Monitoring ecosystem health of Fynbos remnant vegetation in the City of Cape Town using remote sensing

.................................................................................................................... I ZUSAMMENFASSUNG .................................................................................................. III CONTENTS .................................................................................................................. VII LIST OF ABBREVIATIONS .............................................................................................. IX LIST OF FIGURES .......................................................................................................... XI LIST OF TABLES .......................................................................................................... XIII


Introduction and Objectives
"The United Nations declared 2010 to be the International Year of Biodiversity" (Convention on Biological Diversity, 2010). This decision emphasises the importance and the growing attention to the health of ecosystems. "Biodiversity and ecosystems deliver crucial services to humankind -from food security to keeping our waters clean, buffering against extreme weather, providing medicines to recreation and adding to the foundation of human culture" (Nellemann & Corcoran, 2010). This study goes in line with a global trend of monitoring the health of ecosystems and restoring disturbances.
The study area of this work is the City of Cape Town Metropolitan Municipality, which is part of the Cape Floristic Region. This is the smallest, but one of the most species-rich floristic regions in the world. "It represents less than 0.5% of the area of Africa but is home to nearly 20% of the continent's flora" (UNESCO, 2010). It mostly consists of Fynbos, "an evergreen, fire-prone shrubland, confined largely to sandy, infertile soils" (Cowling, Richardson, & Pierce, 2004). With its great biodiversity and high endemism of the indigenous species, the Cape Floristic Region has long been in the focus of researchers. In 2004, the UNESCO declared eight core areas, the 'Cape Floral Region Protected Areas', as a World Heritage Site (UNESCO, 2010). Nevertheless, these areas only cover about 7% (553,000 ha, acc. to UNESCO, 2010) of the whole Floristic Region, while the rest is endangered by a number of threats.
For the natural vegetation in the City of Cape Town Metropolitan Municipality urbanisation has been identified as one of the biggest pressures. While the city boundaries are constantly spreading, the vegetation only remains in fragments. These small habitats have to cope with atmospheric deposition, atypical fire cycles and invasion of exotic plant and animal species. Because of human control, fires cannot spread naturally anymore which affects the life cycle of the Fynbos vegetation (de Klerk, 2008).
Furthermore, all biota are likely to be affected by fragmentation depending on their foraging ranges, seasonal habitat requirements, breeding success, and interactions with other fauna and flora which determine their success and survival, e.g. pollination or seed dispersal. The human influence on the remaining fragments by cattle grazing or other types of land use brings further disturbance to them.
Chapter 1 -Introduction and Objectives 2 The future prospects of the Fynbos biome are rather serious, too. Midgley et al. (2002) analysed the influence of climate change on the Fynbos biome as a whole, and for 330 species of the endemic family Proteaceae. Depending on the climate scenario, he estimated a loss of Fynbos biome area between 51% and 65% by the year 2050.
Furthermore, about 10% of the endemic Proteaceae have ranges that are restricted to the area lost.
This short overview of the situation of Fynbos vegetation points out the need for protection and management in order to conserve the remaining vegetation fragments. Several researchers have worked with remote sensing to analyse the productivity of vegetation in South Africa. Wessels et al. (2007) tested three indicators, the rain-use efficiency, the correlation between EVI and rainfall, and the residual trends method in order to distinguish between degraded and non-degraded areas. He showed that all three of them have certain potential, which is discussed in detail in chapter 5.1. Bai & Dent (2008) applied these indicators to their study about degradation in South Africa and confirmed their suitability. Since these two studies are closely related to the topic of this work, they form the basis for the methodology.
Following the introductory words of this thesis, the second chapter presents the theoretical background. The definition and parameters of ecosystem health are outlined and the fundamentals of remotely sensed data are described as far as they matter for this study. The chapter illustrates the importance of different resolutions and the scope of spectral characteristics. Furthermore, the role of remote sensing in the literature concerning the analysis of ecosystem health is delineated.
Chapter 3 gives a general overview of the study area, which is located in the City of Cape Town Metropolitan Municipality. After an introduction to the study area, local climate, geology and soil, the natural vegetation of the Western Cape lowlands is presented and it is explained why the focus lies on Sand Fynbos. Furthermore, threats that the indigenous vegetation has to cope with, especially urbanisation and invasive alien vegetation are described. Then, the common terms 'degradation' and 'improvement' are discussed on the basis of the special characteristics of the local vegetation.
Chapter 4 introduces the data used in this study and how they were pre-processed. For this purpose, the general atmospheric and geometric corrections of remotely sensed data are depicted. Thereafter, the satellite system is explained from the satellite platform, over the sensor to the used vegetation indices. The second half of this chapter presents further data used in this study, like rainfall data, GIS data, and aerial images. An important set of data described in this section is the ground truth information which was collected during five weeks of field work. The last part of this chapter summarises background information about the fragments that were gathered during the same field trip.
Chapter 5 is a compilation of the methods and indicators used in this study. The first section presents the indicators determining the productivity of vegetation. These are the rain-use efficiency, an EVI-rainfall regression, and the residual trends method. The second section describes the process of delineating the distribution of invasive species. First, the classification of the ground truth data is explained. After this, the analysis of the statistical

Ecosystem Health
The beginnings of the concept of ecosystem health can be found in the writings of the American naturalist Aldo Leopold in the 1940's. He mentioned that "ecosystems can become unhealthy, if overstressed by anthropogenic activities" (Rapport, 2007).
Especially since the 1990's, the subject came into the public focus with a growing awareness for the ecology of the planet and resulted in several studies (e.g. Johnson & Patil, 1998;Covington et al., 1997).
The term 'ecosystem health' is hard to define and because of several parallels, it is often compared to human health. For both, human and ecosystem health, applies that the definition is rather imprecise and in many cases health is only recognised in its absence (Costanza, Norton, & Haskell, 1992). Parameters such as species richness and quantity of the individuals living in the habitats are easy to collect, but key data for a 'wellfunctioning' of the ecosystem are hard to identify. Those parameters cannot just be adopted from other studies, because each ecosystem has its own key factors which have to be found and defined individually. In order to find these, one should first look for signs which differentiate unstressed from stressed ecosystems. One of these could be the loss of biodiversity, declining primary productivity or increases of invasive plant species (Rapport, 2007).
Another problem occurs if the conditions of the ecosystem change, e.g. due to human influence: The ecosystem will change too and this will affect the parameters. Thus, they have to be sufficiently dynamic in order to change accordingly with the developing ecosystem. Rapport, Costanza & McMichael (1998) defined three general measures for assessment of ecosystem health. The first one is 'vigour' or 'vitality', which can be measured in terms of metabolism or primary productivity. The second one is 'organisation', which can be assessed "as the diversity and number of interactions between system components" (Rapport, Costanza, & McMichael, 1998). The third measure is 'resilience', "the degree to which ecosystems can "buffer perturbations and maintain their basic structure and function" (Rapport, Costanza, & McMichael, 1998). An example for these three measures can be given by a case study about ecosystem health indicators in the Great Lakes Basin in North America (Shear, Stadler-Salt, Bertram, & Horvatin, 2003). With the European settlement in the 19 th century, intensive agriculture, commercial forestry and fishery established in a region, that was formerly known to have a high abundance and diversity in fish, forests and mammals. Overfishing and the intensive use of the soil for farming lead to a reduced vitality, i.e. a decline in fish abundance and infertility of agricultural soils. The organisation or community structure of the ecosystem changed with the introduction of exotic fish species. Together with overfishing and decline in water quality this caused the local extinction of native fish species and a reorganisation from nearshore benthic fish associations to offshore pelagic associations. The extinction of the native fish species for wide areas and a flip to eutrophic conditions implied a loss of resilience as the ecosystem would not be able to regenerate from these stresses anymore (Shear, Stadler-Salt, Bertram, & Horvatin, 2003).
These three established measures have to be analysed for each single ecosystem in order to find indicators that give evidence about the state of health of the ecosystem (Rapport, 2007).

Principles of Remote Sensing
Remote sensing sensors have various characteristics and are consequently used for different approaches. These basic properties are presented in the following chapter and can help during the decision progress about which remote sensing system should be used. With unique characteristics, two different sensors can provide completely different information about the same study area; hence the choice of remotely sensed data is very important.

