Assessment of the Contribution of WorldView-2 Strategically Positioned Bands in Bracken fern ( Pteridium aquilinum ( L . ) Kuhn ) Mapping

In the eThekwini Metropolitan Area, mitigation of the Bracken fern (Pteridium aquilinum (L.) Kuhn) invasion within the KwaZulu-Natal Sandstone Sourveld (KZNSS) has been identified as a major environmental priority. To facilitate informed interventions, reliable Bracken fern spatial distribution is necessary. Earlier efforts to map the fern using lower spatial and spectral resolution imagery have been unsuccessful. Consequently, this study sought to determine the reliability of the “new generation” World View-2 (WV-2) image characterised by higher spatial and spectral resolution in delineating the fern invaded areas. The eight band WV2 image was atmospherically corrected and spectrally resized as the SPOT-5 wavebands, additional bands and all bands. The classification accuracy was compared to results from the SPOT-5 image. Results showed that classification based on WV-2s additional bands had superior classification accuracy than the rest of the categories. Furthermore, classification based on all the WV-2s bands and the traditional bands perfomed better than the SPOT-5 image in delineating areas covered by the fern. These findings indicate the value of of the “new generation” imagery characterised by higher spatial and spectral resolution in improving the accuracy of the fern invaded landscapes. South African Journal of Geomatics, Vol. 3, No. 2, August 2014


Introduction
The KwaZulu-Natal Sandstone Sourveld (KZNSS) is one of South Africa's most important grassland ecosystems that provide both social and ecological services to the eThekwini Metropolitan Municipality in KwaZulu-Natal Province, South Africa.In its pristine condition, the grassland is known to be rich in species diversity and provides among other services soil formation, control of erosion, carbon sequestration and recreational opportunities (EThekwiniMunicipality 2012(EThekwiniMunicipality /2013;;Msibi 2011).However, in the recent past, the Sourveld has been converted to, among others cultivated land, commercial plantations, alternative vegetation types and urban development.According to Mucina and Rutherford (2006), only 0.2% of the sourveld is under statutory protection, consequently, long-term sustainability of the vegetation is in doubt.
Encroachment of shrub into the KZNSS landscape has been noted in many parts of the Province (Hudak and Wessman 2001;O'Connor 2005;Wigley et al. 2009).The Bracken fern (Pteridium aquilinum (L.) Kuhn), an aggressive invasive species, has in the recent past been identified as the biggest threat to the KZNSS's remnant patches (Roos et al. 2010;Schneider and Fernando 2010;Msibi 2011).The fern is known to supress resident species, paving way for the emergence of woody plants and forest pioneers (Msibi 2011).Its extensive rooting system enables it to outcompete other species for moisture and nutrients and its typical dense senesced cover impedes germination and growth of other plants.The emerging threats to the KZNSS and its value within eThekwini Municipality necessitates an inventory of the distribution of the fern to facilitate informed intervention.
Traditionally, techniques based on field surveys, aerial photography and review of historical literature among others have been used to determine spatial ecological extents, however, these techniques are often costly, tedious and time consuming (Xie et al. 2008).In the recent past, due to its wide spatio-temporal resolution, availability and lower cost per unit area, remotely sensed data has emerged as a viable tool for land cover mapping (Foody, 2002;Lu et al., 2004;Liu et al. 2004;Kavzoglu and Colkesen 2009;Abd El-Kawy et al. 2011).
Despite the potential of remote sensing in the fern mapping, there is paucity in literature on its successful application (Tong et al. 2006).Fuller et al. (1994) for instance mapped a Bracken infested landscape using Landsat TM.Whereas a satisfactory overall classification accuracy between 80-85% was achieved in other land cover types, the fern classification accuracy was less than 8%.A number of authors (Birnie and Miller 1986;Miller et al. 1989;Miller et al. 1990;Pakeman et al. 1996 among others) attribute such low classification accuracy to the fragmented patches which are often below the pixel extents of commonly used multispectral images like Landsat and SPOT.
The emergence of "new generation" multispectral sensors such as WorldView-2 offer a valuable trade-off between multi/hyper spectral and low/high spatial resolution imagery (Mutanga et al. 2012).Their higher spatial resolution than traditional low and medium resolution imagery and a higher number of bands at strategic sections of the electromagnetic spectrum offer great potential in land cover mapping (Cho et al. 2012).The WorldView-2 sensor for instance is characterised by four additional spectral bands to those contained in SPOT.The strategic location of these bands within the coastal blue, yellow, red-edge, and NIR2 of the electromagnetic spectrum is valuable for vegetation mapping (Mutanga et al. 2012;Cho et al. 2012).
Based on the aforementioned challenges in mapping the fern using lower spatial and spectral resolution imagery, this study explores the potential of new generation WV-2s additional strategically positioned bands in mapping the fern.

