- Email: [email protected]
The use of Landsat imagery for the identification of the remaining West Coast Renosterveld fragments, Western Cape Province, South Africa
Copyright © NISC Pty Ltd
South African Journal of Botany 2005, 71(1): 67–75 Printed in South Africa — All rights reserved
SOUTH AFRICAN JOURNAL OF BOTANY EISSN 1727–9321
The use of Landsat imagery for the identification of the remaining West Coast Renosterveld fragments, Western Cape Province, South Africa IP Newton* and RS Knight Department of Biodiversity and Conservation Biology, University of the Western Cape, Private Bag X17, Bellville 7530, South Africa * Corresponding author, e-mail: [email protected] Received 8 September 2003, accepted in revised form 14 July 2004
Landsat 7 ETM+ imagery from November 1999 was used to map the remaining fragments of West Coast Renosterveld, arguably the most transformed vegetation type in South Africa. A combination of supervised and unsupervised classifications was used. These showed that Renosterveld was not definitively discernible using spectral techniques. Our final map suggested that 7.9% remained, mainly untransformed, within a region defined by combining the boundaries suggested by previous researchers (Cowling and
Heijnis 2001, Low and Rebelo 1996). Using a predictive envelope based upon geological, pedological, altitudinal and rainfall data we estimated that 9.4% of the original extent of Renosterveld remained within the west coast lowlands. We examined boundary effects along the base of the Western Fold Mountains, and showed how soil mixing and coarse-scale maps had the potential to result in an 18.4% overestimate of the amount of spectrally-identified West Coast Renosterveld remaining.
Introduction West Coast Renosterveld is arguably the most transformed vegetation type in South Africa. The latest published estimate of the amount remaining was 5% (Von Hase et al. 2003). Previous estimates of the amount remaining relatively untransformed were 15.4% (Rouget et al. 2003), 9% (Reyers et al. 2001), 4% (Rebelo 1995) and 3% (McDowell and Moll 1992, Low and Rebelo 1996). There have been at least five published versions of the putative original extent of all or part of West Coast Renosterveld since 1953 (Acocks 1953, Boucher 1981, McDowell 1988, Low and Rebelo 1996, Cowling and Heijnis 2001). The beta version of the latest vegetation map of South Africa has recently been released (Mucina and Rutherford 2004), and the alpha version is expected within 2005. The three most recent published estimates of the amount of West Coast Renosterveld remaining have used Landsat Thematic Mapper imagery for this task. Reyers et al. (2001) used the CSIR-ARC (Council for Scientific and Industrial Research; Agricultural Research Council) land-cover classification of South Africa (Thompson 1996) and the original extents of Low and Rebelo (1996). This land-cover map was produced using manual aerial photographic techniques, and was based upon imagery exposed between 1994 and 1996. Rouget et al. (2003) used the classification produced by the ARC for the Institute for Plant
Conservation’s (IPC) ‘Threats to the biodiversity of the Cape floristic region’ project (Lloyd et al. 1999). They used the Broad Habitat Unit (BHU) system of Cowling and Heijnis (2001) as a basis for their original extent. The ARC-IPC classification was performed using 1998 imagery and spectral classification techniques, with manual intervention where required (Lloyd et al. 1999). The Cape Conservation Unit’s ‘Fine scale conservation plan for Cape Lowlands Renosterveld’ project used a combination of the natural vegetation fragments identified by the ARC classification (Lloyd et al. 1999) and our preliminary classification which was produced using a supervised spectral classification technique on November 1999 imagery. Von Hase et al. (2003) used a modified version of the BHU system of Cowling and Heijnis (2001) to define their core area. As part of a temporal and spatial study of West Coast Renosterveld, Landsat 7 ETM+ (Enhanced Thematic Mapper Plus) imagery was used in an attempt to identify, by spectral means, the remaining fragments of this vegetation type. These data were used in part by Von Hase et al. (2003), who performed more detailed field verification than was possible in our study. Thus we will use our data to examine some of the problems related to the spectral identification of West Coast Renosterveld. We also look at how the available ancillary data may affect estimates of the amount remaining.
68
Material and Methods Landsat 7 ETM+ imagery was obtained through the United States Geological Survey Eros Data Center via their website at http://edcwww.cr.usgs.gov/eros-home.html. The images were supplied as level one (systematic) corrected geotiff files. The scenes used were from Landsat path 175, rows 83 and 84, taken on 2/11/1999, which were virtually cloud free (<1%). The cloud that was present was restricted to some light transparent cloud situated to the west of, and over the southern half of, the Perdeberg. The two images were concatenated, based on geographical co-ordinates, and the joined image was windowed to exclude areas north of 32°24’18”S, east of 19°10’19”E, west of 17°58’01”E and south of 34°25’01”S. No radiometric corrections were carried out as the images were exposed within minutes of each other. Pixel resolution was 30m, except for band 6 (thermal band), which was 60m. This was expanded to a 30m resolution for the analyses. The 15m resolution panchromatic band was used for image rectification. Except for an unsupervised classification that was carried out using ER-Mapper (©Earth Resource Mapping Pty, Ltd, Perth, Australia), all other analyses were carried out using Idrisi32 software (©Clark Labs, Clark University, Worcester, MA, USA). Extensive preparatory tests were undertaken using pre- and post-processing filters, and several preliminary supervised and unsupervised classifications were performed. Although the classifications covered the region delimited above, the methods chosen were based upon the quality of the results falling within the combined extents of West Coast Renosterveld, as proposed by McDowell (1988), Low and Rebelo (1996) and Cowling and Heijnis (2001). A number of simple cluster type unsupervised classifications (Eastman 1999) suggested that 12 basic spectral classes existed within the imagery. A preliminary supervised classification was performed, using three Renosterveld training sites based upon the fragment map of McDowell and Moll (1992). A maximum likelihood method, based upon Bayesian probability theory, was used and prior probabilities were specified for each class (Eastman 1999). Prior probabilities were based upon the estimated proportion of the image expected to be covered by each of the 12 classes selected. All seven spectral bands were used in the classification. Each of the raw image bands was pre-processed with a Gaussian 7.x.7 filter before the classification. Three natural vegetation classes were recognised: a Renosterveld class, a Fynbos class, and a Sand Plain Fynbos–Dune Thicket class (nomenclature follows Low and Rebelo 1996). This last class was highly specific to the area to the south and east of the Langebaan lagoon. The Fynbos class was based upon mountain and lowland Fynbos training sites, but the results included alien Acacia vegetation, vineyards at a certain stage of development, and thicket. This preliminary classification was used as a basis for field verification, and was the one used by Von Hase et al. (2003). Approximately 80% of the sites spectrally identified as Renosterveld were visited. This was not a detailed survey
Newton and Knight
of sites as the aim was simply to ensure that the classification was identifying the natural vegetation fragments at an acceptable level of accuracy. On-site brief notes were made on the general structure and overall vegetative cover, where this could be reasonably determined, and whether it tended towards Renosterveld, Fynbos or Thicket. Using the information obtained during the field verification, the Renosterveld training sites were modified, and three different classes of Renosterveld were defined. These were ‘dense Renosterveld’ (>~75% shrub cover), ‘medium density Renosterveld’ (~25% to ~75% shrub cover) and ‘grassy Renosterveld’ (<~25% shrub cover). The preliminary supervised classification was repeated using this new data. In order to ‘spectrally verify’ these results, a number of unsupervised classifications based upon the ISODATA routine of Ball and Hall and other cluster routines such as the H-means and K-means procedures, were used (Eastman 1999). Each of these was run with different input parameters (number of clusters, number of iterations, seeding composite image) and all seven spectral bands were used. We used the 17 cluster; three iteration; band 2 (green); band 3 (red); band 4 (near infrared); false-colour composite seeding image, unsupervised classification to modify our revised supervised classification. An agricultural mask, derived from the agricultural classes identified in the unsupervised classification described in the previous sentence, was placed over the Fynbos class of the revised supervised classification, thus reducing the amount of agricultural land misclassified as Fynbos. The natural vegetation fragments of the two classifications were then combined. After the removal of the ‘grassy Renosterveld’ patches south of 33°48’00”S, all natural classes were combined into a single ‘natural vegetation’ class. The revised classification was converted to latitudelongitude format, using 15 ground control points (GCP). These were based upon the location of road intersections as determined by the 1:50.000 road vector layer, which was purchased from the Chief Directorate: Surveys and Mapping in Mowbray. These were compared with the equivalent intersections on the panchromatic band of the Landsat imagery (15m resolution). The root mean squared (RMS) error of this quadratic transformation was 15.56m. Fragments of natural vegetation of less than three hectares were then removed from the classification. Geological maps at a scale of 1:250.000, produced by the Council for Geoscience (Anonymous 1973, 1990, 1997), and the soil map (also at a scale of 1:250.000) published in Schloms et al. (1983) were used to develop a matrix of substrate types. These were used to assess the probability of the remaining natural vegetation fragments being Renosterveld, as edaphically defined by various authors (e.g. Kruger 1979, Boucher and Moll 1981). To do this the geological and soil maps were divided into groups that were either compatible with or not compatible with Renosterveld. This was done in terms of their description — shales, granites, eutrophic sands and loams were considered as ‘Renosterveld compatible’ (Boucher and Moll 1981). Leached sands, Table Mountain sandstones
South African Journal of Botany 2005, 71: 67–75
69
and wind-blown sands were not. The soil and geological maps were laid over each other to give three classes — soil and geological formations were both compatible with Renosterveld; one was and one was not compatible with Renosterveld, and finally, neither was compatible with the edaphic definition of Renosterveld. This was then laid over the fragment map and the areas of fragments occurring within each of these classes were measured. A digital elevation model (DEM) at 30m resolution was produced for the area from 20m interval contour line vector files, purchased from the Chief Directorate: Surveys and Mapping in Mowbray. The Triangulated Irregular Network module in Idrisi32 was used to develop the DEM (Eastman 1999). The CCWR (Computer Centre for Water Research) median rainfall image (from the compact disc) of Schulze (1997) was used to identify the 300.mm isohyet. Our final fragment map consists of all spectrallyidentified natural vegetation fragments of greater than 3ha, falling within the combined extent of the McDowell (1988), the Low and Rebelo (1996) and the Cowling and Heijnis (2001) proposed extents of West Coast Renosterveld. Fragments of natural vegetation on the west coast lowlands, outside of the above extent but falling within the Renosterveld-compatible soil/geological formations mentioned in the previous paragraph, occurring below 500m in altitude and receiving more than 300mm of rainfall per annum, were also identified.