Scale and Resolution
The extent of the study area and the resolution of the sensor are two almost antagonistic factors, which define the information content that can be extracted from the data. With an increasing extent, the amount of retrieved data increases too, which can also result in higher costs and time involved. In this case, it is advisable to reduce the level of detail in order to maintain the effort on an acceptable level. Remote sensing has four different types of resolution, which are described in the following.
Spatial Resolution is defined as "a measure of the smallest angular or linear separation between two objects that can be resolved by the remote sensing system" (Jensen, 2005).
Each remote sensing sensor has a certain spatial resolution, usually defined by the pixel size, i.e. the length and width of the smallest unit of an image. For example, the Landsat 7 Enhanced Thematic Mapper Plus (ETM+) has a spatial resolution of 30 x 30 m² for its six multispectral bands, 60 x 60 m² for its thermal band and 15 x 15 m² for its panchromatic band. The smaller the spatial resolution, the greater is the level of detail and the easier to discriminate between certain objects. On the other hand, smaller spatial resolution also means more pixels per image and thus higher amount of data that has to be stored (Jensen, 2005).
Temporal Resolution is the inherent repeating coverage of the same area on the earth (Schowengerdt, 2007). Multiple records of the same area for different points in time are useful to map the changes that this area is subject to. "Ideally, the sensor obtains data repetitively to capture unique discriminating characteristics of the object under investigation" (Jensen, 2005). Vegetation, for example, has unique phenological cycles in different geographic regions and depending on its composition. The phenology of vegetation is discussed in section 2.2.3.

Radiometric
Resolution is defined as "the sensitivity of a remote sensing detector to differences in signal strength" (Jensen, 2005). Some remote sensing systems have a higher precision in measuring electromagnetic radiation than other systems. It is measured as the number of gray-scale levels and typically recorded in bits (binary digits). Spectral Resolution is the number and dimension of specific wavelength intervals (called bands or channels) in the electromagnetic spectrum to which a remote sensing instrument is sensitive (Jensen, 2005). The higher the number of bands and the smaller the band width, the higher is the spectral resolution of the sensor. In general, a sensor is called multispectral if it has about 10 bands and hyperspectral if it acquires data in hundreds of spectral bands (Jensen, 2005). The measurements of all spectral bands combined for one location point on the earth are called spectral signature. A land cover type can have a typical or at least similar spectral signature in different regions of the earth (Lillesand, Kiefer, & Chipman, 2004). In the following section, spectral signatures of typical land cover types are described.

Spectral Signatures
If the spatial resolution of a sensor is too low, the colours of the image seem blurry and objects cannot be identified by their shape or spatial detail. In this case, it is often useful to extract the full spectrum of brightness of that area or object to identify it with the help of its spectral signature (also referred to as spectral profile). Although it is often possible to distinguish different types of materials, the distinction is sometimes foiled by factors like natural variability for a given material type, coarse spectral quantisation or modification of signatures by the atmosphere (Schowengerdt, 2007).
Averaged spectral signatures for three basic land cover types are shown in figure 2.1. The x-axis identifies the wavelength or in some other illustrations the number of the individual bands; the y-axis documents the percent reflectance if the data have been calibrated (otherwise the total brightness value) (Jensen, 2005).
Water is probably the easiest land cover type to deduce from its spectral signature. The most distinctive characteristic of water is the energy absorption at near-infrared wavelengths and beyond (Lillesand, Kiefer, & Chipman, 2004). Depending on the turbidity of the water, changes in spectral signature can be dramatic. Chlorophyll concentration or suspended sediment content are the major factors of influence on its spectral profile (Lillesand, Kiefer, & Chipman, 2004).  (Lillesand, Kiefer, & Chipman, 2004).
Bare Soils commonly have spectral signatures that increase quite steady up to 1.8 µm with wavelength. (compare figure 2.1), because the factors that influence soil reflectance act over less specific spectral bands (Lillesand, Kiefer, & Chipman, 2004). The major factors affecting its reflectance are moisture content, soil texture, surface roughness, presence of iron oxide, and organic matter content. For example, soil moisture or the presence of iron oxide decrease the reflectance of the soil while a coarse texturing of a soil increases its reflectance (Lillesand, Kiefer, & Chipman, 2004).
The chlorophyll in green vegetation strongly absorbs energy in the blue and red wavelength bands (referred to as chlorophyll absorption bands) but reflects green radiation and near-infrared. Hence, our eyes perceive healthy vegetation as green in colour (Lillesand, Kiefer, & Chipman, 2004). If the plant is subject to some sort of stress disturbing its normal productivity, the chlorophyll absorption bands are less distinctive. As a result, we see the plant yellow combining green and red reflectance. Going further from the visible to the near-infrared portion of the spectrum at about 0.7 µm, the reflectance of green vegetation increases dramatically. This significant change is called red edge and occurs slightly shifted in stressed vegetation towards shorter wavelengths (Jensen, 2005).
Dips in reflectance occur at 1.4, 1.9, and 2.7 µm because water in the leaves strongly absorbs radiation at these wavelengths. These spectral regions are called water absorption bands (Lillesand, Kiefer, & Chipman, 2004).

Vegetation Phenology
Another possibility besides spectral signatures to distinguish different land cover types, especially different types of vegetation, is the phenology. This is defined as the temporal development of the vegetations' spectral information over the course of the year.
Vegetation grows according to relatively predictable seasonal and annual phenological cycles (Jensen, 2005). During winter or the dry season, it is often easy to distinguish annual from perennial species (compare figure 2.2). Especially if the analyst does not have remotely sensed data with a high temporal resolution for plotting the whole phenological cycle, it is necessary to know the biophysical characteristics of the vegetation in order to choose the right time of the year for the investigation (Jensen, 2005). Carolina. (Jensen, 2005).
In addition to the distinction between different vegetation species, it is also possible to detect changes in annual productivity of a certain species with the help of phenological cycles. For this purpose, correlations between the vegetation phenology and factors that affect the phenology, usually precipitation, are calculated. If differences in annual productivity cannot be explained by rainfall, this might indicate anthropogenic influence, fires etc. (Fabricante, Oesterheld, & Paruelo, 2009). In order to draw the right conclusions, one has to have certain background information about the study area.

Ecosystem Characterisation using Remote Sensing
Since the launch of the first satellite carrying a multispectral sensor, Landsat 1 in 1972, vegetation dynamics has been one of the primary tasks of remote sensing. Several sensors were only developed to display the unique characteristics of vegetation with channels for e.g. the red edge or the water absorption bands (Jensen, 2007).
Especially the analysis of the productivity of vegetation has led to many publications (e.g. Curran, 1981;Holm, Cridland, & Roderick, 2003;Li, Lewis, Rowland, Tappan, & Tiszen, 2004). Holm, Cridland, & Roderick (2003) compared biomass production with the remotely sensed Normalized Difference Vegetation Index (NDVI) in Western Australia and determined that the NDVI can be a good surrogate for ground measuring. They also analysed the relationship between NDVI and rainfall. For this purpose, Holm calculated the rain-use efficiency (RUE), the ratio between annual NDVI and annual rainfall. Today, this is a commonly known indicator and has also been used in Argentina (Guevara, Estevez, & Torres, 1996), the West African Sahel (Nicholson, Tucker, & Ba, 1998), South Africa (Bai & Dent, 2008;Wessels et al., 2007) and many other regions of the world.
Another possibility to analyse the vegetation-rainfall relationship is the correlation or regression of these two variables. Fabricante, Oesterheld, & Paruelo (2009) used NDVI data from the AVHRR sensor in North Patagonia and correlated it with precipitation data from rainfall stations. They observed high correlations of NDVI with precipitation accumulated during a few months of the previous growing season. Li et al. (2004) came to similar conclusions for different vegetation types and discovered several degraded areas in the Senegal.
A further indicator for the analysis of productivity is the residual trends method which was introduced by Evans & Geerken (2004) and became more and more popular in recent years (Herrmann, Anyamba, & Tucker, 2005;Wessels et al., 2007;Bai & Dent, 2008). This method is based on a regression between NDVI and rainfall and negative trends may indicate a degradation of the land. Evans & Geerken (2004) applied the residual trends method in Syria with AVHRR data and found it a useful indicator for human influences although the direction of the trend has to be analysed carefully. Wessels et al. (2007) and Bai & Dent (2008) also utilised this indicator but since this work is based on their research, chapter 5.1 presents a more detailed view on their analysis.
This short review displays that the productivity of vegetation seems to be a suitable measure of ecosystem health that can be analysed by remote sensing. Thus, an application in this work assessing the general measure 'vitality' by Rapport, Costanza, & McMichael (1998) was intended.
Furthermore, remote sensing has received considerable interest in the field of biological invasion in the recent years. Invasive plants may rapidly decrease the biodiversity of an ecosystem and thus there is a need for large-scale monitoring of endangered areas.
Between 1990 and 2000 the number of publications on this topic has grown from about 20 to 80 publications per year (Joshi, Leeuw, & Duren, 2004). The detection and delineation of invasive alien species can be based on their unique spectral or phenological properties, structural characteristics, or the spatial patterns of infestations (Strand et al., 2007). Asner, Jones, Martin, Knapp, & Hughes (2008) studied the discrimination between native and invasive species in Hawaiian forests on the basis of their spectral characteristics. They used an Airborne Visible and Infrared Imaging Spectrometer (AVIRIS) and discovered differences in leaf and canopy properties like water content and pigmentrelated absorption features. Thus, they were able to differentiate between 7 native and 24 introduced tree species, the latter group containing some highly invasive species.
Another application of remote sensing in California focuses on the delineation of the invasive weed yellow starthistle with hyperspectral imagery (Miao et al., 2006). Knowing the unique phenology of the plant, the starthistle was delineated in the earlier growing season because it becomes greener more quickly.
Focusing on the structural characteristics, Rosso, Ustin, & Hastings (2006) used lidar to delineate the distribution of Spartina species in the San Francisco Bay marshes. Lidar is related to radar using a laser beam instead of radio waves and can detect small differences in the height of vegetation. The study displayed the expansion patterns of Spartina and indicated the great potential of lidar for the analysis of wetland topography.
As the following chapter 3.5 presents, the delineation of invasive alien species is an important task in the study area. High densities of alien species dramatically reduce the biodiversity of an area. Thus it is analysed in this study addressing the second of the three general measures 'organisation' by Rapport, Costanza, & McMichael (1998).

