List of Tables
6. RESULTS
6.4 IMAGE CLASSIFICATION
6.4.2 Maximum likelihood classification
Table 6.9: Details of cluster allocations determined through ground truthing for (a) the component dataset and (b) the spectral dataset. Clusters present in both wetland and grassland areas are noted in red.
(a) Component Dataset Date
1991/07/23 2001/03/20 2001/05/07 2002/04/24 2002/07/13 2002/09/15 2002/10/17
Wetland Clusters 5,6,8,11,12 2, 5, 6, 8, 15 4,6,8,10,12,13 2, 4, 8, 9
2 , 3 , 4 , 5 , 8 , 1 4 4 , 5 , 6 , 8 , 11,14 2,5,8,11
Grassland Clusters 2,3,4
3,4,5 2 , 3 , 5 , 7 2,3,8 2 , 3 , 5 , 6 , 7 3,4,14 2, 4, 6, 8
Overlapping Clusters None
5 None 2,8 2,3,5 4,14 2,8 (b) Spectral Dataset
Date 1991/07/23 2001/03/20 2001/05/07 2002/04/24 2002/07/13 2002/09/15 2002/10/17
Wetland Clusters 5,6,7,11,14,15 4, 7, 9, 14 2 , 4 , 7 , 8 , 10,15 2 , 4 , 5 , 9 , 1 4 , 15 2,4,7,13
3,4,7,10 3,6,7,15
Grassland Clusters 2,3,9
2,3 2,3,6 2,3,4 2 , 3 , 4 2,4,7 2,3,4
Overlapping Clusters None
None 2 2,4 2,4 4,7 3
While it is possible that the different clusters produced by the cluster classification represented various ratios of water and vegetation, interpreting and classifying these clusters into meaningful classes proved extremely difficult. Even a simple classification into three classes yielded large areas for which it was not possible to make an allocation.
Given the confusion caused by the large numbers of overlapping clusters, it was not possible to adequately map surface water in the study area using cluster analysis.
being assigned to the class to which their probability of membership is the greatest. In performing this classification, those pixels which had a low probability of belonging to any of the classes were left unclassified. As shown in Table 6.10, the number of unclassified pixels ranged from 34% in July 2002 to 77% in September 2002. That so many pixels were left unclassified could have been due to a number of factors. Firstly, the training sites might have been poorly chosen and, hence, did not adequately represent the intended land cover classes. Secondly, the water, wetlands and grasslands in the study area are possibly a lot more complex than could be captured with just three types of training site. This landscape complexity was also reflected in the results of the cluster and mixture analyses discussed elsewhere in this chapter. In addition, the fact that the numbers of unclassified pixels varied so greatly from date to date suggests the presence of seasonal and other effects (e.g. fire) on the Eastern Shores landscape.
Table 6.10: Percentage pixels left unclassified by the maximum likelihood classifier.
Date 1991/07/23 2001/03/20 2001/05/07 2002/04/24 2002/07/13 2002/09/15 2002/10/17
Component Data
55 73 59 51 37 77 58
Spectral Data
55 70 57 49 34 75 61
As with the unsupervised cluster classification, virtually all the pixels classified as water were found in Lake Bhangazi (See Appendix 5). These water areas ranged from a maximum of 290 ha in July 1991 to a minimum of 227 ha in October 2002, with the lower value being 78% of the size of the larger area (Figure 6.6). While these values are similar to the values obtained by the unsupervised cluster analysis, they tend to be slightly lower for each of the study dates. When compared to the wetland and grassland classes, the range in size of the open water class is rather small, with the smallest grassland and wetland sizes being only 18% and 20% respectively of their largest areas. Once again, this could be due to the training sites not adequately capturing the heterogeneity and/or seasonality of the landscape. Of particular interest in both Figures 6.6(a) and 6.6(b) is that the variation in the
sizes of the wetland and grassland classes appears to be negatively correlated with the number of unclassified pixels. This is borne out by regression analysis, which shows a correlation of-0.995 between the number of unclassified pixels and the combined totals of the wetland and grassland classes. Given that the sizes of the water classes are very similar to those obtained by the cluster classification, it would appear that most of the unclassified pixels are either grassland or wetland pixels and not water.
A cross-tabulation of the spectral data classification against that of the component data provides an indication of how closely the classifications of the spectral and component datasets match. Table 6.11 shows how in July 1991, pixels classified as water, grassland or wetland using the component data were either given the same classification using the spectral data or were left unclassified. This contrasts with the cross-tabulation of the July 2002 data where it can be seen that some degree of confusion exists between wetland and grassland classes. The figures highlighted in bold in Table 6.11(b) show pixels that,
a ) 8000
6000
nT x
"£• 4000
<
2000
1991/07/23 2001/03/20 2001/05/07 2002/04/24 2002/07/13 2002/09/15 2002/10/17 Date
— — Unclassified — — Grassland
— Water — — Wetland b ) 8000
6000
"5T x
a, 4000
<
2000
0
1991/07/23 2001/03/20 2001/05/07 2002/04/24 2002/07/13 2002/09/15 2002/10/17 Date
Figure 6.6: Results of the maximum likelihood classification for (a) the component data and (b) the spectral data. Areas for each land cover class are shown in hectares.
when classified as being either grassland or wetland using the spectral data, were classified as wetland or grassland respectively using the component dataset. This 'cross-classification' of grassland and wetland occurred in all four of the 2002 study dates but not at all in 1991 and 2001. Water pixels, by contrast, when classed as water using one of the datasets, were always given the same classification using the other dataset, or where left unclassified. This was true for each of the study dates.
Table 6.11: Cross tabulation of maximum likelihood classification results of the component data against the spectral data for (a) July 1991 and (b) July 2002.
(a) Component Data
Unclassified Water Grassland
Wetland
Unclassified 58304
30 2364 1999
Water 65 3152 0 0
Grassland 2585
0 35033 0
Wetland 1323
0 0 9310 (b)
5S
Component Data Unclassified
Water Grassland
Wetland
Unclassified 33725
28 2974 5347
Water 174 2794 0 0
Grassland 2713
0 37674 2892
Wetland 2303
0 1194 22347
The large numbers of unclassified pixels and the cross-classifications that occurred between grassland and wetland pixels from the two different datasets, made it difficult to interpret the results with any degree of confidence. The fact that so many pixels had been left unclassified made it impossible to map seasonal and spatial variations in water and wetlands with any certainty.