180
Open Renosterveld Dense Renosterveld Grassy Renosterveld Fynbos Thicket/SP Fynbos
160 140 120 100 80 60
6
7 d Ba n
5
Ba nd
nd
4 Ba
nd Ba
Ba
nd
2 nd
1 Ba
nd
3
40 Ba
Figure 1 shows the places referred to in the text, the roads traversed during field verification and the location of those remnants of natural vegetation identified within the combined West Coast Renosterveld extents of Low and Rebelo (1996) and Cowling and Heijnis (2001). The two extents were combined as it was felt that both excluded important Renosterveld areas. For example, the Low and Rebelo (1996) extent excludes some of the Darling hills, while that of Cowling and Heijnis excludes Blouberg and Signal Hills. Figure 2 shows the spectral signatures of the revised supervised classification. The subdivision of the original ‘Renosterveld’ classification into three separate units was apparently valid, as the ‘grassy Renosterveld’ was distinctive in band 4, and the ‘dense Renosterveld’ in band 5. The wide spread of the Fynbos class, which tended to encompass most of the spread of the other natural vegetation classes, was worrying. However, our unsupervised classifications indicated that the spectral differences between Renosterveld and Fynbos were not distinct. A problem that was largely confined to the southeastern part of the region was the classification of ‘wasteland’ (mainly a mixture of alien grasses and light alien Acacia infestations) as Renosterveld. In the revised supervised classification these were mostly allocated to the ‘grassy Renosterveld’ class. Although the removal of this class from the southern parts of the final map helped resolve much of this problem, some highly transformed fragments
Figure 1: Spatially-degraded Landsat image showing the natural vegetation fragments (in black) of greater than three hectares remaining within the combined original extents (white lines) of West Coast Renosterveld, as proposed by Low and Rebelo (1996) and Cowling and Heijnis (2001). Black lines indicate roads taken during field verification. Figure is oriented north up. 1V-E-K = VoelvleiElandsberg PNR-Kranskop fragment
DIGITAL BRIGHTNESS
Results and Discussion
SPECTRAL BAND Figure 2: Spectral signatures (mean ± 1 standard deviation) of the natural vegetation classes used in the revised supervised classification. Thicket/SP Fynbos = Thicket and Sand Plain Fynbos
70
remained. A number of other problems related to correctly identifying natural land cover were encountered. These included a pine plantation on the Stellenbosch-Somerset West road that was classified as natural vegetation in all classifications. Part of the Elandsberg private nature reserve (PNR) had experienced a fire early in March 1999. The burnt area was partly classified as ‘cereals’, due to the sparse vegetation present at the time of exposure of the imagery. An area on the southern slopes of the Heuningberg was classified as ‘urban’. This was due to a fire caused by a passing train in late 1999. The pine plantation was digitised into the ‘plantation’ class, the burnt part of Elandsberg was digitised in as Renosterveld, and the Heuningberg section was left untouched as it was not certain whether the area would be left to grow naturally or would be converted to agriculture. Manual correction of fragments was minimal. None of Jarman’s (1981) floristically-defined communities (as determined in a field study by Boucher and Jarman 1977) in the Langebaan area was recognised by her Landsat multispectral scanner (MSS) classification. However, eight of her 14 spectral classes corresponded closely to communities defined by re-analysing the area in terms of the community structure description of Campbell et al. (1981). Renosterveld as an entity has not been recognised by authors developing structural vegetation classes for remote sensing purposes (Edwards 1983, Thompson 1996). It falls into the CSIR-ARC level 1 (resolvable at Landsat resolution) category of ‘shrubland and low Fynbos’ (Thompson 1996). In our study, it was generally found that where there was a predominance of broad-leaved shrubs (e.g. Protea, Rhus) the area was classified as Fynbos. Where fine-leaved shrubs and grass predominated the classification tended towards a Renosterveld classification. The actual boundary (broadleaved:fine-leaved species) between these two spectral classes was not established, but it appeared that the boundary between these two classes was, from a human perspective, quite subtle, and was probably also affected by slope and water availability as well. The unsupervised classifications generally identified two to four spectral clusters that predominantly represented the natural vegetation occurring within our imagery. The larger fragments were represented by more than one cluster, and within the extents of West Coast Renosterveld there were two or three major fragment groupings. In the first, the same cluster(s) represented the Elandsberg PNR, the Porseleinberg, the Heuningberg and the Koringberg. The Thicket–Sand Plain Fynbos region to the east of the Langebaan lagoon, the lower slopes of the Piketberg and the slopes along the Olifants River Valley were also usually included in this group. The second group included the Darling hills and most of the Western Fold mountainous regions. In the third group, the Paarlberg, the southern Cape Peninsula and the Hottentots-Holland shared the same clusters. The Perdeberg was always represented by two or three clusters, dividing it roughly into a northern and a southern half. As this spectral division of the Perdeberg into two parts was not evident in classifications using February
Newton and Knight
2001 imagery, it was assumed that this was caused by the very light cloud present in the November 1999 imagery. Of interest is that the supervised classification identified the northern half as Renosterveld and the southern half as Fynbos, whereas the unsupervised classifications allocated ‘lowland’ vegetation clusters to the southern half and ‘highland’ vegetation clusters to the north. The Kasteelberg always consisted of two main classes, one on the northeast slopes, the other on the southwest slopes. This was believed to be primarily caused by shadow on the steep southwestern side. Lloyd et al. (1999) noted a similar artifact on such mountain slopes. Based upon the combined Low and Rebelo (1996) and Cowling and Heijnis (2001) extents, the estimated natural vegetation cover of the unsupervised classifications ranged from 4.8% to 10.3% (average = 7.2%). The ‘lowland’ class varied between 0.9% and 7.9% (average = 4.5%), while the ‘highland’ class ranged from 1.5% to 4.1% (average = 2.9%). These ‘lowland’ and ‘highland’ classes were not directly related to the natural vegetation classes as floristically defined by ecologists, but referred to their predominance in representing lowland or mountain vegetation within the full image area. There were also clusters that were present in areas known to be natural vegetation, but as they were more common in areas of non-natural vegetation they were excluded from the area calculations. The unsupervised classification used to verify the revised supervised classification had a ‘lowland’ natural vegetation component of 3.2%, and a ‘highland’ one of 2.3%. Our preliminary and revised supervised classifications respectively gave 5.3% and 6.7% for the Renosterveld class(es), and 11.2% and 12.6% for the Fynbos class. After treatment with the agricultural mask from the unsupervised classification, the Fynbos class of the revised supervised classification was reduced to 1.9%. Table 1 shows the relationship between the revised supervised classification and the unsupervised classification used. Combining the two classifications benefitted the final map by helping to separate out the ‘wasteland’ patches and to better consolidate fragment borders. It has also increased the total area identified as natural vegetation by 43%. Ideally, one would like to compare the spectral classifications with vegetation structure to determine the relationship between these factors. To do this properly, it would be necessary to obtain vegetation structure from accurately-identified ground control points at a time very close to the exposure of the satellite imagery. Our field verification (carried out in 2001 and 2002) was very generalised and aimed at an overall assessment of a fragment or portion thereof. Figure 3 shows the locations of the 149 sites classified as Renosterveld in our preliminary classification where the field notes were detailed enough for a reasonable assessment. Table 2 shows a structural breakdown of these sites into 12 vegetation structures. The Renosterveld structural descriptions should not be confused with the Renosterveld classes developed for the revised supervised classification, although the training class sites are represented in the appropriate structural description as
South African Journal of Botany 2005, 71: 67–75
71
Table 1: Matrix showing the percentage overlap between the natural classes of the revised supervised and the unsupervised classifications that were combined to form the final fragment map Total area of all natural fragments of unsupervised classification = 41 172.8ha Total area of all natural fragments of supervised classification = 57 313.4ha Total area of combined classifications = 62 825.3.ha Unsupervised classes Lowland
Highland
No overlap
Total
Supervised classes1 Dense Renosterveld Open Renosterveld Grassy Renosterveld Fynbos3 No overlap
2.2 16.9 3.7 9.8 0.8
4.4 9.3 0.2 10.4 8.0
0.9 15.3 8.2 0.1 –
7.5 41.4 22.0 20.3 8.8
Total
33.3
32.2
34.5
43.22
1 The class names describe the vegetation structure existing at the training sites in November 2001. The three Renosterveld classes are divided at the 25% and 75% percentage canopy cover of the shrub layer level 2 The percentage of vegetation found only in one or other of the classifications 3 Fynbos class after modification with the agricultural mask developed from the unsupervised classification