Study Area
The study area is situated in South Africa's Western Cape Province, in and adjacent to the  Eight sections with sixteen test sites (depicted in red in figure 3.1) were selected from the study area. These were specified by SANBI being the major Sand Fynbos fragments. Table   3.1 gives an overview of the different fragments with their size, protection status and in which section they are located. The boundaries and the protection status were derived from the shape file about the biodiversity network (for more information see chapter 4.7). Eight of the fragments are formally protected and the others are intended to follow. actually consists of three fragments, they were not sampled during the field trip and hence summarised without unique names. The Blaauwberg Conservation Area (BCA) is actually a bigger fragment, but since only the eastern part was said to contain Sand Fynbos, the rest was omitted.
Detailed descriptions of the fragments are found in chapter 4.6 containing the information gathered during the field trip.

Climate
According to the Koeppen-Geiger Climate Classification, Cape Town has a Mediterranean climate (Csb) with mild, wet winters and dry, hot summers (Kottek, Grieser, Beck, Rudolf, & Rubel, 2006). This is unique in Southern Africa and is caused by the pole-ward migration of the Hadley Cells. During the hot summers, from November to March, frequent tradewinds occur. These are strong winds, named Southeaster after their main wind direction.
In winter, north-westerly winds bring most of the annual rain to the Cape Flats with an annual average of about 500 mm rainfall. Due to the mountainous landscape of Cape Town with the Table Mountain and the Cape Fold Belt, orographic rainfall occurs in these areas, especially in spring and autumn, which causes high local differences in the mean annual rainfall (Cowling, Richardson, & Pierce, 2004).
The mean annual temperature is 16.2 °C for the predominant Cape Flats Sand Fynbos

Geology and Soil
To understand the geomorphology, today visible at the Cape Peninsula and Cape Flats, it requires a little background. About 450 million years ago, the region was a large coastal plain with several river deltas. During this time, the Table Mountain Group sandstones were sedimented and lithified by pressure. When the continents collided to form the supercontinent Pangaea about 250 million years ago, these sediments were folded and raised to create the Cape Fold Belt. This mountain range was much higher than it is today, visible as a dashed line in figure 3.3 (Compton, 2004).  (Compton, 2004).
Theory says that today's landscape might be an inverted version of the past. The Cape Flats were once a mountain ridge of the Cape Fold Belt and the visible remnants of the Table Mountain Group sandstones, which are the Table Mountain itself  Since the Tertiary, the sea level fluctuated -120 to +200m from the present mean sea level. Thus, the Cape Flats got flooded several times and marine sediments accumulated on the Flats (Compton, 2004).
The soils of the study area are commonly nutrient-poor sandy soils. These sands can be aeolian or fluvial deposits and depending on location and lime content, they can be acidic or neutral. In some parts of the area, the underlying Malmesbury Shale or the Cape Granite leads to a clay-rich soil.
The geologic diversity and the resulting soils are an important factor for the evolution of the botanic diversity in the study area (Mucina & Rutherford, 2006).

Natural Vegetation
There are three major vegetation types in the study area: Fynbos, Renosterveld and Strandveld. In this work, the focus lies on one type of Fynbos vegetation, Sand Fynbos.
This special vegetation type is due to its rarity of high conservation interest. Furthermore, the decision to focus on Sand Fynbos was made in order to eliminate as many variables as possible since it was expected that the other vegetation types of the study area have different phenologies. However, in the following will be given a short introduction to all three vegetation types in order to understand the differences. Many of the Proteaceae are adapted to fire and only release their seeds after burning, when the general vegetation cover is low and the soil is well fertilised (Mucina & Rutherford, 2006).
The major type of Fynbos in the study area is Sand Fynbos, which occurs on "acidic tertiary, grey regic [≙ undeveloped soil] sands" (Mucina & Rutherford, 2006). It is dominated by dense ericoid shrubland with emergent proteoid shrubs and restios. The latter are the prevailing type in areas that are seasonally waterlogged.

This work focuses on Sand Fynbos vegetation which divides into Atlantis Sand Fynbos and
Cape Flats Sand Fynbos in the study area. The northern part of the study area is usually Atlantis Sand Fynbos, while the southern part is Cape Flats Sand Fynbos. Since the differences are rather minor in the study area, these are not explained any further (Mucina & Rutherford, 2006). Renosterveld occurs on the clay-rich soils derived from shale and granite. It is an "evergreen, fire-prone shrubland or grassland dominated by small cupressoid-leaved evergreen asteraceous shrubs (principally renosterbos) with an understorey of grasses (Poaeceae) and a high biomass and diversity of geophytes" (Mucina & Rutherford, 2006).
In comparison to Fynbos, it is lacking Proteaceae and Ericaceae and should have less than 5-10% cover of Restionaceae. Because of the dominance of fine, fast-growing grasses, the fire frequency (3-5 years) is higher than it is in Fynbos. Renosterveld may transform to Fynbos, where annual rainfall increases and leaching leads to a nutrient-poor soil.
Strandveld consists of medium dense to closed shrubland with a succulent character in the arid areas. It occurs on mineral-rich soils usually close to the sea, but never under direct influence of sea spray. "Unlike in Fynbos or Renosterveld, fire plays a lesser role in the Strandveld communities" (Mucina & Rutherford, 2006). The main reason is the succulent nature of the Strandveld vegetation, impeding the spread of fire (Mucina & Rutherford, 2006).
In the following chapters, the term 'Fynbos' is used as an abbreviation for 'Sand Fynbos'.

Threats for the local Ecosystems
The natural vegetation of the study area has faced and is still facing many threats -most of them are man-made. Some of them directly affect the Fynbos vegetation like urban sprawl or invasive vegetation and some of them are more subtle like climate change.
In the following, the two major threats to Fynbos vegetation in the study area besides climate change are presented.

Urbanisation of Cape Town
In the past half century, Cape Town has undergone some major changes. In 1950, the southern half of the study area was more or less uninhabited, but during the following apartheid era it got populated by the black community being forced out of the city centre by the government. In the following years, the townships and existing villages, today known as the Southern Suburbs, expanded and grew together to cover most of the southern Cape Flats (Wilkinson, 2000).
The northern half of the study area is only populated with some smaller villages, but agriculture plays a major role here. During the last century, the growing demand for food led to a massive extension of the agricultural area and thus to a shrinking of the natural vegetation. Grain and wine are the major cultivation products and stock farming is quite common (Wilkinson, 2000).
Between 1985 and 2005, the physical extent of Cape Town grew by 40%. The rate at which the city is developing almost doubled in the last thirty years and is now increasing at an average rate of 1232 hectares per year (Spatial Planning and Urban Design Department, 2009).

