scattergram displays of spectral data in multidimensional space but conceded that the method suffered from user subjectivity. This problem was apparent in the current project where the selection of endmembers from component data scattergrams produced unsatisfactory results. It is quite likely that another user would select a different set of endmembers from the same scattergrams, and in so doing obtain different (and perhaps more accurate) results from the mixture analysis. This was demonstrated by Theau & Duguay (2004) who, in using spectral mixture analysis to map lichen in northern Canada, tested different sets of scattergram-selected endmembers and obtained slightly different results each time. The scattergram approach to selecting endmembers obviously needs to be performed with care and requires numerous iterations before producing accurate endmembers. Nevertheless, a number of studies have reported success with this method. Small (2001) found that two different sets of endmembers selected from scattergrams both provided reasonable estimates of vegetation cover fractions. Similarly, Theau & Duguay (2004) selected shade, lichen and canopy endmembers from the vertices of a TM 3 and 4 scattergram and reported high accuracies in mapping lichen cover.
The other method of selecting image endmembers involves locating endmembers on the images themselves. This is similar to the training site technique used in supervised classification but differs in that endmember sites represent purest examples of endmembers and do not contain mixtures of other elements (as can happen with traditional training sites). Locating water or shade endmembers is easy, with most studies using water bodies for this purpose (e.g. Elmore et al, 1999; Lu et al, 2004). Finding vegetation endmembers is rather more difficult as pure examples rarely exist in nature. For instance, trees are invariably accompanied by shade while the 30 m pixel size of Landsat TM images makes it almost inevitable that some degree of mixing occurs within pixels. In this study, maximum NDVI values were used to define the woody vegetation endmembers. In doing this it was assumed that the presence of shade would lower NDVI values and that those pixels with the highest NDVI values represented forest with the least amount of shade. The non-woody
vegetation endmembers were selected from the modal pixel values of the training sites on the assumption that these were the most representative pixels. These non-woody vegetation endmembers where also assumed to be representative of all the vegetation species found in the grassland and wetland environments in the study area. In a similar manner, Elmore et al. (1999) used riparian vegetation along a river to define an endmember representing all vegetation in their study area while Lu et al. (2004) used areas of dense pasture to define an endmember for all green vegetation.
Both Adams at al. (1995) and Small (2001) noted the ease with which fraction images could be interpreted, being more intuitive than digital numbers (DN) and also facilitating the comparison of multitemporal images. Interpretation of fraction images is further aided by classifying the fractions into different classes. This generally involves applying user- defined land cover thresholds to the endmember fraction images to produce a land cover map. In this study, where water was the land cover of interest, the water fraction images were classified into ten classes representing water concentrations at 10% intervals. This greatly reduced the volume of information to be processed and facilitated the interpretation of results. Compare this to Lu et al. (2004) who used fraction images to produce maps of mature forest, advanced secondary succession, initial secondary succession, pasture, agricultural lands, bare lands and water. No attempt was made to map varying concentrations of vegetation within pixels, the fractions being used instead to define boundaries for the different land cover classes. Similar methods were also employed by Adams et al. (1995), Mertes et al. (1995) and Roberts et al. (1999). In all these studies, spectral mixture analysis was used to determine sub-pixel mixing, which in turn was used to define class boundaries in a manner not possible with hard classifiers. The results were land cover maps, not maps showing concentrations of a particular feature. This, then, is the crucial difference between the current study and other studies reviewed in the literature:
this study used mixture analysis to produce maps of water concentration whereas most other studies produced maps depicting the distribution of different feature types.