the class type. Table 2 shows how the combining of the supervised and unsupervised classifications has improved the overall accuracy of the classification, despite fragments of less than 3ha being lost. It appears that the unsupervised classification identified the indigenous shrub-dominated vegetation better than the preliminary supervised classification, and incorporated fewer ‘wasteland’, tree and grassland fragments. Combining the two classifications has resulted in a slight improvement in the percentage of Renosterveld identified, and has reduced the proteoid and other unwanted components by a small amount. However, if the actual sample numbers are examined, it will be seen that the supervised and final map percentages have been calculated on a greater number of sample sites, implying that although there has been very little improvement in the percentage accuracy by combining the coverages, this accuracy is now based on a larger area. The two classes identified in the unsupervised classification were based upon clusters that were predominantly found where natural vegetation occurred. The presence of unwanted vegetation types within them argues that these unwanted types are spectrally indistinguishable from the indigenous shrubdominated areas. Experimental subdivision of the Renosterveld class of the preliminary classification had also shown that many of the ‘wasteland’ patches identified had identical spectral signatures to areas of natural vegetation within the Elandsberg PNR. It is therefore reasonable to assume that with the imagery available, a 70% success rate in identifying Renosterveld and an 80% success rate in identifying indigenous shrublands is about the limit attainable. Table 2 also looks at the figures (based upon our final map) in terms of fragment area. The discrepancy between the fragment numbers is due to some of the larger fragments having been sampled more than once. Where a single fragment has had two or more vegetation structures
Figure 3: Distribution of the sample sites (stars) used to estimate the accuracy with which fragments (grey) of West Coast Renosterveld were spectrally identified. Figure is oriented north up
3
64 15 7 1 4 15 5 8 1 13 16 149 91 14 105 16 13 15
Number of samples
43.0 10.1 4.7 0.7 2.7 10.1 3.4 5.4 0.7 8.7 10.7 100.0 61.1 9.4 70.5 10.7 8.7 10.1
% of total samples
12 2 0 0 0 3 2 2 0 4 3 28 14 4 18 3 4 3
42.9 7.1 0.0 0.0 0.0 10.7 7.1 7.1 0.0 14.3 10.7 100.0 50.0 14.3 64.3 10.7 14.3 10.7
33 8 6 1 4 7 1 5 1 2 4 72 52 7 59 4 2 7
45.8 11.1 8.3 1.4 5.6 9.7 1.4 6.9 1.4 2.8 5.6 100.0 72.2 9.7 81.9 5.6 2.8 9.7
Unsupervised classification Highland class Lowland class Number of % of total Number of % of total samples samples samples samples
56 15 7 1 3 11 4 6 1 6 8 118 82 11 93 8 6 11
Number of samples
47.5 12.7 5.9 0.8 2.5 9.3 3.4 5.1 0.8 5.1 6.8 100.0 69.5 9.3 78.8 6.8 5.1 9.3
% of total samples
Number of Total area fragments (ha) of sampled fragments sampled 50 14 852 13 4 925 7 2 898 1 1 815 3 852 9 4 661 3 2 532 3 6 339 1 17 6 1 160 8 1 892 104 41 944 74 25 343 7 8 889 81 34 231 8 1 892 6 1 160 9 4 661
Final fragment map % of total number of fragments sampled 48.1 12.5 6.7 1.0 2.9 8.7 2.9 2.9 1.0 5.8 7.7 100.0 71.2 6.7 77.9 7.7 5.8 8.7
% of total area of fragments sampled 35.4 11.7 6.9 4.3 2.0 11.1 6.0 15.1 0.0 2.8 4.5 100.0 60.4 21.2 81.6 4.5 2.8 11.1
1
A single stratum of shrubs, may have grass patches scattered within it. Where emergent shrubs occur, these are non-proteoid (e.g. Rhus, Chrysanthemoides, Olea, Euclea). Shrub density not recorded 2 As for ‘1’ above, but percentage canopy cover (PCC) of shrubs greater than 75% 3 As for ‘1’ above, PCC between 25% and 75% 4 Total vegetation cover less than 20% — only encountered in the Elandsberg PNR 5 A substantial component of succulents (Euphorbia, Cotyledon, Aizoaceae) 6 Two strata of shrubs, the emergents belonging to the Proteaceae 7 Low stratum of shrubs, but typically non-Renosterveld groups (e.g. Proteaceae, Ericaceae, Restionaceae). In some cases the low height may have been due to fires within the previous three to four years 8 Quite dense non-proteoid shrubs, such as found along runnels on hills 9 Large trees predominating. Either invasive species from nearby plantations (pines, eucalypts), dense growths of alien acacias or suburban areas with many trees, mainly noted in the Stellenbosch region 10 Typically grassland with less than 50% PCC of alien acacias or eucalypts
Renosterveld structure1 Dense Renosterveld2 Medium density Renosterveld Very sparse Renosterveld4 Succulent Renosterveld5 Grassy/herbaceous Proteoid structure6 Low Fynbos7 Thicket8 Trees predominating9 ‘Wasteland’10 Total Total Renosterveld Total other natural Total natural shrub Total ‘Wasteland’ Total trees Total grass/herbaceous
Vegetation structure
Preliminary classification
Table 2: Vegetation structure recorded at 149 sites classified as Renosterveld in the preliminary supervised classification, the number of these remaining in the final fragment map and the number present in each class of the unsupervised classification that was combined with the supervised classification to form the final map. For the final fragment map, it also shows the total number and area of the fragments sampled within each structure class
72 Newton and Knight
South African Journal of Botany 2005, 71: 67–75
73
Table 3: Area of natural vegetation fragments, as identified in this study, occurring within the extents of the three most recently-published proposed original extents of West Coast Renosterveld Author of extent used Cowling and Heijnis (2001) Boland . Swartland. Total.
Fragment area (ha)
% of original extent remaining
Proposed original extent (ha)
31 516 16 150 47 665
13.1 3.9 7.3
241 456 410 946 652 402
Low and Rebelo (1996)
45 733
7.5
613 589
Combined: Low and Rebelo (1996) and Cowling and Heijnis (2001)
53 416
7.9
673 714
Mucina and Rutherford (2004)1 Peninsula shale Renosterveld. Swartland alluvium Renosterveld. Swartland granite bulb veld. Swartland shale Renosterveld. Swartland silcrete Renosterveld. Total.
433 1 585 15 902 34 181 812 52 912
14.6 25.4 16.8 6.9 8.1 8.7
Swartland alluvium fynbos2.
10 596
23.3
45 447
Our predictive extent3
74 694
9.4
798 228
2 6 94 493 9 607
968 244 655 944 977 788
These values were calculated using the beta version. They may change with the release of the final version Data for this unit have been included as it incorporates most of the Voelvlei-Elandsberg-Kranskop fragment (see text for details). It would be better described as a Fynbos-Renosterveld mosaic 3 As derived from soil and geological maps, rainfall and altitude; see text for details 1 2
allocated to it, it has been assumed that the fragment had equal proportions of each of the structures noted. These affected, large heterogeneous fragments mainly occurred along the foot of the mountain slopes where Fynbos and Renosterveld adjoin. This is why, in terms of area, the percent component of the ‘natural shrub’ class is very similar in the fragment number and fragment area analyses, but the proportions of ‘Renosterveld’ and ‘other natural’ classes are different. It also demonstrates that the ‘wasteland’ and alien tree component is lower in terms of fragment area than in fragment number. With respect to the interval between the exposure of the imagery and field verification, a number of fires are known to have occurred during the intervening period (e.g. parts of the Kasteelberg and Piketberg). Of significance in our case would be the identification of grassy fragments during the field verification phase. Of the nine remaining in the final map, three are known to have burned since 1999 while three others are managed/used for grazing, and had a sparse covering of shrubs, suggesting that they should have been included in the Renosterveld class for accuracy assessment purposes. These latter three were the areas used as the ‘grassy Renosterveld’ training sites. The history of the remaining three is unknown. Assessing the status of ‘West Coast Renosterveld’ as an entity has become more complex since 2001, when Cowling and Heijnis divided the region into two units. The beta version of the NBI vegetation map of South Africa has further divided this vegetation, creating five units that could potentially be referred to as Coastal Lowland Renosterveld (Mucina and Rutherford 2004). In order that workers may
assess how the use of the different boundary parameters may affect estimates of the amount of natural vegetation remaining, and to suggest which of the subclasses is potentially the most threatened, we have calculated the area and percentage remaining for each of the three most recent proposals (Table 3). The Low and Rebelo (1996) and BHU system (Cowling and Heijnis 2001) assessed separately gave similar figures for the amount remaining; using the extents combined increased estimates by 5.1% and 7.6% respectively. Restricting our estimate to those fragments occurring within the five Renosterveld sites as defined by the beta version of the NBI’s vegetation map (Mucina and Rutherford 2004) gave a lower estimate of the amount remaining. A major reason for this is the allocation of much of the Voelvlei-Elandsberg-Kranskop (V-E-K) block and the area around Saron to Swartland Alluvium Fynbos. The V-E-K fragment is the largest block of natural vegetation in the area and had been included as West Coast Renosterveld in previous studies. It would be better described as a FynbosRenosterveld mosaic (see next paragraph). West Coast Renosterveld is probably unique as an endangered vegetation class. Its strong preference for finegrained, nutrient-rich soils (Kruger 1979) means that there are abrupt transitions from Renosterveld to Fynbos across edaphic boundaries (e.g. Linder 1976). Most of the remaining fragments are found around the perimeter of the region assigned to this class, or on the hills scattered across the landscape. The Western Fold Mountains, which form the eastern boundary of this class and the Piketberg are predominantly sandstone formations (Anonymous 1973, 1997). Along these boundaries we are dealing not only with
74
Newton and Knight
abrupt edaphic interfaces, but also with sandstone-derived debris eroding over the adjoining shale-derived plains. This will have led to soil mixtures and sandstone coverings of variable depth. The effect of this is that the (unavoidable) use of generalised boundaries may have led to a significant overestimation of the amount of West Coast Renosterveld remaining. We used the V-E-K block mentioned in the previous paragraph to estimate this error. From the spectral classification, the area of this block was measured at 7 476ha. Estimates from field surveys have suggested that 870ha of Voelvlei is floristically Renosterveld (Rebelo 1995). In Elandsberg estimates have varied from ‘less than 900ha’ (Chris Burgers quoted by Tansley 1982) to 1600ha by Krug (2004). For Kranskop, Benjamin Walton (pers. comm. to IPN) estimated that about half the area could be floristically described as Renosterveld. These figures therefore suggest that c. 3 730ha (or roughly half) of the V-E-K block are nonRenosterveld. We calculated the area of those fragments along the boundary of the combined extents of Low and Rebelo (1996) and Cowling and Heijnis (2001) that abutted the Western Fold and the Piketberg mountains. If we