Invasive Vegetation
Beside the rapid decrease of the area with potential natural vegetation due to agriculture and urbanisation, the remaining vegetation fragments are under heavy pressure by the invasion of alien vegetation.
In the early nineteenth century, colonists introduced a wide variety of plants to South Africa. At first, these were predominantly agricultural crops. With the spread of the settlement an increasing need for firewood and building materials occurred. The local Fynbos vegetation, which lacks trees and grows slowly, could not fill this role. For this reason, tree species were brought to the country, especially from Australia and South America (Cowling, Richardson, & Pierce, 2004).
"The drift-sands on the Cape Flats [...] posed an especially vexing problem. They had for years resisted attempts at control and made travel across them hazardous and unpleasant" (Day, Siegfried, Louw, & Jarman, 1979). In order to stabilise the dunes and for the production of firewood, Acacia saligna (compare figure 3.5) and Acacia cyclops were introduced in the 1830s from Australia. For the need of building materials, different pine species were brought to South Africa. To date, 113 naturalized alien grass species, especially from Europe were introduced to South Africa; some of them on purpose, e.g.
for pasture or as ornamental grasses and some by accident with imported seeds or wool (Milton, 2004). Musil, Milton, & Davis (2005) hold that the invasion by alien grasses is still underestimated compared to other invaders. However, a lot of different plant species were brought to the country, but no other species affected the natural vegetation of the study area till today like the two mentioned acacia species (Day, Siegfried, Louw, & Jarman, 1979).
Acacia cyclops and Acacia saligna have been the most extensive invaders of Fynbos reserves. According to a survey (1983)(1984)(1985) documented by Cowling, Richardson, & Pierce (2004), Acacia cyclops covered 52.6 % and Acacia saligna 37 % of all grid cells tested. Today, especially Acacia saligna is regarded as "the most important invasive weed in the Cape Floristic Region" (Maslin & McDonald, 2004). It is well adapted to the dry and nutrient-poor soils of the Cape Flats. The main root may grow to 16 m deep (Maslin & McDonald, 2004) and it is able to fix atmospheric nitrogen (Marsudi, 1999). This nitrogenfixing attribute can change the whole soil regime. Yelenik, Stock & Richardson (2004) have shown that long-term invasions and as a consequence the enrichment of the soils not only hinder the recovery of Fynbos after a clearing, but also lead to higher growth rates of a weedy grass species. This may cause serious problems for restoration of the native vegetation.
In order to control the invasion of Acacia saligna, different approaches were applied. The most obvious one is to cut down the trees or burn them. Depending on the size of the area, there are three general ways to do this: 'burn standing', 'fell and burn' and 'fell, remove and burn'. Holmes, Richardson, Wilgen & Gelderblom (2000) suggest that 'burn standing' is the best treatment to promote Fynbos recovery, since it causes the least change to vegetation variables.
Another approach is the introduction of biological control agents, i.e. natural enemies of the invaders. 73 species of herbivorous natural enemies have been introduced to South Africa to control the dispersal of all kinds of alien plants (Cowling, Richardson, & Pierce, 2004). In 1987, the rust fungus Uromycladium tepperianum was introduced to control the Acacia saligna and established throughout the stands of the weed. The fungus releases hormones causing the tree to rapidly produce galls with an average diameter of 1-2 cm (Serdani, 2001). The infection with U. tepperianum itself does not kill the tree, but the high gall production stresses it to such an extent that it cannot cope with other environmental stresses like droughts and dies or shows reduced vigour especially in seed production (Wood & Morris, 2007).
The seed-feeding beetle Melanterius compactus has been introduced in 2001 to increase the pressure on the acacia populations (Wood & Morris, 2007). Statistics about the success of the beetle are outstanding to this date.
Despite great successes with these biological control agents, it is still necessary to cut down and/or burn the acacia to completely clear invaded areas. This requires good management over several years.

Monitoring 'Degradation' and 'Improvement' in the Context of the Study Area
Land degradation is one of the most common topics in ecological research (e.g. Blaikie & Brookefield, 1987;Meadows & Hoffman, 2003;Pickup, Bastin, & Chewings, 1998). It is generally defined by the FAO as the "temporary or permanent decline in the productive capacity of the land" (Stocking & Murnaghan, 2001). Most of these definitions refer to land as an economic resource and link degradation with the term 'land use'. Thus, land degradation is usually indirectly measured as the value for human use. For example Bai & Dent (2008) primarily define land degradation "as a long-term decline in ecosystem function" and then cut it down on the measurement of net primary productivity (NPP) with degradation as a decline of NPP and improvement as an increase in NPP. This might be appropriate for a nationwide or global approach but chapter 2.1 already depicted that this may not be appropriate for the analysis of complex ecosystems.
After an analysis of the conditions of the study area in combination with the general behaviour of EVI it becomes apparent that the general approach of Wessels et al. (2007) and Bai & Dent (2008), which form the basis for this study, have to be modified. Since the focus is on areas with natural or semi-natural vegetation and rangelands are excluded, it is not reasonable to look at these from an economic point of view. Measuring the value of a natural area for human benefit is rather vague and not part of this study.
As mentioned in the previous chapter, a main problem of the study area is the invasion of alien vegetation, especially of Acacia saligna and alien grasses. Because degradation through direct anthropogenic influences is rather minor on the fragments, the invasion of alien species is the main problem and thus the first indicator for degradation of Fynbos vegetation. In this context, alien plants as well as an exceedingly high cover of grasses indicate degradation, since the grass fraction of pristine Fynbos vegetation is rather minor.

Data and Data Pre-Processing
In this work, a combination of remotely sensed data, GIS data and precipitation data was used for the analysis. In addition, ground truth data was collected during field work in January -February 2010. The data and the way it was processed are presented in the following section.

Satellite Data
The remotely sensed data used in this study is described in the following. The basis for this remote sensing approach is the data of the MODIS sensor on the platform Terra. This data is pre-processed by NASA and different vegetation indices are calculated and provided online.

Terra -Satellite
The satellite Terra, also known as EOS AM-1, was launched in December, 1999 by the American NASA and started to collect data in February, 2000 (NASA Science, 2010). It is flying in a 705-km (on the equator) high sun-synchronous orbit and crosses the equator each day at 10.30 a.m. from north to south at an inclination of 98° (Lillesand, Kiefer, & Chipman, 2004). Table 4.1 gives an overview of Terra's characteristics.

MODIS -Sensor
The MODerate Resolution Imaging Spectrometer MODIS is installed on both Terra and Aqua satellites. It provides long-term observations to derive an enhanced knowledge of global dynamics and processes of land-cover, ocean and the lower atmosphere (Jensen, 2005). With its one to two days global repetition rate, it is eminently suited for questions about the phenological cycle of vegetation. Thus, this sensor was chosen for the study in order to correlate its data with constant rainfall data.
MODIS is a whiskbroom scanner, collecting signals with a rotating mirror along scan lines at right angles to the flight line. Thus, it has a wide field of view of ±55° off-nadir (the base point opposite to the zenith) and reaches a swath width of 2,330 km. As mentioned in the theoretical background, it has a high radiometric resolution of 12 bit. Collecting data in 36 spectral bands, it also offers a good spectral resolution (compare table 4.2). 20 of these bands detect solar radiation in the range of 0.4 to 3 µm and the other 16 bands collect thermal signals from 3 to 15 µm (Jensen, 2005). Depending on the band, MODIS' coarse spatial resolution ranges from 250 x 250 m² for the bands 1 and 2 to 500 x 500 m² for the bands 3 through 7 and 1 x1 km² for the other bands from 8 to 36. The MODIS data have one of the most comprehensive calibration settings and are characterized by improved geometric rectification and radiometric calibration (Lillesand, Kiefer, & Chipman, 2004). Furthermore, the products about the land surface are already atmospherically corrected. With these stringent calibration standards, a large variety of data products is derived from this sensor with the following main application areas (after Lillesand, Kiefer, & Chipman, 2004): • Cloud masks for the assessment of climate and potential climate change.
• Aerosol concentration and optical properties.
• Cloud properties like optical thickness and effective particle radius.
• Vegetation and land surface cover, several kinds of atmospherically corrected vegetation indices like net primary productivity, leaf area index, normalised difference vegetation index and enhanced vegetation index.
• Snow and sea-ice cover and reflectance.
• Surface temperature with an absolute accuracy goal of 0.3 to 0.5 °C over oceans and 1 °C over land.
• Ocean colour, spectral radiance (visible and near-infrared bands) leaving the ocean.
• Concentration and fluorescence of chlorophyll in surface water.
These products are available online at no charge, e.g. at the USGS Global Visualization Viewer (USGS, 2010).

Vegetation Indices
In this study, the Enhanced Vegetation Index (EVI) is used to determine the relationship between primary production and rainfall. The EVI is part of the MODIS product MOD13Q1, which also contains a second vegetation index, the so-called Normalised Difference Vegetation Index (NDVI). Both indices are in a 250 x 250 m² spatial resolution and are delivered as a 16-days composite, which means there are 23 composites, socalled time steps per year. As mentioned in the previous chapter, MODIS actually has a global repetition rate of one to two days, but because of effects like cloud cover and viewing geometry these data are not always useful. Thus, the observations are collected over 16 days and composited by the MODIS VI algorithm applying a filter based on quality, cloud, and viewing geometry (Huete et al., 2002).
Since the NDVI forms the basis for the EVI, it is described in the following, too.

