List of Tables
7. DISCUSSION
7.5 SURFACE WATER ON THE EASTERN SHORES
Inundation by water can occur in a number of ways: through direct precipitation, from groundwater exchange, or via overland flow from surrounding slopes or streams. For the Eastern Shores study area, where no large streams are found, rainfall and groundwater are the main determinants of surface water inundation. While short-term rainfall events undoubtedly increase surface inundation, it is clear that these increases are temporary and that seasonal rainfall patterns, expressed through groundwater levels, are the main controlling factors of inundation on the Eastern Shores. That this is true is shown by the high correlation between surface water extent and cumulative rainfall over periods of six months and more. Seasonal rainfall patterns appear to have long-term effects on groundwater levels, leading to patterns of inundation that reflect rainfall events that might have occurred many months previously.
Also of importance in this study is the possible presence of water in areas excluded by the forest mask. In defining the forest mask, the maximum extent of forest during the period covered by the satellite images was used, leading to a possible underestimation of the water resource on the Eastern Shores. While many hectares of plantation forest were felled during the period between July 1991 and October 2002, none of these areas were analysed in the study. With the removal of trees from these areas, it is very likely that surface water would reappear. The forest mask also excluded large patches of swamp forest from the analysis10.
10. Smith (2001) found that swamp forests on the Eastern Shores covered an area of just over 6 km2.
Swamp forests are characterised by the presence of waterlogged ground, and surface water would certainly occur in these areas. Excluding these forests from analysis would therefore result in an underestimation of surface water on the Eastern Shores. The amount of surface water thus excluded could also be estimated using methods developed by Walsh (2004) who used a hybrid classification system supported by ancillary DEM data to map swamp forests in Maputaland.
Apart from Lake Bhangazi, virtually all the surface water on the Eastern Shores occurs in conjunction with emergent vegetation. These areas, which are either permanently saturated or else undergo regular or occasional inundation, are likely to be characterised by distinctive vegetation communities. It is therefore illuminating to compare the areas of inundation obtained in this study with the land cover mappings obtained by C SIR/ARC (1998) and Smith (2001). As mentioned previously, the CSIR/ARC (1998) study mapped unimproved grassland, wetlands, and thicket & bushland as the main cover types on the Eastern Shores, while Smith (2001) found the area to be dominated by hygrophilous grasslands, woody grasslands, and sedge & grass swamp. Care should be taken when comparing these results with the current study as the classifications were produced using different methods and with satellite imagery from different dates. This fact, along with possible misregistration between the datasets, can lead to differences in the outputs from the different classifications.
A cross-tabulation of the pixels from water class 10 (i.e. 91 to 100% water concentration) against the equivalent classes in the CSIR/ARC (1998) and Smith (2001) classifications reveals high agreement between the three different classifications (Table 7.4). When compared to the Smith (2001) water body class there was a 90.1% agreement while a slightly lower value of 88.4% was obtained against the CSIR/ARC (1998) open water class. That there is such good agreement between the three methods is not surprising given the unique spectral characteristics of open water and the relative ease with which it can be mapped. All three classifications were based on Landsat TM images and, hence, used data with the same spectral characteristics.
Table 7.4: Percentage of pixels from water class 10 (91 to 100%) that occurred in the water land cover classes of the CSIR/ARC (1998) and Smith (2001) classifications.
Date
1991/07/23 2001/03/20 2001/05/07 2002/04/24 2002/07/13 2002/09/15 2002/10/17 Average
CSIR/ARC (1998) Class: Water bodies
87 88 87 89 92 87 89 88.4
Smith (2001) Class: Open water
90 89 88 91 94 89 90 90.1
Moving on to water classes 1 to 9, it is enlightening to examine the peak distribution of each class within the Smith (2001) and CSIR/ARC (1998) classifications. To this end, an analysis of the land cover class in which each water class occurred most often is presented in Table 7.5. From this table it can be seen that pixels in the different water classes occurred most often in the hygrophilous grassland class of the Smith (2001) dataset and in the unimproved grassland class of the CSIR/ARC (1998) classification. Smith (2001) defined hygrophilous grasslands as being waterlogged for most of the year, a fact confirmed by the results from this study where on 44 occasions out of 63 (9 classes, 7 study dates) the majority of pixels in a water concentration class were located in hygrophilous grasslands. According to the Smith (2001) classification, 44% of the (unmasked) study area comprised hygrophilous grasslands with woody grasslands making up 35% and sedge &
grass swamp another 11%. The high concordance between the hygrophilous grassland mapping from Smith (2001) and the water classes from the current study indicates that the different classification methods11 have produced complimentary results: the spectral mixture analysis has mapped water where Smith (2001) found hygrophilous grasslands, and vice versa.
.Smith (2001) used a maximum likelihood classifier, supplemented with ancillary datasets, to produce the land cover map of Maputaland. Mapping accuracy of hygrophilous grasslands was improved through the incorporation of a DEM and various aquatic vegetation datasets that provided indicators of the likelihood of occurrence of this land cover type.
Table 7.5: Land cover class in which each water concentration class occurred most often across the seven study dates for (a) the Smith (2001) dataset and (b) the CSIR/ARC (1998) dataset. The figures show how often the maximum number of water class pixels occurred in a particular land cover class. Percentages in brackets refer to the size of each land cover class as shown in Figure 7.1. Water class 10 is not shown since it was shown in Table 7.4.
(a) Smith (2001)
Hygrophilous grasslands (44%) Woody grassland (35%)
Open water (3%)
Sedge & grass swamp (11 %) Beach (<1%)
Mud (2%)
Water Class 1
4 3
2 7
3 6
1 4
7 5
5 1 1
6 5
1 1
7 6
1 8
3 2 1 1
9 1 1 4 1
Total 44
7 5 3 2 2 (b) CSIR/ARC (1998)
Unimproved grassland (62%) Wetlands (18%)
Thicket & bushland (10%) Water bodies (3%)
Water Class 1
7 2
6 1
3 5 2
4 4 3
5
J
3 1
6 5
1 1
7 6 1
8 6 1
9 6
1 Total
48 12 2 1
In analysing the comparison with the CSIR/ARC (1998) classification it should be remembered that this land cover map was produced for the whole of South Africa and land cover classes had, of necessity, to be broad enough to capture the main land cover types in the country. The unimproved grassland class, for instance, represented a generic grassland land cover and no attempt was made to capture the subtle variations in grassland found throughout the country. It is not surprising, therefore, that most of the water classes from this study coincided with the unimproved grassland class, with a smaller peak being located in the wetland class. Nevertheless, based on this comparison it is clear that the CSIR/ARC (1998) mapping of wetlands is poor.
Comparing the results from this study against both cumulative rainfall and the land cover maps of CSIR/ARC (1998) and Smith (2001) allowed the accuracy of the spectral mixture analysis technique to be appraised. All indications are that this technique is able to unravel
the complex water/vegetation mixtures on the Eastern Shores and, in so doing, allow accurate maps of water inundation to be produced. The protocol for performing this mapping is presented in the next section.