assume that 50% of the area of these fragments is not supporting Renosterveld, our estimate of the amount of Renosterveld remaining would be reduced to 6.5%, a reduction of 18.4% on our original estimate. Of necessity all the broad-scale vegetation mapping exercises have been based upon generalised boundaries, yet have still provided a good indication of the relative status of each of the vegetation classes. Where detailed studies are being performed, it is necessary to look outside of these generalised boundaries for areas having the same physical parameters as the vegetation type being studied. Using our geological-pedological probability map, we allocated to every fragment of natural vegetation of greater than 3ha identified on the western lowlands occurring below 500m in altitude and receiving more than.300mm of rainfall a year a probability of supporting typical Renosterveld species. An upper rainfall limit was not imposed due to the very steep increase in rainfall along the Western Fold Mountains and the coarse (1’) grid used for rainfall interpolation by the CCWR (Dent et al. 1989). The use of the 500m altitude limit was considered a reasonable surrogate. Many of our outlying fragments fall within the Swartland shale Renosterveld areas of the beta version of the NBI vegetation map (Mucina and Rutherford 2004) that have expanded the boundaries of previously published West Coast Renosterveld extents. Our incorporation of the soil map of Schloms et al. (1983) has meant that we have identified fragments in some of the communities identified as Sand Fynbos by Mucina and Rutherford (2004). Based upon those areas identified by either or both the soil and geological maps we would estimate that 74 694ha (9.4% of a predicted original extent of 978 228ha) of West Coast Renosterveld remains. Restricting our estimates to those areas where both the geological and the soil maps predict compatible conditions, the equivalent figures were 54 624ha or 8.1% of a predicted original extent of 671 136ha of West Coast Renosterveld remaining. These figures are subject to the generalisations inherent in the soil, geological, rainfall and altitude data and the error level of our classification. Figure 4 shows our probability map for all the fragments identified as potentially supporting West Coast Renosterveld. Only detailed field studies will determine which of the fragments are genuine Renosterveld, and which are Fynbos or Thicket that have been included due to boundary generalisations and spectral misinterpretations. In addition, the level of disturbance within fragments can only be assessed in situ. McDowell and Moll (1992) reported that several farmers admitted they were planting dryland leguminous species within their Renosterveld fragments to improve their forage value. Donaldson et al. (1999) reported that the owner of the farm where they were working treated fragments less than 10ha in extent as part of the adjoining wheat fields, being burnt and sprayed as and when the surrounding wheat fields were. Although they were working in South Coast Renosterveld, it is not unreasonable to expect west coast farmers to do the same.
Figure 4: The location of fragments of natural vegetation on the west coast lowlands predicted to support Renosterveld, based upon geological, pedological, rainfall and altitudinal parameters (see text for details). Figure is oriented north up
Acknowledgements — This material is based upon work supported by the National Research Foundation (NRF) of South Africa, under Grant No. 2053674, awarded to Professor Suzanne J Milton. This work was carried out as part of the requirements of a PhD thesis at
South African Journal of Botany 2005, 71: 67–75
UWC by IPN, who acknowledges the receipt of an NRF bursary. Thanks go to Richard Cowling and an anonymous referee for their suggestions.
References Acocks JPH (1953) Veld types of South Africa. Memoirs of the Botanical Survey of South Africa 28: 1–192 Anonymous (1973) 3218 Clanwilliam: 1:250 000 Geological Series. Government Printer, Pretoria Anonymous (1990) 3318 Cape Town: 1:250 000 Geological Series. Government Printer, Pretoria Anonymous (1997) 3319 Worcester: 1:250 000 Geological Series. Council for Geoscience, Pretoria Boucher C (1981) Floristic and structural features of the coastal foreland vegetation south of the Berg River, Western Cape Province, South Africa. In: Moll E (ed) Proceedings of a Symposium on the Coastal Lowlands of the Western Cape, March 19–20, 1981. University of the Western Cape, Bellville, pp 21–26 Boucher C, Jarman ML (1977) The vegetation of the Langebaan area, South Africa. Transactions of the Royal Society of South Africa 42: 241–272 Boucher C, Moll EJ (1981) South African Mediterranean shrublands. In: Di Castri F, Goodall DW, Specht RL (eds) Ecosystems of the World 11: Mediterranean-type Shrublands. Elsevier, Amsterdam, pp 233–248 Campbell BM, Cowling RM, Bond WJ, Kruger FJ (1981) Structural characterization of vegetation in the Fynbos biome. South African National Scientific Programmes Report 52: 1–18 Cowling RM, Heijnis CE (2001) The identification of Broad Habitat Units as biodiversity entities for systematic conservation planning in the Cape floristic region. South African Journal of Botany 67: 15–38 Dent MC, Lynch SD, Schulze RE (1989) Mapping mean annual and other Rf statistics over southern Africa. WRC Report 109/1/89, Agriculture Catchments Research Unit Report No. 27 Donaldson J, Nanni I, Zachariades C, Kemper J (1999) Effects of habitat fragmentation on pollinator diversity and plant reproductive success in Renosterveld shrublands of South Africa. Conservation Biology 16: 1267–1276 Eastman JR (1999) Guide to GIS and Image Processing, Vol. 2. Clark University, Worcester, Massachusetts Edwards D (1983) A broad-scale structural classification of vegetation for practical purposes. Bothalia 14: 705–712 Jarman ML (1981) Remote Sensing Applications in Vegetation Mapping with Special Reference to the Langebaan Area, South Africa. MSc Thesis, Department of Botany, University of Cape Town, South Africa Krug CB (2004) The Renosterveld restoration project: understanding and restoring West Coast Renosterveld. Veld and Flora June 2004: 68–70 Kruger FJ (1979) South African heathlands. In: Specht RL (ed) Ecosystems of the World 9A. Heathlands and Related Shrublands: Descriptive Studies. Elsevier, Amsterdam, pp 19–80
Edited by RM Cowling
75
Linder HP (1976) A Preliminary Study of the Vegetation of the Piketberg Mountain, Cape Province. Honours project, Botany Department, University of Cape Town, South Africa Lloyd JW, Van den Berg EC, Van Wyk E (1999) The mapping of threats to biodiversity in the Cape floristic region with the aid of remote sensing and geographic information systems. Unpublished draft final report, December 1999. Report no. GW/A1999/54. Agricultural Research Council — Institute for Soil, Climate and Water, Pretoria, South Africa Low B, Rebelo AG (1996) Vegetation of South Africa, Lesotho & Swaziland. Department of Environmental Affairs and Tourism: Pretoria McDowell C (1988) Factors Affecting the Conservation of Renosterveld by Private Landowners. PhD Thesis, Department of Botany, University of Cape Town, South Africa McDowell C, Moll E (1992) The influence of agriculture on the decline of West Coast Renosterveld, south-western Cape, South Africa. Journal of Environmental Management 35: 173–192 Mucina L, Rutherford MC (2004) Vegetation map of South Africa, Lesotho and Swaziland: shapefiles of basic mapping units. Beta version 3.0, January 2004. National Botanical Institute, Cape Town Rebelo AG (1995) Renosterveld: conservation and research. In: Low AB, Jones FE (eds) The Sustainable Use and Management of Renosterveld Remnants in the Cape Floristic Region (Proceedings of a Symposium). FCC Report 1995/4. Flora Conservation Committee, Botanical Society of South Africa, Kirstenbosch, Cape Town, pp 32–42 Reyers B, Fairbanks DHK, Van Jaarsveld AS, Thompson MW (2001) South African vegetation priority conservation areas — a coarse filter approach. Diversity and Distributions 7: 79–89 Rouget M, Richardson DM, Cowling RM, Lloyd JW, Lombard AT (2003) Current patterns of habitat transformation and future threats to biodiversity in terrestrial ecosystems of the Cape floristic region, South Africa. Biological Conservation 112: 63–85 Schloms BHA, Ellis F, Lambrechts JJN (1983) Soils of the Cape coastal platform. South African National Scientific Programmes Report 75: 70–86 + 3 maps Schulze RE (1997) South African atlas of agrohydrology and climatology. Report TT82/96, Water Research Commission, Pretoria Tansley S (1982) Koppie Conservation Project. Unpublished project for the South African Nature Foundation Thompson M (1996) A standard land-cover classification scheme for remote-sensing applications in South Africa. South African Journal of Science 92: 34–42 Von Hase A, Rouget M, Maze K, Helme N (2003) A fine-scale conservation plan for Cape lowlands Renosterveld: technical report (main report). Report No. CCU 2/03, Cape Conservation Unit, Botanical Society of South Africa
South African Journal of Botany 2005, 71(1): 67–75 Printed in South Africa — All rights reserved
SOUTH AFRICAN JOURNAL OF BOTANY EISSN 1727–9321
The use of Landsat imagery for the identification of the remaining West Coast Renosterveld fragments, Western Cape Province, South Africa IP Newton* and RS Knight Department of Biodiversity and Conservation Biology, University of the Western Cape, Private Bag X17, Bellville 7530, South Africa * Corresponding author, e-mail: [email protected] Received 8 September 2003, accepted in revised form 14 July 2004
Landsat 7 ETM+ imagery from November 1999 was used to map the remaining fragments of West Coast Renosterveld, arguably the most transformed vegetation type in South Africa. A combination of supervised and unsupervised classifications was used. These showed that Renosterveld was not definitively discernible using spectral techniques. Our final map suggested that 7.9% remained, mainly untransformed, within a region defined by combining the boundaries suggested by previous researchers (Cowling and
Heijnis 2001, Low and Rebelo 1996). Using a predictive envelope based upon geological, pedological, altitudinal and rainfall data we estimated that 9.4% of the original extent of Renosterveld remained within the west coast lowlands. We examined boundary effects along the base of the Western Fold Mountains, and showed how soil mixing and coarse-scale maps had the potential to result in an 18.4% overestimate of the amount of spectrally-identified West Coast Renosterveld remaining.