NDVI
The NDVI is commonly known as a remotely sensed surrogate for net primary production (NPP) and has been used in numerous studies (e.g. Wessels et al., 2007;Bai & Dent, 2008;Fabricante, Oesterheld, & Paruelo, 2009). It is one of the most important vegetation indices because it is able to monitor "seasonal and inter-annual changes in vegetation growth and activity" (Jensen, 2005). Taking advantage of the red edge in vegetation signatures (compare chapter 2.2), the NDVI is calculated as follows (Jensen, 2005): where: The possible range of values for the NDVI goes from -1 to 1. Areas with a high vegetation cover will generally yield high values for the index having high near-infrared reflectance and low reflectance in the red band. With a larger reflectance in the visible bands than in the near-infrared band, clouds, water, and snow yield negative index values. Areas with open ground and rocks tend to have similar reflectances in both bands and result in an NDVI near zero (Lillesand, Kiefer, & Chipman, 2004).
However useful the NDVI is, it should be pointed out that it has certain disadvantages.
This ratio-based index is nonlinear and can be affected by certain noise effects such as atmospheric scattering. Another problem is the saturation effect in high-biomass regions like forests. The NDVI dynamic range is compressed for these high-biomass regions and rather stretched under low-biomass conditions. The third disadvantage of the NDVI is its sensitivity to canopy background such as visible soil. The NDVI values seem particularly higher with darker canopy backgrounds (Jensen, 2005). This also makes it a rather "poor indicator of vegetation biomass if the ground cover is low, as in arid and semi-arid regions" (Schowengerdt, 2007).

EVI
The EVI was developed by the MODIS Land Discipline Group especially for the use with MODIS data. As the name suggests, the EVI is an enhanced version of the NDVI and is calculated as follows (Schowengerdt, 2007): where: The bands ρ NIR , ρ red and ρ blue are atmospherically corrected or partially corrected (Rayleigh and ozone absorption) surface reflectances. L is the soil adjustment factor, is a gain factor and the two coefficients 1 and 2 describe an atmospheric correction using the blue band to correct the red band. Huete et al. (2002) has shown that the EVI has an improved sensitivity to high biomass and does not saturate as easy as the NDVI.
Furthermore, atmospheric influences are reduced for the EVI and the monitoring of vegetation is improved through a decoupling of the canopy background signal (Jensen, 2005;Huete et al., 2002). Because of these advantages over the NDVI, the EVI was chosen for this study. The whole times series from the start of the data collection in February 2000 till the end of 2009 was used.

Pre-Processing of Remotely Sensed Data
Depending on the remote sensing system, the data is usually not delivered 'ready-foruse'. Several pre-processing steps like atmospheric correction or georectification have to be done in order to use it for further applications. In some cases, these pre-processings are already calculated by the provider, in some not, but they always have to be done first.

Atmospheric Correction
While travelling through the atmosphere from the earth's surface to the sensor, the electromagnetic radiation signals are modified by scattering and absorption by gases and aerosols (Song, Woodstock, Seto, Pax Lenney, & Macomber, 2001).
1. Path radiance (L1). This is light scattered from the atmosphere and received by the sensor without having made ground-contact.
2. Reflected radiation (L2). Direct and diffuse solar radiation is reflected by the pixel i.e. the earth's surface and transmitted to the sensor. The diffuse solar radiation is light scattered in the atmosphere so that it is redirected to the pixel and reflected towards the sensor.
3. Adjacency radiance (L3). Solar radiation is reflected by neighbouring pixels and scattered by the atmosphere right in the direction of the sensor.
The task of atmospheric correction is to remove the effects of the path radiance, the adjacency radiance and the diffuse solar radiation in order to retrieve the information from the currently viewed pixel contained in L2 (Richter, 2010).
Atmospheric correction is not always necessary, depending on the type of analytical method used. For certain classifications, which only use a single date, it is unnecessary to atmospherically correct them as long as the training data and the data to be classified are in the same relative scale (Jensen, 2005).

Geometric Correction
Geometric Correction is the second important type of correction and is usually necessary because the remotely sensed data are not in their proper geometric x, y location (Jensen, 2005). The remotely sensed data typically exhibits internal and external geometric errors, i.e. distortions of the image (Jensen, 2005). Internal geometric errors are mostly systematic distortions and can be caused by the earth's rotation or the scanning system itself. Those distortions are usually identified and corrected rather easy. On the other side, external geometric errors are caused by changes in altitude or attitude of the sensor. If the remote sensing platform, i.e. the aircraft or spacecraft gradually changes its altitude or if the elevation of the ground changes, the scale of the imagery changes, too (Jensen, 2005). While satellite platforms are not affected by atmospheric effects, aircraft platforms constantly contend with atmospheric turbulence or wind and thus change their attitude, i.e. axial position or pitch up and down (Jensen, 2005). These geometric distortions can be corrected using ground control points (GCP) and appropriate mathematical models (Jensen, 2005).

MODIS Processing
Since the MODIS EVI data is calculated with an atmospheric correction, it is not necessary to correct them any further. However, some processing steps have to be done in order to use and correlate them with the rainfall data. At first, MODIS data always have to be Subsetting was used to reduce the size of the MODIS data to the area of interest, i.e. the study area. Resampling changes the spatial resolution (pixel size) of the data. The exact pixel size of the MODIS EVI 250m product is 232m, so the data was resampled to this size.
Because these MRT processings have to be done for each single file, two small IDL (Interactive data Language) scripts were used in order to automate these functions.
Furthermore, it is essential to analyse the quality of the whole time series used in this study and correct errors if necessary. For this purpose, Colditz et al. (2008) developed a tool called Time Series Generator (TiSeG), which uses the quality-assurance science data set included in the MOD13Q1 product. The time series is analysed for outliers, which can still be contained in the data despite the corrections during the pre-processing of the provider. These outliers or as they are called in the tool 'invalid pixels' may result from insufficient atmospheric correction and prevail during the rainy season (approximately April -September for Cape Town), while the cloud cover is dense. The invalid pixels are temporally interpolated after certain rules that have to be set by the user. For detailed information on the tool and its parameters see Colditz et al. (2008). In general, the interpolation rules for the rainy season are set a bit more flexible so that bigger gaps of invalid values can be filled. The whole EVI time series used in this study was processed with this tool. Finally, the data format has to be converted to a more practicable format. The MODIS data are delivered in the so called Hierarchical Data Format (.hdf). Since most of the remote sensing programs have their problems with this format, the data were converted to the common GeoTIFF format. Figure 4.2 shows an example for a readily processed EVI scene for the year 2008.

Rainfall Data
The rainfall data were provided by the South African Weather Service (SAWS) from 19 rainfall stations. These data were delivered in Excel-format for the years 1999 to 2008.
Each rainfall station was listed in annual sheets containing daily precipitation records with rows for the days and columns for the months. In order to integrate these into a Geographic Information System (GIS), the data had to be modified. These modifications were implemented by me in an IDL (Interactive Data Language) script (compare appendix A.2). First, the script reads in the precipitation data from the Excel file and then the annual sheets are converted so that the daily records of the whole time series are lined up in a column for each rainfall station. Furthermore, the daily records are aggregated to 16-days sums in order to match the MODIS data. Finally, the script exports the converted data to a text file. Afterwards, a point shape file was generated manually with the software ESRI ArcGIS 9.3 using the coordinates of the climate stations. The rainfall records were imported from the text file to the GIS and linked with the shape file. Finally, each 16-days sum was interpolated to a raster in ArcGIS using the Natural Neighbour method.
232m, the same cell size as for the EVI data was chosen and the output files were snapped to the EVI data so that they perfectly match them. Furthermore, the 16-days sums were aggregated to annual sums of rainfall for the processing of the rain-use efficiency. Figure 4.3 shows the readily processed annual rainfall sum of 2008.

Aerial Images
The The images were interpreted visually and used to select the pixels for ground truthing.

Ground Truth Data
The ground truth data acquisition was performed between 11th January and 14th February 2010, which is towards the end of the dry season.
After some consultations with the local botanists at the SANBI office, a pre-selection of the fragments had to be done. Because of diverse reasons, not all sites could be visited. It turned out that for Tokai State Forest, the area under potential Fynbos vegetation is too small for a serious remote sensing analysis. Furthermore, it was recommended not to visit the Haasendal fragment due to safety issues in the area. The fragments Plattekloof and Kenilworth Race Course were also eliminated from the schedule because of time and size issues. In addition, the Plattekloof fragment with its long and narrow shape is not even covering the width of one MODIS pixel and thus its signal must be influenced by neighbouring effects of the surrounding built-up area.  In order to improve the estimation, the sampling area was divided into four 10 x 10 m² squares. The vegetation types were documented in detail, so they could be aggregated to more coarse vegetation groups during the appraisal. Afterwards, the percentage cover of the four 10 x 10 m² sub-squares and then the cover of the five 20 x 20 m² squares were aggregated to one percentage cover for each GTP.
With this procedure, it was intended to get a representative vegetation cover of the surrounding area. Although, the centre of a pixel was always chosen for the sampling of the GTP, it was considered that the pixel may have a certain shift due to geo-correction issues. Therefore, pixels in homogenous areas of the fragment were preferred for the GTP in order to cope with a possible shift.
For the rather homogenous fragment Riverlands Plain, the sampling procedure was simplified and a transect of nine pixel middle points was collected.
This way, 21 GTPs were sampled and a total number of 126 sample points were collected (compare table 4.3).  b. Riverlands Plain is the easternmost sub-fragment and named by the author of this work due to its flat grassy appearance. Some sparse, young acacia stands can be found, but they are obviously cut down on a regularly basis.

