List of Tables
7. DISCUSSION
7.2 IMAGE PREPARATION AND TRANSFORMATION
The image preparation phase commenced with the rectification and georeferencing of the Landsat images for each study date. As recommended by Lunetta (1999) in his discussion on change detection studies, this was a two step process in which a chosen reference image was georegistered to geographic coordinates followed by an image-to-image registration that rectified the remaining images relative to the reference image. The RMS error of 0.80 pixels (24 m) obtained from the rectification of the reference image compared favourably with similar studies. Both Dewidar & Frihy (2003) and Lu et al. (2004) reported RMS errors of below 15 m while Frazier et al. (2003) obtained values of less than 20 m. Slightly larger RMS errors were obtained by Munyati (2000) (< 28.5 m) and Ringrose et al. (2003)
(42 metres). Lunetta (1999) suggested that RMS errors for the relative, or image-to-image, rectifications should not exceed 0.5 pixels. This was achieved in this study where each of the RMS error values from the relative rectification were below 0.5 pixels (see Table 6.2 on page 89). A number of other studies employing relative rectification obtained similar RMS errors, ranging from Frazier et al. (2003), who reported values of less than 10 m (0.33 pixels), to the <0.5 pixels of Boresjo Bronge (1999) and the ±0.5 pixels measured by Sader et al. (1995). The fact that the accuracies achieved during image rectification were so similar to results presented in the literature suggests that this crucial first stage in the preparation of the satellite images had been performed satisfactorily. This was an important outcome in this study as it allowed multidate images to be compared with confidence.
The preparation of the images continued with the application of radiometric correction procedures. As with the geometric rectification, results obtained were similar to those obtained by other researchers applying relative radiometric normalisation techniques. The types of areas selected as bright and dark targets closely mirrored those selected by Munyati (2000) who used irrigation reservoirs as dark targets and bare, unvegetated soil for the bright targets. As in the current study, the correlations (reported as R2 values) obtained by Munyati (2000) for the relative normalisation of three Landsat TM images were very high, namely 0.991, 0.998 and 0.999 (Table 7.1). These R2 values are comparable to those of the current study, which were all above 0.998. Eckhardt et al. (1990), in a study of irrigated lands in Nevada, used a deep clear lake as a dark target, and a dry lake bed and a badlands area as the two bright targets. The spectral characteristics of these targets were very similar to the water and beach targets used in the current study. R2 values reported by Eckhardt et al. (1990) for the normalisation of three SPOT HRV images were all above 0.998. Jensen et al. (1995) used water and unvegetated, bare soil as pseudoinvariant targets in their research into vegetation in the Florida Everglades. Their resulting correlation coefficients were all in the range 0.95 to 0.99.
Both the image rectification and radiometric normalisation produced results similar to those obtained by other studies. This is not surprising given that the same correction techniques were used. Furthermore, the fact that the results were so similar indicated that the procedures had been correctly applied and that the images had been adequately prepared for the image classification stage of this study.
Table 7.1: Comparison of correlation coefficients obtained by studies employing relative radiometric normalisation techniques.
Study Current study Jensen et al. (1995) Jensen et al. (1995) Eckhardtera/. (1990) Munyati (2000) Munyati (2000)
Range ofR2 values 0.998 to 1.000 0.96 to 0.99 0.95 to 0.99
0.998802 to 0.999000 0.991 to 0.999
0.907 to 0.988
Sensor Landsat TM Landsat MSS SPOT XS SPOT HRV Landsat TM Landsat MSS
The final procedure carried out prior to image classification involved the use of ancillary and derived data to mask out areas in which surface water was unlikely to be found.
Comparing these results to previous studies is difficult as the specifics of individual studies dictated differences in the types of masking that were used. Although a review of the literature revealed no studies that used masking in exactly the same way it was applied in this study, a few did use digital elevation models (DEMs) in a similar way. Thompson et al. (2002) devised a methodology for mapping South African wetlands that used terrain- based modelling to determine areas where water was likely to accumulate. Their terrain model used DEM-derived slope and hydrological variables to exclude non-wetland areas in much the same way that the current study used a slope criterion of 5% to mask out unsuitable zones. Another study using a DEM-derived mask was described by Sader et al.
(1995), who also used a 5% slope threshold to mask out areas too steep for water to accumulate. This method of Sader et al. (1995) was identical to the one employed in the current study. The forest mask used in this study relied on NDVI values to separate forested from non-forested areas. A similar image-based method of forest delineation was used by Sader et al. (1995), although in their case they used an unsupervised classification of Landsat TM bands 3, 4 and 5, and not NDVI values, to extract forests. In contrast, Boresjo Bronge (1999) employed a completely different approach to producing a forest mask. In this case, forest data extracted from scanned 1:5 0000 topographical maps was used to produce a forest mask that aided the classification of boreal vegetation in Sweden.
A disadvantage of this method was that it was not image-based but was dependant on the accuracy of the topographical maps.
The value of masking was emphasised by Lunetta (1999) who noted that the reduction of data dimensionality associated with masking enhanced the accuracy of image classification.
In this study, the application of slope, sand and forest masks reduced the dimensionality of the remaining spectral data to three main components, namely water, forests and grasslands. This can be seen in scattergrams of the PCA data (see Figure 5.6 on page 80), where the reduction of data to these three components is evident from the triangular pattern exhibited by the plotted pixels. While the use of PCA component data in this study was ultimately not successful, the results from the PCA analysis were important in that they verified the presence of three major components in the spectral data. This finding had important implications for the selection of endmembers in the spectral mixture analysis.