Introduction West Coast Renosterveld is arguably the most transformed vegetation type in South Africa. The latest published estimate of the amount remaining was 5% (Von Hase et al. 2003). Previous estimates of the amount remaining relatively untransformed were 15.4% (Rouget et al. 2003), 9% (Reyers et al. 2001), 4% (Rebelo 1995) and 3% (McDowell and Moll 1992, Low and Rebelo 1996). There have been at least five published versions of the putative original extent of all or part of West Coast Renosterveld since 1953 (Acocks 1953, Boucher 1981, McDowell 1988, Low and Rebelo 1996, Cowling and Heijnis 2001). The beta version of the latest vegetation map of South Africa has recently been released (Mucina and Rutherford 2004), and the alpha version is expected within 2005. The three most recent published estimates of the amount of West Coast Renosterveld remaining have used Landsat Thematic Mapper imagery for this task. Reyers et al. (2001) used the CSIR-ARC (Council for Scientific and Industrial Research; Agricultural Research Council) land-cover classification of South Africa (Thompson 1996) and the original extents of Low and Rebelo (1996). This land-cover map was produced using manual aerial photographic techniques, and was based upon imagery exposed between 1994 and 1996. Rouget et al. (2003) used the classification produced by the ARC for the Institute for Plant
Conservation’s (IPC) ‘Threats to the biodiversity of the Cape floristic region’ project (Lloyd et al. 1999). They used the Broad Habitat Unit (BHU) system of Cowling and Heijnis (2001) as a basis for their original extent. The ARC-IPC classification was performed using 1998 imagery and spectral classification techniques, with manual intervention where required (Lloyd et al. 1999). The Cape Conservation Unit’s ‘Fine scale conservation plan for Cape Lowlands Renosterveld’ project used a combination of the natural vegetation fragments identified by the ARC classification (Lloyd et al. 1999) and our preliminary classification which was produced using a supervised spectral classification technique on November 1999 imagery. Von Hase et al. (2003) used a modified version of the BHU system of Cowling and Heijnis (2001) to define their core area. As part of a temporal and spatial study of West Coast Renosterveld, Landsat 7 ETM+ (Enhanced Thematic Mapper Plus) imagery was used in an attempt to identify, by spectral means, the remaining fragments of this vegetation type. These data were used in part by Von Hase et al. (2003), who performed more detailed field verification than was possible in our study. Thus we will use our data to examine some of the problems related to the spectral identification of West Coast Renosterveld. We also look at how the available ancillary data may affect estimates of the amount remaining.
68
Material and Methods Landsat 7 ETM+ imagery was obtained through the United States Geological Survey Eros Data Center via their website at http://edcwww.cr.usgs.gov/eros-home.html. The images were supplied as level one (systematic) corrected geotiff files. The scenes used were from Landsat path 175, rows 83 and 84, taken on 2/11/1999, which were virtually cloud free (<1%). The cloud that was present was restricted to some light transparent cloud situated to the west of, and over the southern half of, the Perdeberg. The two images were concatenated, based on geographical co-ordinates, and the joined image was windowed to exclude areas north of 32°24’18”S, east of 19°10’19”E, west of 17°58’01”E and south of 34°25’01”S. No radiometric corrections were carried out as the images were exposed within minutes of each other. Pixel resolution was 30m, except for band 6 (thermal band), which was 60m. This was expanded to a 30m resolution for the analyses. The 15m resolution panchromatic band was used for image rectification. Except for an unsupervised classification that was carried out using ER-Mapper (©Earth Resource Mapping Pty, Ltd, Perth, Australia), all other analyses were carried out using Idrisi32 software (©Clark Labs, Clark University, Worcester, MA, USA). Extensive preparatory tests were undertaken using pre- and post-processing filters, and several preliminary supervised and unsupervised classifications were performed. Although the classifications covered the region delimited above, the methods chosen were based upon the quality of the results falling within the combined extents of West Coast Renosterveld, as proposed by McDowell (1988), Low and Rebelo (1996) and Cowling and Heijnis (2001). A number of simple cluster type unsupervised classifications (Eastman 1999) suggested that 12 basic spectral classes existed within the imagery. A preliminary supervised classification was performed, using three Renosterveld training sites based upon the fragment map of McDowell and Moll (1992). A maximum likelihood method, based upon Bayesian probability theory, was used and prior probabilities were specified for each class (Eastman 1999). Prior probabilities were based upon the estimated proportion of the image expected to be covered by each of the 12 classes selected. All seven spectral bands were used in the classification. Each of the raw image bands was pre-processed with a Gaussian 7.x.7 filter before the classification. Three natural vegetation classes were recognised: a Renosterveld class, a Fynbos class, and a Sand Plain Fynbos–Dune Thicket class (nomenclature follows Low and Rebelo 1996). This last class was highly specific to the area to the south and east of the Langebaan lagoon. The Fynbos class was based upon mountain and lowland Fynbos training sites, but the results included alien Acacia vegetation, vineyards at a certain stage of development, and thicket. This preliminary classification was used as a basis for field verification, and was the one used by Von Hase et al. (2003). Approximately 80% of the sites spectrally identified as Renosterveld were visited. This was not a detailed survey
Newton and Knight
of sites as the aim was simply to ensure that the classification was identifying the natural vegetation fragments at an acceptable level of accuracy. On-site brief notes were made on the general structure and overall vegetative cover, where this could be reasonably determined, and whether it tended towards Renosterveld, Fynbos or Thicket. Using the information obtained during the field verification, the Renosterveld training sites were modified, and three different classes of Renosterveld were defined. These were ‘dense Renosterveld’ (>~75% shrub cover), ‘medium density Renosterveld’ (~25% to ~75% shrub cover) and ‘grassy Renosterveld’ (<~25% shrub cover). The preliminary supervised classification was repeated using this new data. In order to ‘spectrally verify’ these results, a number of unsupervised classifications based upon the ISODATA routine of Ball and Hall and other cluster routines such as the H-means and K-means procedures, were used (Eastman 1999). Each of these was run with different input parameters (number of clusters, number of iterations, seeding composite image) and all seven spectral bands were used. We used the 17 cluster; three iteration; band 2 (green); band 3 (red); band 4 (near infrared); false-colour composite seeding image, unsupervised classification to modify our revised supervised classification. An agricultural mask, derived from the agricultural classes identified in the unsupervised classification described in the previous sentence, was placed over the Fynbos class of the revised supervised classification, thus reducing the amount of agricultural land misclassified as Fynbos. The natural vegetation fragments of the two classifications were then combined. After the removal of the ‘grassy Renosterveld’ patches south of 33°48’00”S, all natural classes were combined into a single ‘natural vegetation’ class. The revised classification was converted to latitudelongitude format, using 15 ground control points (GCP). These were based upon the location of road intersections as determined by the 1:50.000 road vector layer, which was purchased from the Chief Directorate: Surveys and Mapping in Mowbray. These were compared with the equivalent intersections on the panchromatic band of the Landsat imagery (15m resolution). The root mean squared (RMS) error of this quadratic transformation was 15.56m. Fragments of natural vegetation of less than three hectares were then removed from the classification. Geological maps at a scale of 1:250.000, produced by the Council for Geoscience (Anonymous 1973, 1990, 1997), and the soil map (also at a scale of 1:250.000) published in Schloms et al. (1983) were used to develop a matrix of substrate types. These were used to assess the probability of the remaining natural vegetation fragments being Renosterveld, as edaphically defined by various authors (e.g. Kruger 1979, Boucher and Moll 1981). To do this the geological and soil maps were divided into groups that were either compatible with or not compatible with Renosterveld. This was done in terms of their description — shales, granites, eutrophic sands and loams were considered as ‘Renosterveld compatible’ (Boucher and Moll 1981). Leached sands, Table Mountain sandstones
South African Journal of Botany 2005, 71: 67–75
69
and wind-blown sands were not. The soil and geological maps were laid over each other to give three classes — soil and geological formations were both compatible with Renosterveld; one was and one was not compatible with Renosterveld, and finally, neither was compatible with the edaphic definition of Renosterveld. This was then laid over the fragment map and the areas of fragments occurring within each of these classes were measured. A digital elevation model (DEM) at 30m resolution was produced for the area from 20m interval contour line vector files, purchased from the Chief Directorate: Surveys and Mapping in Mowbray. The Triangulated Irregular Network module in Idrisi32 was used to develop the DEM (Eastman 1999). The CCWR (Computer Centre for Water Research) median rainfall image (from the compact disc) of Schulze (1997) was used to identify the 300.mm isohyet. Our final fragment map consists of all spectrallyidentified natural vegetation fragments of greater than 3ha, falling within the combined extent of the McDowell (1988), the Low and Rebelo (1996) and the Cowling and Heijnis (2001) proposed extents of West Coast Renosterveld. Fragments of natural vegetation on the west coast lowlands, outside of the above extent but falling within the Renosterveld-compatible soil/geological formations mentioned in the previous paragraph, occurring below 500m in altitude and receiving more than 300mm of rainfall per annum, were also identified.