Additional Information about the Fragments
c. Riverlands Floodland is also named by the author. Garden Cities, a development company, but not used for any certain purpose and thus no records exist. However, according to the Friends of BCA, a non profit organisation for the health of the BCA, the area has not been burnt the last ten years.
A spot in the northeast of the actual study area was cleared by the friends of the BCA in 2005 and seems to be covered with intact Fynbos vegetation to date.
Unfortunately, the area is too small to include it in the analysis.
4. Rondebosch Common is a conservation area in the middle of the city area. This fragment is protected for a long time but since it is surrounded by dense city area, it is still used as a local recreation site and for walking dogs etc. It is mainly covered with grasses, herbaceous plants and a few shrubs. Joanne Eastman, a member of the Friends of the Rondebosch Common, assured in an interview that the last clearings of acacia happened in the 1990s (compare appendix A.5). At that time infestations of alien grasses were also cleared and some more during the last ten years but not as severe as it was in the 1990s. The last fire was in 1999 but did not burn the whole Fynbos is too small for the analysis with remote sensing and was thus omitted from the study.
10. Schoongezicht is a family-owned farm and is said to have some intact Fynbos vegetation in the northern part. Since this fragment was recommended by the City's Environmental Management Branch after the field work was undertaken, it could not be sampled.
In general, the background information about the fragments is very sparse which makes an interpretation of the indicators for the whole time series difficult. Thus, this work focuses in the following chapters on the analysis of the most recent data and the time series became secondary.

Additional GIS Data
In addition to the self-made shape files for the ground truth data and the rainfall stations, several other datasets have been used in this study.

Methodology
As mentioned in chapter 2.1, ecosystem health is a very complex issue and indicators have to be found separately for each area of interest. Rapport (2007)

Productivity of Vegetation
Since the methodology of this study is based on the papers by Wessels et al. (2007) and Bai & Dent (2008), three common methods were used to analyse the productivity of the vegetation: the rain-use efficiency, a regression between rainfall and EVI and the residual trends method (RESTREND).

Rain-use Efficiency
The rain-use efficiency has been used in numerous studies as an index of degradation (Nicholson, Tucker, & Ba, 1998;Wessels et al., 2007;Bai & Dent, 2008;Symeonakis & Drake, 2004). The RUE is the ratio between annual NPP and annual rainfall. Since NPP is not easy to obtain, the NDVI is commonly used as a surrogate and has been proven comparable in several studies (e.g. Bai & Dent, 2008). It has been proposed that the RUE is able to normalize the inter-annual variability in the NPP which is caused by rainfall variability and consequently provides an index of degradation that is almost independent of the effects of precipitation (Nicholson, Tucker, & Ba, 1998).
For the north-eastern part of South Africa, Wessels et al. (2007) employed the rain-use efficiency to compare degraded and non-degraded areas. These areas have been selected on the basis of the National Land Cover (NLC) project (Fairbanks, Thompson, Vink, Newby, & Van Den Berg, HM, Everard, 2000). In order to isolate effects of degradation from local variations in soil, terrain etc., he selected comparable degraded and non-degraded test sites on the basis of Land Capability Units (for further information see Klingebiel & Montgomery, 1961). For the calculation of the RUE, Wessels worked with the ∑NDVI and rainfall data for the region's growth season. His results show that degraded areas have significantly lower RUE than the non-degraded areas. However, he also pointed out that the RUE strongly correlated with rainfall and cannot be expected to be independent of the effects of precipitation. Thus, RUE had a rather high variability over time and space.
Bai & Dent (2008) based their RUE analysis for South Africa on Wessels' paper and judged that long-term trends of RUE should distinguish between rainfall variability and land degradation. To get around the correlation of RUE and rainfall, they developed a so-called RUE-adjusted NDVI which basically masked out the areas where productivity was not determined by rainfall.
In this study, RUE was used as a first indicator for the state and development of the fragments and not for the final conclusion whether a fragment degraded over time or not. Thus and because the whole study area is rather small, further regional selections like Bai & Dent employed have not been used here.
As mentioned before, the EVI was used instead of the NDVI in this study. As Wessels et al. In a second approach, the difference between each year and its subsequent year was calculated and stacked to an 8-layer 'tendency stack'. This one was averaged for each pixel to get a single mean layer showing the general tendencies over time. At this point, it must be mentioned that I am aware of the fact that analysing a time period of nine years is too short to speak of the term 'trend'. Thus, I did not use the term 'trend' and the term 'tendency' was employed in this study. The tendencies were intended to give an indication of potential developments in the study area, but since the background information about the fragments are rather sparse, they only play a minor role in this study.
Furthermore, the values of the RUE mean tendency were extracted for the GTPs and normalised by the mean RUE of the time series. The results are displayed in a diagram with the mean percentage increase or decrease for the ground truth polygons.

EVI-rainfall Regression
The relationship between biomass and rainfall is an important indicator for the vitality of vegetation and has been analysed with remote sensing (e.g. Fabricante, Oesterheld, & Paruelo, 2009;Li, Lewis, Rowland, Tappan, & Tiszen, 2004). In this work, a linear regression was calculated in order to determine the response of the vegetation to precipitation.
The regression analysis is one of the most flexible and commonly used statistical analysis techniques. It is used to analyse the relationship between a dependent variable and one or more independent variables (Backhaus, Erichson, Plinke, & Weiber, 2008). The regression is especially employed to quantitatively describe and explain relationships and to estimate values of the dependent variable. In this work, the dependant variable will be represented by the EVI data and the independent variable by the rainfall data.
A simple regression for the estimation of the dependent variable is defined as follows (Backhaus, Erichson, Plinke, & Weiber, 2008):  Klein & Roehrig (2006) and due to calculation constraints. Klein & Roehrig (2006) Evans & Geerken (2004) developed a method to monitor degradation or improvement of vegetated areas which is independent of the effects of rainfall variability. This method, the so-called residual trends was adopted by Wessels et al. (2007) and Bai & Dent (2008) and shall be used in this work, too. In order to remove the effects of rainfall, the difference between the observed EVI and the linear regression predicted EVI is calculated.

Residual Trends (RESTREND)
These differences between observed and predicted EVI are referred to as residuals (Evans & Geerken, 2004). In order to get trends, the residuals are correlated over time. Wessels et al. (2007) used the paired areas again for the RESTREND method to show the different behaviour for degraded and non-degraded areas. Wessels and also Bai & Dent (2008) found the RESTRENDs to be independent of the effects of rainfall and showed that it can be a good indicator for the monitoring of 'degradation' or more commonly for monitoring changes decoupled from rainfall. According to their results, negative trends indicate 'degradation'; positive trends indicate an 'improvement' of the vegetated area.
In order to calculate the RESTREND in this work, the linear regression from the previous section was used. The IDL program used for the regression also contained a function to compute the residuals and to correlate them over time. The residual trends were only calculated for the whole time series and not for 2-year periods to show trends over the longest period possible. At this point it must be mentioned again that a 9-year period is often not enough to monitor long-term effects and the results have to be treated with caution. However, since remote sensing so far has no longer time series with such a good combination of temporal and spatial resolution, MODIS is often used for the monitoring of phenological trends (e.g. Fensholt, Rasmussen, Nielsen, & Mbow, 2009;He, J. Zhang, & Q. Zhang, 2009). The IDL program also delivers a layer of significance for the residuals of each tested lag and sum combination. Again, this method is pixel based in order to compare local variations which can be aggregated later on.

Distribution of Invasive Species
The second focus of this work is the distribution of invasive species into Fynbos, in particular the Acacia saligna and different Grass species. While these invasive species suppress the local Fynbos vegetation, they have the potential to reduce the biodiversity of the region and consequently degrade the ecosystem health (compare Van Der Wal et al., 2008). Therefore, it is important to establish an early warning system which is able to indicate areas that have been infested with alien vegetation to initiate appropriate measures. The results of this work shall be the basis for such a monitoring system and will give a first overview of the condition of the Fynbos fragments.
In order to delineate the distribution, a knowledge-based threshold method, the ENVI decision tree has been used. Basic statistical measures derived from the vegetation phenology as well as the methods from the previous sections establish the basis for this analysis.

