Equation 9 Pixel Height Mean Difference
5.1 SPECTRAL ANALYSIS
5.1.1 MULTISPECTRAL BAND SELECTION FOR CLASSIFICATION
5 RESULTS and ANALYSIS
Whilst the methods section has conveyed the practicality of this study approach, this chapter consolidates the observations and various outputs from the above executed methodology framework to determine shoreline changes in Kommetjie and Milnerton. The results are sectioned to address the analytics that went into both pre-processing and finalizing the results.
The spectral properties of coastal areas and land cover changes as discussed in section 2.2.1.1 on page 18 are demonstrated here both statistically and visually. The same approach is used to address the various parameters associated with multidecadal shoreline changes, as well as an effective comparison of the different coastal elevation modelling techniques towards capturing coastal zone geomorphological changes.
Figure 23 Spectral transect over Kommetjie Aerial imagery source: City of Cape Town 2018
Evidently all the bands can detect land cover change to differing extents. The Short-Wave Infrared (SWIR) band has the lowest reflectance of four over water whilst the opposite is true reaching a maximum of 144 as it transitions over land. The Green and Near Infrared (NIR) bands can detect more subtleties in the shallow coastal area than the SWIR band. They also detect the beach area at pixel 8 as opposed to SWIR’s pixel 9. Both the Green and NIR bands reach their maximum within the beach area at pixels 74 and 97 respectively, however, because they have the same general profile and the difference between them remains consistent, using them together in an index may not give the best contrast in land cover changes. At pixel 13, the profiles all level off and this could be a response from vegetation or buildings in the area.
Figure 24 Example of Landsat TM spectral signatures
The second graph in Figure 24 compares a different group of bands along the same transect.
The little variation in the thermal band suggests poor change detection capabilities, which is to be expected with it lower spatial resolution, however, it was designed to detect temperatures, which coincides with its slight increase as land cover changes from water to beach at pixel 6 like the blue and red bands (Chander et al., 2009). All the bands have a consistent lower response over the ocean with the Red band having the lowest reflectance of 12 whilst the Blue band achieves a high of 243 over the beach area. Again, it is clear that the Blue and Red bands have the same general profile which again brings into question whether they would be compatible in an index calculation
5.1.1.1 SPECTRAL INDEX SELECTION
The 4 indices tested were Normalised Differential Vegetation Index (NDVI), Normalised Differential Water Index (NDWI), Bare Land Index (BLI), Bare Soil Index (BSI). Before assessing their suitability, the 3 variations of NDWI, which is used to distinguish waterbodies in remotely sensed imagery, specified in the literature review were first investigated (Mcfeeters, 1996; Nguyen et al., 2021; Xu, 2006).
Figure 25 Comparison of NDWI formula variations
The visual and statistical effects of interchanging image bands in a differential index formula for the detection of water bodies are depicted in Figure 25 above. All three equations can differentiate between land and ocean components; however, their level of sensitivity differs.
The MNDWI equation had clearly assigned all water bodies the same value of 1 whereas NDWI and MNDWI2 can detect wave fluctuations of the ocean leading to the speckled images which may lead to spectral confusion during classification. MNDWI2 depicts land surfaces such as the beach and mountain areas in a similar tone whilst the urbanized areas appear significantly darker, the opposite is true for NDWI. MNDWI2 had the highest standard deviation of 0.364991 however it was the closest to its mean of 0.382247 meaning the land covers were more correlated and harder to distinguish. The MNDWI1 equation reaffirmed the literature review by having the greatest divergence and was subsequently carried forward to compare it with the rest of the indices shown in Figure 26 on page 91.
Figure 26 Comparison of Indices
Figure 26 above takes a broader look at the Cape Town area to compare the indices. It is abundantly clear that they can all distinguish water and land components. The results echo the opinion of the literature review as the NDVI formula depicts vegetation or land as positive values whilst water is negative (Gao, 1996; Xu, 2006). The BLI was able to discern different land covers, however could not be used because Sentinel-2 data does not have already usable thermal bands and operate based on Land Surface Temperature (LST). For the sake of uniformity, this index was disregarded going forward. The BSI misclassified the mountain shadows as waterbodies and its deviation in contrast to its mean was too large, which was due to the use of the Blue band because as second image in Figure 26 indicates, it had the highest reflectance values overall. NDVI and MNDWI were the most viable indices to use, however statistically MNDWI was able to detect more land cover variability with a higher divergence between its mean and standard deviation. For these reasons and the recommendation of past research, MNDWI was the chosen index for this study.