180
Open Renosterveld Dense Renosterveld Grassy Renosterveld Fynbos Thicket/SP Fynbos
160 140 120 100 80 60
6
7 d Ba n
5
Ba nd
nd
4 Ba
nd Ba
Ba
nd
2 nd
1 Ba
nd
3
40 Ba
Figure 1 shows the places referred to in the text, the roads traversed during field verification and the location of those remnants of natural vegetation identified within the combined West Coast Renosterveld extents of Low and Rebelo (1996) and Cowling and Heijnis (2001). The two extents were combined as it was felt that both excluded important Renosterveld areas. For example, the Low and Rebelo (1996) extent excludes some of the Darling hills, while that of Cowling and Heijnis excludes Blouberg and Signal Hills. Figure 2 shows the spectral signatures of the revised supervised classification. The subdivision of the original ‘Renosterveld’ classification into three separate units was apparently valid, as the ‘grassy Renosterveld’ was distinctive in band 4, and the ‘dense Renosterveld’ in band 5. The wide spread of the Fynbos class, which tended to encompass most of the spread of the other natural vegetation classes, was worrying. However, our unsupervised classifications indicated that the spectral differences between Renosterveld and Fynbos were not distinct. A problem that was largely confined to the southeastern part of the region was the classification of ‘wasteland’ (mainly a mixture of alien grasses and light alien Acacia infestations) as Renosterveld. In the revised supervised classification these were mostly allocated to the ‘grassy Renosterveld’ class. Although the removal of this class from the southern parts of the final map helped resolve much of this problem, some highly transformed fragments
Figure 1: Spatially-degraded Landsat image showing the natural vegetation fragments (in black) of greater than three hectares remaining within the combined original extents (white lines) of West Coast Renosterveld, as proposed by Low and Rebelo (1996) and Cowling and Heijnis (2001). Black lines indicate roads taken during field verification. Figure is oriented north up. 1V-E-K = VoelvleiElandsberg PNR-Kranskop fragment
DIGITAL BRIGHTNESS
Results and Discussion
SPECTRAL BAND Figure 2: Spectral signatures (mean ± 1 standard deviation) of the natural vegetation classes used in the revised supervised classification. Thicket/SP Fynbos = Thicket and Sand Plain Fynbos
70
remained. A number of other problems related to correctly identifying natural land cover were encountered. These included a pine plantation on the Stellenbosch-Somerset West road that was classified as natural vegetation in all classifications. Part of the Elandsberg private nature reserve (PNR) had experienced a fire early in March 1999. The burnt area was partly classified as ‘cereals’, due to the sparse vegetation present at the time of exposure of the imagery. An area on the southern slopes of the Heuningberg was classified as ‘urban’. This was due to a fire caused by a passing train in late 1999. The pine plantation was digitised into the ‘plantation’ class, the burnt part of Elandsberg was digitised in as Renosterveld, and the Heuningberg section was left untouched as it was not certain whether the area would be left to grow naturally or would be converted to agriculture. Manual correction of fragments was minimal. None of Jarman’s (1981) floristically-defined communities (as determined in a field study by Boucher and Jarman 1977) in the Langebaan area was recognised by her Landsat multispectral scanner (MSS) classification. However, eight of her 14 spectral classes corresponded closely to communities defined by re-analysing the area in terms of the community structure description of Campbell et al. (1981). Renosterveld as an entity has not been recognised by authors developing structural vegetation classes for remote sensing purposes (Edwards 1983, Thompson 1996). It falls into the CSIR-ARC level 1 (resolvable at Landsat resolution) category of ‘shrubland and low Fynbos’ (Thompson 1996). In our study, it was generally found that where there was a predominance of broad-leaved shrubs (e.g. Protea, Rhus) the area was classified as Fynbos. Where fine-leaved shrubs and grass predominated the classification tended towards a Renosterveld classification. The actual boundary (broadleaved:fine-leaved species) between these two spectral classes was not established, but it appeared that the boundary between these two classes was, from a human perspective, quite subtle, and was probably also affected by slope and water availability as well. The unsupervised classifications generally identified two to four spectral clusters that predominantly represented the natural vegetation occurring within our imagery. The larger fragments were represented by more than one cluster, and within the extents of West Coast Renosterveld there were two or three major fragment groupings. In the first, the same cluster(s) represented the Elandsberg PNR, the Porseleinberg, the Heuningberg and the Koringberg. The Thicket–Sand Plain Fynbos region to the east of the Langebaan lagoon, the lower slopes of the Piketberg and the slopes along the Olifants River Valley were also usually included in this group. The second group included the Darling hills and most of the Western Fold mountainous regions. In the third group, the Paarlberg, the southern Cape Peninsula and the Hottentots-Holland shared the same clusters. The Perdeberg was always represented by two or three clusters, dividing it roughly into a northern and a southern half. As this spectral division of the Perdeberg into two parts was not evident in classifications using February
Newton and Knight
2001 imagery, it was assumed that this was caused by the very light cloud present in the November 1999 imagery. Of interest is that the supervised classification identified the northern half as Renosterveld and the southern half as Fynbos, whereas the unsupervised classifications allocated ‘lowland’ vegetation clusters to the southern half and ‘highland’ vegetation clusters to the north. The Kasteelberg always consisted of two main classes, one on the northeast slopes, the other on the southwest slopes. This was believed to be primarily caused by shadow on the steep southwestern side. Lloyd et al. (1999) noted a similar artifact on such mountain slopes. Based upon the combined Low and Rebelo (1996) and Cowling and Heijnis (2001) extents, the estimated natural vegetation cover of the unsupervised classifications ranged from 4.8% to 10.3% (average = 7.2%). The ‘lowland’ class varied between 0.9% and 7.9% (average = 4.5%), while the ‘highland’ class ranged from 1.5% to 4.1% (average = 2.9%). These ‘lowland’ and ‘highland’ classes were not directly related to the natural vegetation classes as floristically defined by ecologists, but referred to their predominance in representing lowland or mountain vegetation within the full image area. There were also clusters that were present in areas known to be natural vegetation, but as they were more common in areas of non-natural vegetation they were excluded from the area calculations. The unsupervised classification used to verify the revised supervised classification had a ‘lowland’ natural vegetation component of 3.2%, and a ‘highland’ one of 2.3%. Our preliminary and revised supervised classifications respectively gave 5.3% and 6.7% for the Renosterveld class(es), and 11.2% and 12.6% for the Fynbos class. After treatment with the agricultural mask from the unsupervised classification, the Fynbos class of the revised supervised classification was reduced to 1.9%. Table 1 shows the relationship between the revised supervised classification and the unsupervised classification used. Combining the two classifications benefitted the final map by helping to separate out the ‘wasteland’ patches and to better consolidate fragment borders. It has also increased the total area identified as natural vegetation by 43%. Ideally, one would like to compare the spectral classifications with vegetation structure to determine the relationship between these factors. To do this properly, it would be necessary to obtain vegetation structure from accurately-identified ground control points at a time very close to the exposure of the satellite imagery. Our field verification (carried out in 2001 and 2002) was very generalised and aimed at an overall assessment of a fragment or portion thereof. Figure 3 shows the locations of the 149 sites classified as Renosterveld in our preliminary classification where the field notes were detailed enough for a reasonable assessment. Table 2 shows a structural breakdown of these sites into 12 vegetation structures. The Renosterveld structural descriptions should not be confused with the Renosterveld classes developed for the revised supervised classification, although the training class sites are represented in the appropriate structural description as
South African Journal of Botany 2005, 71: 67–75
71
Table 1: Matrix showing the percentage overlap between the natural classes of the revised supervised and the unsupervised classifications that were combined to form the final fragment map Total area of all natural fragments of unsupervised classification = 41 172.8ha Total area of all natural fragments of supervised classification = 57 313.4ha Total area of combined classifications = 62 825.3.ha Unsupervised classes Lowland
Highland
No overlap
Total
Supervised classes1 Dense Renosterveld Open Renosterveld Grassy Renosterveld Fynbos3 No overlap
2.2 16.9 3.7 9.8 0.8
4.4 9.3 0.2 10.4 8.0
0.9 15.3 8.2 0.1 –
7.5 41.4 22.0 20.3 8.8
Total
33.3
32.2
34.5
43.22
1 The class names describe the vegetation structure existing at the training sites in November 2001. The three Renosterveld classes are divided at the 25% and 75% percentage canopy cover of the shrub layer level 2 The percentage of vegetation found only in one or other of the classifications 3 Fynbos class after modification with the agricultural mask developed from the unsupervised classification