Materials and Methods
This study was conducted in Giba Gorge, within eThekwini Metropolitan Municipality, KwaZulu-Natal, South Africa (Figure 1).Due to pressure from other land uses on the KZNSS, private landowners and the Municipality started the Giba Gorge Environmental Precinct cooperative project to manage a common conservation area.However, currently, the biggest threat to the Giba Gorge Environmental Precinct is the displacement of natural habitat by other vegetation forms such as eucalyptus and unplanned fires.The latter has particularly been associated with the fern's invasion (Lindenmayer et al. 2010).undertaken in IDRISI Andes using the Chavez's COST model (Chavez, 1996).In both the image datasets, a total of 623 reference points was generated, 70% (436) for training and 30% (187) for validation.

Results
The classification based on the four strategically positioned bands (coastal blue, yellow, rededge and NIR2) and the SPOT 5 image produced the highest and lowest overall accuracy respectively (Table 2 and 3).The overall classification accuracy based on the WV-2 eight bands was lower than the WV-2s strategically positioned bands while the overall accuracy based on WV-2s traditional bands (blue, green, red and NIR) was higher than the accuracy achieved using SPOT 5 image (Table 2 and 3).A summary of the fern accuracies for the four image categories are reported in Table 2 while the confusion matrices for the major land cover-types in the study area reported in Table 3 a and b.  2) in concert with field data showed better classification accuracy using additional bands image, with smallest fern patches correctly identified in grass dominated areas (Figure 3C).Classifications based on WV-2s eight bands and traditional bands, Figure 2a

Discussion
This study explored the potential of new generation WV-2 sensor in mapping the fern.The Results in this study indicate suitability of the WV-2s improved spectral resolution in vegetation mapping in corroboration with studies reported by eg Mehner et al. (2004) and Holland and Aplin (2013).In comparison to other band combinations (see table 2), the strategically positioned bands have been known to be more sensitive to different levels of chlorophyll, foliage mass and leaf area index, and therefore suitable for discriminating different vegetation types (Daughtry and Walthall 1998;Cochrane 2000;Schmidt and Skidmore 2003).Consequently, these bands are valuable in discriminating the fern from other vegetation types.
The perfomance of all the WV-2s band combinations (see table 2) were superior to SPOT 5 image in discriminating the fern.This can be attributed to the lower spatial and spectral resolution that characterise the SPOT 5 image.Commonly, reliable classification based on imagery characetrised by lower spatial and spectral resolutions is impeded by the mixed pixel problem.As seen in this study (Figure 2 and 4), SPOT 5 imagery may be unsuitable for discriminating the fern from heterogeneous landscapes due to spectral confusion and mixed pixel problem.
Results in this study show the suitability of the additional and strategically positioned bands in WV2 imagery for mapping the fern.Areas around Giba Gorge are generally characterised by mild subtropical conditions during the year.In this regard, most of the months are ideal for acquiring imagery for mapping the fern.Since no site site specific fern's spectral uniqueness has ever been reported, use of similar data sets in diverse landscapes can be expected to produce similar results.However, typically, the fern is vulnerable to extreme winter conditions, causing a dieback.Consequently, in areas characterised by late winter frost conditions, image acquisition for mapping should be captured during the winter onset as the fern's folier response to extreme winter conditions may impede reliabled delineation.

Conclusions
This study explored the potential of WV-2 image data in mapping the fern.The position and the number of spectral bands were assessed and compared to the commonly used multispectral image (SPOT-5).The results showed that the additional bands in WV-2 are valuable in discriminating the fern from other vegetation types.The added spectral dimensions
and b respectively and the SPOT 5 image (Figure 2 d) were less effective in delineating areas covered by the fern.
position and the number of bands were assessed and compared to the four bands that characterise the SPOT 5 image.Results revealed that WV-2 additional bands (coastal blue, yellow, red-edge and NIR2) can improve the mapping accuracy.A number of studies(Chen 2010;Ozdemira and Karnielib 2011; Cho et al. 2012)  note the potential of WV-2 additional bands in vegetation mapping.Using pixel-based approach, Chen (2010) demonstrated that the four additional bands were most suitable for differentiating tree species while Cho et al.(2012) identified WV-2s yellow band as the most influential in vegetation mapping.Ozdemira and Karnielib (2011) used WV-2s image texture to predict forest structure parameters and identified the WV-2s additional bands (yellow, red-edge and NIR2)as most suitable for predicting forest structure.Other studies(Dlamini 2010;Omar 2010) note the value of WV-and red-edge portions in distinguishing vegetation species.The absorption of chlorophyll in the coastal blue band facilitates discrimination based on leaf chlorophyll content while extended NIR (1 and 2) broadens the spectrum for vegetation analysis.

Table 2 :
Summary results of the maximum likelihood classification showing only the bracken class and its accuracies (OA-Overall Accuracy, UA-User's Accuracy, PU-Producers's Accuracy).

Table 3 :
The confusion matrices from the maximum likelihood classification of the strategically positioned bands WV-2 image (A) and SPOT 5 image (B) -(UA-User's Accuracy, PU-Producers's Accuracy).Visual inspection of the classified imagery (Figure