Classification Scheme
The first step towards a suitable classification is a good aggregation of the sampling data.
Since the data collected during the field trips is too detailed for a suitable remote sensing classification, it has to be grouped to more appropriate classes. As mentioned above, this part of the study focuses on the delineation of invasive species in order to form the basis for an early warning system for the dispersal of these invasive species within the fragments. Thus, it is sufficient to define three different classes: 'Acacia', 'Fynbos' and 'Grass'. The ground truth polygons (GTP) were distributed manually to these classes according to their predominant vegetation cover (compare appendix A.6). In order to be classified as Acacia, a polygon must have at least 70 % acacia cover. The same applies for between. The GTP classified as Fynbos, had a Fynbos cover of at least 34 %, less than 5 % acacia cover and less or equal than 10 % grass cover on them. Polygons that did not match any of these criteria were marked as 'Mixed' and not taken into account for the classification.

Statistical Measures of the Phenology
Land cover classifications based on satellite imagery utilise the spectral or phenological characteristics to assign a certain class to a pixel. In this approach, these characteristics are derived from the phenological cycle of the vegetation. Obviously there seem to be some general differences between the three classes concerning the general amount of EVI and the reaction to rainfall. The polygon classified as Acacia seems to have a generally higher EVI with rather strong reactions to the two rainfall peaks while the Fynbos GTP only reacts to the first higher amount of rainfall and then stays on its rather low level of EVI.

Delineation using the ENVI Decision Tree
The Decision Tree is a tool of the image processing software ENVI by ITT. This tool performs a classification based on a series of binary decisions to place pixels into classes (ITT, 2008). Each decision divides the concerned pixels into two new classes using a

Results and Discussion
After the presentation of the adopted methods in the previous section, this chapter summarises the results. Furthermore, the results and the suitability of the indicators are displayed.

Productivity of Vegetation
The methodology for the productivity of vegetation was described in chapter 5.1. The following section presents the results and the corresponding discussion about the usefulness of the indicators.   In order to analyse these graphs, one has to assume a certain consistency in the vegetation cover. The RUE of Fynbos seems to be constantly lower than the one of Acacia, which confirms the interpretation of figure 6.2. Only for the first two years, the RUE values are quite similar for these two classes and even a bit higher for Fynbos in 2000.

Rain-use Efficiency
This Again, the grass covered Rondebosch Common has the lowest values of RUE, but as mentioned before has very high precipitation values. In contrast, Riverlands Plain, which is in the northern part of the study area with similar rainfall amounts as the Fynbos and Acacia fragments, has a high RUE value. This discrepancy can be compared to the one between the Rondebosch polygons and the northern grass covered polygons mentioned above. Thus, it seems that in years with sufficient rainfall, grass is as rain-use efficient as the acacia.    Like Wessels et al. (2007) suggested, the rain-use efficiency is still highly influenced by local and temporal variability of rainfall. Thus, the interpretation is difficult.
Despite the fact that local variations of the RUE occur in the study area, the analysis of the ground truth polygons showed differences between invasive vegetation and the The analysis of the vegetation fragments shows the same trends as the one of the ground truth polygons. Because of the high local precipitation variability in the study area due to the surrounding mountains, only fragments of the same region within the study area can be compared. For those, the same rule applies namely that acacia and grass covered fragments in 2008 with a normal amount of rainfall have higher RUE than the fragments with an intact Fynbos cover.
Thus, the analysis of the RUE implies, that an increased RUE for years with average rainfall indicates a degradation, i.e. an invasion of the Fynbos fragments by Acacia saligna.
The interpretation of the RUE tendencies is rather difficult. Since information about the general development of the fragments or management measures is insufficient, there are too many variables that influence the RUE tendencies, also including the rainfall variability.
The GTPs with intact Fynbos exhibited a quite homogenous positive tendency but it would be too vague to conclude that these polygons improved over the time series since the RUE of Fynbos strongly depends on the amount of rainfall. Furthermore, the GTP of Mamre with a currently high acacia cover had a high positive tendency too, which does not allow the establishing of a general rule.

EVI-rainfall Regression
As mentioned in section 6.1.2, the EVI-rainfall regression was conducted over the whole time from 2000 to 2008 and for 2-year periods. Figure     The second step was to apply the regression between EVI and rainfall to the whole time   Most likely, the correlation of Acacia saligna with rainfall (uninfluenced by e.g. neighbouring effects) lies somewhere between the one of Fynbos and Grass.
The very low homogenous correlation of the Mamre GTP and the corresponding fragment seems rather confusing at first glance since its vegetation cover has high local differences with dense areas of Acacia saligna and young but intact Fynbos areas. The explanation could be the heavy grazing that takes place on this fragment. The reportedly high number of livestock, especially pigs, sheep and goats on the Mamre fragment indicate an overgrazing of the fragment. During the rainy season, this would cause an uncoupling between EVI and rainfall in the measuring progress. Furthermore a fire in 2009 was recorded for the fragment, but no exact extent or location on the fragment (compare appendix A.4).
The interpretation of the whole time series is difficult once again since detailed background information is missing. The changes within the Fynbos polygons for the 2- year period compared to the whole time series seem major, but considering that all these values lie at 0.2 R² or underneath, there is just no real relationship between rainfall and productivity for both periods and hence the analysis of the differences is difficult but also unnecessary.
The analysis of the regression indicated that a higher coefficient of determination is closely connected with a change in vegetation cover from indigenous to invasive species, although certain issues that affect the signal of the polygons have to be considered. Thus, the regression can be a good indicator for the delineation of degraded, i.e. invaded areas.

RESTREND
As mentioned in section 5.1.3, the residual trends were only calculated on the basis of the regression for the whole time series and not for shorter time periods like the regression itself. Figure 6.10 shows the mean significance of the residual trends computed by the IDL program. The RESTREND is significant only for small areas, which could be expected after the results of the regression. Some parts of Kohler Bricks and Blaauwberg have significant RESTREND values, but altogether the residual trends are not consistent enough for an interpretation. In addition, some agricultural areas show high enough significance to work with but they are not of interest in this study. Although it proved to be a useful indicator for degradation of rangelands for Wessels et al. (2007), the results of the RESTREND display that this is not a suitable method for Fynbos vegetation. With very low correlations, the Fynbos vegetation cannot show any significant trends for the RESTREND method. It could be more useful to analyse the annual regression and the resulting residual trends instead of the seasonal regression but the time series of nine years is too short for this approach. This might be a possibility in some years, when the MODIS time series has reached a reasonable length, provided that the sensor is working long enough.
However, the different indicators of productivity presented in this chapter all point to the same conclusion: high biomass production of the tested fragments does not stand for Fynbos vegetation but rather for the infestation with invasive species such as Acacia saligna or different grass species. Thus, their distribution shall be delineated in the following chapter.

Distribution of Invasive Species
As mentioned during the methodology in chapter 5.2, the second aim of this work is to determine the distribution of invasive species, especially Acacia saligna and different Grass species. These species are one of the most important factors regarding ecosystem health in this region and have to be analysed.
In the following, the results of the different statistical measures of the vegetation phenology are presented and discussed and subsequently the delineation of the invasive species using the ENVI Decision Tree.

Statistical Measures of the Phenology
The statistical measures of the EVI have been calculated for every year and the values for     The last of the four diagrams in figure 6.14 presents the relation between maximum EVI In contrast to the grass vegetation, the Acacia saligna is an evergreen plant which results in high mean and median values. The standard deviations for the Acacia polygons are rather low and express the constant appearance over the course of the year. Similar to Fynbos, the Acacia is adapted very well to the low nutrient conditions of Fynbos soils, but has a higher biomass production and thus higher minima and maxima of EVI.
As mentioned above, certain outliers of the classes could be explained by small local fires but since the background information about these regions is rather poor, these fires cannot be confirmed and one can only conjecture about the reasons. Despite this, the statistical measures reflect typical characteristics of the different vegetation types and therefore a delineation of the invasive species should be possible on their basis.