the class type. Table 2 shows how the combining of the supervised and unsupervised classifications has improved the overall accuracy of the classification, despite fragments of less than 3ha being lost. It appears that the unsupervised classification identified the indigenous shrub-dominated vegetation better than the preliminary supervised classification, and incorporated fewer ‘wasteland’, tree and grassland fragments. Combining the two classifications has resulted in a slight improvement in the percentage of Renosterveld identified, and has reduced the proteoid and other unwanted components by a small amount. However, if the actual sample numbers are examined, it will be seen that the supervised and final map percentages have been calculated on a greater number of sample sites, implying that although there has been very little improvement in the percentage accuracy by combining the coverages, this accuracy is now based on a larger area. The two classes identified in the unsupervised classification were based upon clusters that were predominantly found where natural vegetation occurred. The presence of unwanted vegetation types within them argues that these unwanted types are spectrally indistinguishable from the indigenous shrubdominated areas. Experimental subdivision of the Renosterveld class of the preliminary classification had also shown that many of the ‘wasteland’ patches identified had identical spectral signatures to areas of natural vegetation within the Elandsberg PNR. It is therefore reasonable to assume that with the imagery available, a 70% success rate in identifying Renosterveld and an 80% success rate in identifying indigenous shrublands is about the limit attainable. Table 2 also looks at the figures (based upon our final map) in terms of fragment area. The discrepancy between the fragment numbers is due to some of the larger fragments having been sampled more than once. Where a single fragment has had two or more vegetation structures
Figure 3: Distribution of the sample sites (stars) used to estimate the accuracy with which fragments (grey) of West Coast Renosterveld were spectrally identified. Figure is oriented north up
3
64 15 7 1 4 15 5 8 1 13 16 149 91 14 105 16 13 15
Number of samples
43.0 10.1 4.7 0.7 2.7 10.1 3.4 5.4 0.7 8.7 10.7 100.0 61.1 9.4 70.5 10.7 8.7 10.1
% of total samples
12 2 0 0 0 3 2 2 0 4 3 28 14 4 18 3 4 3
42.9 7.1 0.0 0.0 0.0 10.7 7.1 7.1 0.0 14.3 10.7 100.0 50.0 14.3 64.3 10.7 14.3 10.7
33 8 6 1 4 7 1 5 1 2 4 72 52 7 59 4 2 7
45.8 11.1 8.3 1.4 5.6 9.7 1.4 6.9 1.4 2.8 5.6 100.0 72.2 9.7 81.9 5.6 2.8 9.7
Unsupervised classification Highland class Lowland class Number of % of total Number of % of total samples samples samples samples
56 15 7 1 3 11 4 6 1 6 8 118 82 11 93 8 6 11
Number of samples
47.5 12.7 5.9 0.8 2.5 9.3 3.4 5.1 0.8 5.1 6.8 100.0 69.5 9.3 78.8 6.8 5.1 9.3
% of total samples
Number of Total area fragments (ha) of sampled fragments sampled 50 14 852 13 4 925 7 2 898 1 1 815 3 852 9 4 661 3 2 532 3 6 339 1 17 6 1 160 8 1 892 104 41 944 74 25 343 7 8 889 81 34 231 8 1 892 6 1 160 9 4 661
Final fragment map % of total number of fragments sampled 48.1 12.5 6.7 1.0 2.9 8.7 2.9 2.9 1.0 5.8 7.7 100.0 71.2 6.7 77.9 7.7 5.8 8.7
% of total area of fragments sampled 35.4 11.7 6.9 4.3 2.0 11.1 6.0 15.1 0.0 2.8 4.5 100.0 60.4 21.2 81.6 4.5 2.8 11.1
1
A single stratum of shrubs, may have grass patches scattered within it. Where emergent shrubs occur, these are non-proteoid (e.g. Rhus, Chrysanthemoides, Olea, Euclea). Shrub density not recorded 2 As for ‘1’ above, but percentage canopy cover (PCC) of shrubs greater than 75% 3 As for ‘1’ above, PCC between 25% and 75% 4 Total vegetation cover less than 20% — only encountered in the Elandsberg PNR 5 A substantial component of succulents (Euphorbia, Cotyledon, Aizoaceae) 6 Two strata of shrubs, the emergents belonging to the Proteaceae 7 Low stratum of shrubs, but typically non-Renosterveld groups (e.g. Proteaceae, Ericaceae, Restionaceae). In some cases the low height may have been due to fires within the previous three to four years 8 Quite dense non-proteoid shrubs, such as found along runnels on hills 9 Large trees predominating. Either invasive species from nearby plantations (pines, eucalypts), dense growths of alien acacias or suburban areas with many trees, mainly noted in the Stellenbosch region 10 Typically grassland with less than 50% PCC of alien acacias or eucalypts
Renosterveld structure1 Dense Renosterveld2 Medium density Renosterveld Very sparse Renosterveld4 Succulent Renosterveld5 Grassy/herbaceous Proteoid structure6 Low Fynbos7 Thicket8 Trees predominating9 ‘Wasteland’10 Total Total Renosterveld Total other natural Total natural shrub Total ‘Wasteland’ Total trees Total grass/herbaceous
Vegetation structure
Preliminary classification
Table 2: Vegetation structure recorded at 149 sites classified as Renosterveld in the preliminary supervised classification, the number of these remaining in the final fragment map and the number present in each class of the unsupervised classification that was combined with the supervised classification to form the final map. For the final fragment map, it also shows the total number and area of the fragments sampled within each structure class
72 Newton and Knight
South African Journal of Botany 2005, 71: 67–75
73
Table 3: Area of natural vegetation fragments, as identified in this study, occurring within the extents of the three most recently-published proposed original extents of West Coast Renosterveld Author of extent used Cowling and Heijnis (2001) Boland . Swartland. Total.
Fragment area (ha)
% of original extent remaining
Proposed original extent (ha)
31 516 16 150 47 665
13.1 3.9 7.3
241 456 410 946 652 402
Low and Rebelo (1996)
45 733
7.5
613 589
Combined: Low and Rebelo (1996) and Cowling and Heijnis (2001)
53 416
7.9
673 714
Mucina and Rutherford (2004)1 Peninsula shale Renosterveld. Swartland alluvium Renosterveld. Swartland granite bulb veld. Swartland shale Renosterveld. Swartland silcrete Renosterveld. Total.
433 1 585 15 902 34 181 812 52 912
14.6 25.4 16.8 6.9 8.1 8.7
Swartland alluvium fynbos2.
10 596
23.3
45 447
Our predictive extent3
74 694
9.4
798 228
2 6 94 493 9 607
968 244 655 944 977 788
These values were calculated using the beta version. They may change with the release of the final version Data for this unit have been included as it incorporates most of the Voelvlei-Elandsberg-Kranskop fragment (see text for details). It would be better described as a Fynbos-Renosterveld mosaic 3 As derived from soil and geological maps, rainfall and altitude; see text for details 1 2
allocated to it, it has been assumed that the fragment had equal proportions of each of the structures noted. These affected, large heterogeneous fragments mainly occurred along the foot of the mountain slopes where Fynbos and Renosterveld adjoin. This is why, in terms of area, the percent component of the ‘natural shrub’ class is very similar in the fragment number and fragment area analyses, but the proportions of ‘Renosterveld’ and ‘other natural’ classes are different. It also demonstrates that the ‘wasteland’ and alien tree component is lower in terms of fragment area than in fragment number. With respect to the interval between the exposure of the imagery and field verification, a number of fires are known to have occurred during the intervening period (e.g. parts of the Kasteelberg and Piketberg). Of significance in our case would be the identification of grassy fragments during the field verification phase. Of the nine remaining in the final map, three are known to have burned since 1999 while three others are managed/used for grazing, and had a sparse covering of shrubs, suggesting that they should have been included in the Renosterveld class for accuracy assessment purposes. These latter three were the areas used as the ‘grassy Renosterveld’ training sites. The history of the remaining three is unknown. Assessing the status of ‘West Coast Renosterveld’ as an entity has become more complex since 2001, when Cowling and Heijnis divided the region into two units. The beta version of the NBI vegetation map of South Africa has further divided this vegetation, creating five units that could potentially be referred to as Coastal Lowland Renosterveld (Mucina and Rutherford 2004). In order that workers may
assess how the use of the different boundary parameters may affect estimates of the amount of natural vegetation remaining, and to suggest which of the subclasses is potentially the most threatened, we have calculated the area and percentage remaining for each of the three most recent proposals (Table 3). The Low and Rebelo (1996) and BHU system (Cowling and Heijnis 2001) assessed separately gave similar figures for the amount remaining; using the extents combined increased estimates by 5.1% and 7.6% respectively. Restricting our estimate to those fragments occurring within the five Renosterveld sites as defined by the beta version of the NBI’s vegetation map (Mucina and Rutherford 2004) gave a lower estimate of the amount remaining. A major reason for this is the allocation of much of the Voelvlei-Elandsberg-Kranskop (V-E-K) block and the area around Saron to Swartland Alluvium Fynbos. The V-E-K fragment is the largest block of natural vegetation in the area and had been included as West Coast Renosterveld in previous studies. It would be better described as a FynbosRenosterveld mosaic (see next paragraph). West Coast Renosterveld is probably unique as an endangered vegetation class. Its strong preference for finegrained, nutrient-rich soils (Kruger 1979) means that there are abrupt transitions from Renosterveld to Fynbos across edaphic boundaries (e.g. Linder 1976). Most of the remaining fragments are found around the perimeter of the region assigned to this class, or on the hills scattered across the landscape. The Western Fold Mountains, which form the eastern boundary of this class and the Piketberg are predominantly sandstone formations (Anonymous 1973, 1997). Along these boundaries we are dealing not only with
74
Newton and Knight
abrupt edaphic interfaces, but also with sandstone-derived debris eroding over the adjoining shale-derived plains. This will have led to soil mixtures and sandstone coverings of variable depth. The effect of this is that the (unavoidable) use of generalised boundaries may have led to a significant overestimation of the amount of West Coast Renosterveld remaining. We used the V-E-K block mentioned in the previous paragraph to estimate this error. From the spectral classification, the area of this block was measured at 7 476ha. Estimates from field surveys have suggested that 870ha of Voelvlei is floristically Renosterveld (Rebelo 1995). In Elandsberg estimates have varied from ‘less than 900ha’ (Chris Burgers quoted by Tansley 1982) to 1600ha by Krug (2004). For Kranskop, Benjamin Walton (pers. comm. to IPN) estimated that about half the area could be floristically described as Renosterveld. These figures therefore suggest that c. 3 730ha (or roughly half) of the V-E-K block are nonRenosterveld. We calculated the area of those fragments along the boundary of the combined extents of Low and Rebelo (1996) and Cowling and Heijnis (2001) that abutted the Western Fold and the Piketberg mountains. If we
assume that 50% of the area of these fragments is not supporting Renosterveld, our estimate of the amount of Renosterveld remaining would be reduced to 6.5%, a reduction of 18.4% on our original estimate. Of necessity all the broad-scale vegetation mapping exercises have been based upon generalised boundaries, yet have still provided a good indication of the relative status of each of the vegetation classes. Where detailed studies are being performed, it is necessary to look outside of these generalised boundaries for areas having the same physical parameters as the vegetation type being studied. Using our geological-pedological probability map, we allocated to every fragment of natural vegetation of greater than 3ha identified on the western lowlands occurring below 500m in altitude and receiving more than.300mm of rainfall a year a probability of supporting typical Renosterveld species. An upper rainfall limit was not imposed due to the very steep increase in rainfall along the Western Fold Mountains and the coarse (1’) grid used for rainfall interpolation by the CCWR (Dent et al. 1989). The use of the 500m altitude limit was considered a reasonable surrogate. Many of our outlying fragments fall within the Swartland shale Renosterveld areas of the beta version of the NBI vegetation map (Mucina and Rutherford 2004) that have expanded the boundaries of previously published West Coast Renosterveld extents. Our incorporation of the soil map of Schloms et al. (1983) has meant that we have identified fragments in some of the communities identified as Sand Fynbos by Mucina and Rutherford (2004). Based upon those areas identified by either or both the soil and geological maps we would estimate that 74 694ha (9.4% of a predicted original extent of 978 228ha) of West Coast Renosterveld remains. Restricting our estimates to those areas where both the geological and the soil maps predict compatible conditions, the equivalent figures were 54 624ha or 8.1% of a predicted original extent of 671 136ha of West Coast Renosterveld remaining. These figures are subject to the generalisations inherent in the soil, geological, rainfall and altitude data and the error level of our classification. Figure 4 shows our probability map for all the fragments identified as potentially supporting West Coast Renosterveld. Only detailed field studies will determine which of the fragments are genuine Renosterveld, and which are Fynbos or Thicket that have been included due to boundary generalisations and spectral misinterpretations. In addition, the level of disturbance within fragments can only be assessed in situ. McDowell and Moll (1992) reported that several farmers admitted they were planting dryland leguminous species within their Renosterveld fragments to improve their forage value. Donaldson et al. (1999) reported that the owner of the farm where they were working treated fragments less than 10ha in extent as part of the adjoining wheat fields, being burnt and sprayed as and when the surrounding wheat fields were. Although they were working in South Coast Renosterveld, it is not unreasonable to expect west coast farmers to do the same.
Figure 4: The location of fragments of natural vegetation on the west coast lowlands predicted to support Renosterveld, based upon geological, pedological, rainfall and altitudinal parameters (see text for details). Figure is oriented north up
Acknowledgements — This material is based upon work supported by the National Research Foundation (NRF) of South Africa, under Grant No. 2053674, awarded to Professor Suzanne J Milton. This work was carried out as part of the requirements of a PhD thesis at
South African Journal of Botany 2005, 71: 67–75
UWC by IPN, who acknowledges the receipt of an NRF bursary. Thanks go to Richard Cowling and an anonymous referee for their suggestions.
References Acocks JPH (1953) Veld types of South Africa. Memoirs of the Botanical Survey of South Africa 28: 1–192 Anonymous (1973) 3218 Clanwilliam: 1:250 000 Geological Series. Government Printer, Pretoria Anonymous (1990) 3318 Cape Town: 1:250 000 Geological Series. Government Printer, Pretoria Anonymous (1997) 3319 Worcester: 1:250 000 Geological Series. Council for Geoscience, Pretoria Boucher C (1981) Floristic and structural features of the coastal foreland vegetation south of the Berg River, Western Cape Province, South Africa. In: Moll E (ed) Proceedings of a Symposium on the Coastal Lowlands of the Western Cape, March 19–20, 1981. University of the Western Cape, Bellville, pp 21–26 Boucher C, Jarman ML (1977) The vegetation of the Langebaan area, South Africa. Transactions of the Royal Society of South Africa 42: 241–272 Boucher C, Moll EJ (1981) South African Mediterranean shrublands. In: Di Castri F, Goodall DW, Specht RL (eds) Ecosystems of the World 11: Mediterranean-type Shrublands. Elsevier, Amsterdam, pp 233–248 Campbell BM, Cowling RM, Bond WJ, Kruger FJ (1981) Structural characterization of vegetation in the Fynbos biome. South African National Scientific Programmes Report 52: 1–18 Cowling RM, Heijnis CE (2001) The identification of Broad Habitat Units as biodiversity entities for systematic conservation planning in the Cape floristic region. South African Journal of Botany 67: 15–38 Dent MC, Lynch SD, Schulze RE (1989) Mapping mean annual and other Rf statistics over southern Africa. WRC Report 109/1/89, Agriculture Catchments Research Unit Report No. 27 Donaldson J, Nanni I, Zachariades C, Kemper J (1999) Effects of habitat fragmentation on pollinator diversity and plant reproductive success in Renosterveld shrublands of South Africa. Conservation Biology 16: 1267–1276 Eastman JR (1999) Guide to GIS and Image Processing, Vol. 2. Clark University, Worcester, Massachusetts Edwards D (1983) A broad-scale structural classification of vegetation for practical purposes. Bothalia 14: 705–712 Jarman ML (1981) Remote Sensing Applications in Vegetation Mapping with Special Reference to the Langebaan Area, South Africa. MSc Thesis, Department of Botany, University of Cape Town, South Africa Krug CB (2004) The Renosterveld restoration project: understanding and restoring West Coast Renosterveld. Veld and Flora June 2004: 68–70 Kruger FJ (1979) South African heathlands. In: Specht RL (ed) Ecosystems of the World 9A. Heathlands and Related Shrublands: Descriptive Studies. Elsevier, Amsterdam, pp 19–80
Edited by RM Cowling
75
Linder HP (1976) A Preliminary Study of the Vegetation of the Piketberg Mountain, Cape Province. Honours project, Botany Department, University of Cape Town, South Africa Lloyd JW, Van den Berg EC, Van Wyk E (1999) The mapping of threats to biodiversity in the Cape floristic region with the aid of remote sensing and geographic information systems. Unpublished draft final report, December 1999. Report no. GW/A1999/54. Agricultural Research Council — Institute for Soil, Climate and Water, Pretoria, South Africa Low B, Rebelo AG (1996) Vegetation of South Africa, Lesotho & Swaziland. Department of Environmental Affairs and Tourism: Pretoria McDowell C (1988) Factors Affecting the Conservation of Renosterveld by Private Landowners. PhD Thesis, Department of Botany, University of Cape Town, South Africa McDowell C, Moll E (1992) The influence of agriculture on the decline of West Coast Renosterveld, south-western Cape, South Africa. Journal of Environmental Management 35: 173–192 Mucina L, Rutherford MC (2004) Vegetation map of South Africa, Lesotho and Swaziland: shapefiles of basic mapping units. Beta version 3.0, January 2004. National Botanical Institute, Cape Town Rebelo AG (1995) Renosterveld: conservation and research. In: Low AB, Jones FE (eds) The Sustainable Use and Management of Renosterveld Remnants in the Cape Floristic Region (Proceedings of a Symposium). FCC Report 1995/4. Flora Conservation Committee, Botanical Society of South Africa, Kirstenbosch, Cape Town, pp 32–42 Reyers B, Fairbanks DHK, Van Jaarsveld AS, Thompson MW (2001) South African vegetation priority conservation areas — a coarse filter approach. Diversity and Distributions 7: 79–89 Rouget M, Richardson DM, Cowling RM, Lloyd JW, Lombard AT (2003) Current patterns of habitat transformation and future threats to biodiversity in terrestrial ecosystems of the Cape floristic region, South Africa. Biological Conservation 112: 63–85 Schloms BHA, Ellis F, Lambrechts JJN (1983) Soils of the Cape coastal platform. South African National Scientific Programmes Report 75: 70–86 + 3 maps Schulze RE (1997) South African atlas of agrohydrology and climatology. Report TT82/96, Water Research Commission, Pretoria Tansley S (1982) Koppie Conservation Project. Unpublished project for the South African Nature Foundation Thompson M (1996) A standard land-cover classification scheme for remote-sensing applications in South Africa. South African Journal of Science 92: 34–42 Von Hase A, Rouget M, Maze K, Helme N (2003) A fine-scale conservation plan for Cape lowlands Renosterveld: technical report (main report). Report No. CCU 2/03, Cape Conservation Unit, Botanical Society of South Africa