Delineation using the ENVI Decision Tree
As mentioned in the previous chapter, several indicators could have been used for the delineation of invasive species. But since the Decision Tree is a simple threshold method only the ones that seemed to be the most useful were integrated in the tree.  The first decision (compare figure 6.16) separates most of the Grass pixels into the right branch. It is based on the maximum R² which has to be greater than 0.38 in order to fulfil the condition. Since the border between Grass and Acacia is rather narrow, a second node is inserted to eliminate the Acacia from the Grass pixels. The mean EVI is used in this case with a threshold of 0.2565; everything greater than this value is classified as Acacia.
The left branch also starts with a mean EVI node in order to separate the Acacia from the rest. In this case, everything less than 0.244 mean EVI is assumed not to be Acacia. The final decision is to delineate the remaining Grass pixels from the Fynbos. Every pixel with a standard deviation greater than 0.31 is assumed to be Grass. The remaining pixels should represent natural vegetation. At this point, it must be mentioned that Acacia classified pixels could still have a certain grass cover as well as Grass classified pixels could have a certain percentage of acacia but since the focus lies on the identification of disturbed, i.e. invaded areas in general, this is not of importance.    some Grass in between. From the aerial images and the field trips, it can be assured that the Acacia cover is quite accurately defined. Though the Fynbos vegetation in the western part of the fragment exits, it is rather mixed with grassy vegetation and some acacias. The pixels of the eastern part classified as Fynbos are also correct, but the vegetation cover is generally rather low, probably due to the heavy grazing or a fire in 2009.
The third fragment Blaauwberg (compare figure 6.18c) is also only formally protected but not fenced in and only few management measures have been taken so far. As the classification depicts, the land cover of this fragment is a mix of Acacia saligna and grassy areas with almost no visible Fynbos vegetation on it. However, the distribution of the two vegetation types is very patchy and the fragment basically contains mixed pixels, except for some larger acacia covered areas in the north. Thus, the delineation between the two classes seems inaccurate compared to the aerial images, but this is a matter of scale.
However, the important information is that the Blaauwberg fragment is completely invaded and has to be managed. All in all, the delineation of invasive species worked really well for the fragments. So far, a validation could not be done since additional GTPs are not available and also due to time constraints of this work. However, the aerial images and the overall impression of the fragments from the field trips suggest that the delineation of the alien vegetation from healthy Fynbos worked. Furthermore, a normal validation as it is common in remote sensing would not be meaningful for this small scale. As expected, the scale is a certain problem since a lot of pixels are classified as one of these three classes while they are actually rather mixed. Refining the classification and adding three transition classes could improve the delineation but again this could not be realised due to time constraints and might be part of future research.
This delineation of the invaded areas is a first step towards a monitoring of ecosystem health on the Sand Fynbos fragments. It clearly depicts that a long term management is closely related to the condition of the vegetation. There are no larger areas of Sand Fynbos vegetation in the study area that are not managed. Mamre seems to be an exception; while the grazing on the fragment not only seems to keep the Fynbos short it also hinders the dispersal of the acacia.

Conclusion and Perspectives
This study was intended to give a first overview of possibilities to monitor ecosystem health of remnant Sand Fynbos vegetation with the help of remote sensing. Ecosystem health is a very complex issue and indicators have to be found separately for each area of interest.
The area under intact Sand Fynbos vegetation has decreased so rapidly over the past decades that there is an urgent need today to protect and monitor the remaining fragments. Because these fragments are scattered over the whole municipality and their management is decentralised, ground monitoring is time and cost intensive. Remote sensing could be a cheap and comprehensive option for this purpose.
Since potential indicators for ecosystem health are numerous, not all of them could be covered or analysed in greater detail within the scope of this study. Thus, this topic is by far not exhausted and still has a lot of perspectives for future research.
This work focused on two of the three general measures of ecosystem health, established by Rapport, Costanza, & McMichael (1998): vitality and organisation. Vitality is basically measured in terms of primary productivity and organisation is represented by the biotic diversity of an ecosystem.
The first general measure, the vitality of vegetation was described by three indicators: the rain-use efficiency, an EVI-rainfall regression and the residual trends method.
The rain-use efficiency has been used in numerous studies as an index of degradation (e.g. Bai & Dent, 2008 The second indicator of vitality, the regression between EVI and rainfall was employed to investigate the reaction of the local vegetation to precipitation. Generally, it reflects the results of RUE. The photosynthetic activity of Sand Fynbos is not particularly dependant on rainfall and thus its general coefficient of determination is very low. Since the degraded vegetation cover, i.e. Acacia saligna and grasses, correlates rather well with rainfall, a distinction on the basis of this relationship is possible. Both the 2-years time period as well as the whole time series indicate this conclusion. Thus, the EVI-rainfall regression has great potential for the discrimination between intact Fynbos vegetation and invaded areas and is considered as a suitable indicator. The third indicator of vitality, the residual trends method (RESTREND) with its independence of rainfall variability displayed promising results in former studies (Wessels et al., 2007;Bai & Dent, 2008). For the special conditions of the study area, this method is not suited very well. Since the Sand Fynbos vegetation does not correlate with rainfall, any further developments based on this regression are not useful. Thus, the RESTREND method cannot deliver any further significant results for the application in the study area.
The analysis of the vitality displays the general problem of the term 'degradation'. Most publications so far refer to land as an economic resource and link degradation with the term 'land use'. Thus, degradation is commonly measured by decreasing productivity of vegetation. This approach is not appropriate for the analysis of complex ecosystems.
The special conditions of this study area with the unique indigenous Sand Fynbos vegetation and the invasive alien species showed that degradation requires a unique definition and its detection has to be determined individually. Since the degraded areas with its invasive alien species have a higher productivity than the Sand Fynbos vegetation, the interpretation of the indicators must be conducted differently from the literature. The second general measure by Rapport, Costanza, & McMichael (1998), the organisation of an ecosystem is represented by its biodiversity. Since this diversity dramatically decreases with the introduction of dominant alien species, like Acacia saligna in the study area, the distribution of these invasive species was delineated. On the basis of these combined indicators, a convenient delineation of the degraded areas could be generated. First, it is advisable to analyse the EVI-rainfall regression and second the annual phenology of the vegetation. Since the resolution of the MODIS data is rather coarse and mixed pixels were not defined in this delineation, the overall accuracy is probably not very high, but the classification gives a good first impression of potentially degraded areas.
These results can be used for specific ground monitoring and planning of management measures and the developed indicators form a suitable basis for future research. This work made an important first step to display the difficulties of the study area for remote sensing, so they can be taken into account in further studies.
In order to draw the right conclusions from the interpretation of the whole time series, consistent background information about the development of the fragments are necessary. These can only be achieved by an accurate documentation of management measures, fires and other factors that influence the vegetation cover.
The following section displays improvements of the used methods and recommended approaches for further research.
The applied delineation of invasive species could be enhanced with the introduction of mixed pixels to the classification. For this purpose, the influence of different compositions of Sand Fynbos vegetation, i.e. different proportions of restioid, ericoid and proteoid Fynbos has to be investigated. Furthermore, the decision progress for the delineation could be automated by the use of the R Regression Tree which is based on a programming language for statistical analysis (Bates et al., 2010). However, the general accuracy of the applied delineation has to be tested in other regions of the study area.
Now that the unique characteristics of the different vegetation types are determined, the use of Landsat images with a 30m spatial resolution could be considered. Two images, one from the growing season and one from the dry season could be compared in order to distinguish between the evergreen Sand Fynbos vegetation and annual grasses.
The developed indicators could be applied to other remote sensors with higher spatial resolutions, in order to achieve a higher geometric accuracy of the delineation of invasive species and to improve the extraction of 'pure' spectral information. RapidEye would be a suitable satellite system for further research since its temporal and spectral resolution are sufficiently high to monitor the vegetation phenology and its spatial resolution with up to 5m pixel size is significantly higher than the one of the MODIS sensor (RapidEye AG, 2010).
'Resilience', the third of the three general measures for ecosystem health by Rapport, Costanza, & McMichael (1998) could be determined on the basis of consistent background information. The recovery of vegetation after a fire or after the cutting of Acacia saligna could be monitored by the use of the EVI-rainfall regression or the RUE.
Another potential indicator for ecosystem health could be a fragmentation analysis (e.g. Fahrig, 2003). For this purpose, the size, surrounding land cover and connectivity between the fragments has to be investigated. Since the Sand Fynbos fragments are rather small and scattered over the whole city area, it would be important to see if the individual fragments are able to keep up their biotic diversity and if the seed exchange with other fragments is still possible.
Furthermore, the application of lidar (compare Rosso, Ustin, & Hastings, 2006) for the delineation of Acacia saligna seems promising for future research. Since the acacia is growing very fast and is generally taller than the Sand Fynbos vegetation, a distinction by the height of the plants should be possible and could deliver very accurate results.
However, this method is rather expensive because lidar systems are only airborne and flights for certain study areas have to be ordered specifically.
Although a higher spatial resolution would have been desirable, this study shows that remote sensing can be a suitable basis for a reasoned and sustainable management of Sand Fynbos fragments. This strict management becomes essential as Fynbos vegetation cannot regenerate on its own from the infestation of alien species anymore.

A.4 Email Interview with Ismail Ebrahim
The following email interview about the Mamre fragment with Ismail Ebrahim, employee of the CREW programme at SANBI, was conducted by Dr. Nicky Allsopp on the 10 th May,

2010.
Date of becoming part of city reserve system?

A.5 Email Interview with Joanne Eastman
The following summary of an email interview about the Rondebosch Common with Joanne Eastman, a member of the Friends of the Rondebosch Common, was conducted by Dr. Nicky Allsopp on the 27 th July, 2010.