This article was downloaded by: [University of New England] On: 01 March 2012, At: 14:27 Publisher: Taylor & Francis Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
International Journal of Remote Sensing Publication details, including instructions for authors and subscription information: http://www.tandfonline.com/loi/tres20
Optimal band selection from hyperspectral data for Lantana camara discrimination a
a
a
Subhashni Taylor , Lalit Kumar , Nick Reid & Craig R. G. Lewis b a
Ecosystem Management, School of Environmental and Rural Science, University of New England, Armidale, New South Wales, 2351, Australia b
Animal Genetics and Breeding Unit, University of New England, Armidale, New South Wales, 2351, Australia Available online: 01 Mar 2012
To cite this article: Subhashni Taylor, Lalit Kumar, Nick Reid & Craig R. G. Lewis (2012): Optimal band selection from hyperspectral data for Lantana camara discrimination, International Journal of Remote Sensing, 33:17, 5418-5437 To link to this article: http://dx.doi.org/10.1080/01431161.2012.661093
PLEASE SCROLL DOWN FOR ARTICLE Full terms and conditions of use: http://www.tandfonline.com/page/terms-andconditions This article may be used for research, teaching, and private study purposes. Any substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to anyone is expressly forbidden. The publisher does not give any warranty express or implied or make any representation that the contents will be complete or accurate or up to date. The accuracy of any instructions, formulae, and drug doses should be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims, proceedings, demand, or costs or damages whatsoever or howsoever caused arising directly or indirectly in connection with or arising out of the use of this material.
International Journal of Remote Sensing Vol. 33, No. 17, 10 September 2012, 5418–5437
Optimal band selection from hyperspectral data for Lantana camara discrimination
Downloaded by [University of New England] at 14:27 01 March 2012
SUBHASHNI TAYLOR*†, LALIT KUMAR†, NICK REID† and CRAIG R. G. LEWIS‡ †Ecosystem Management, School of Environmental and Rural Science, University of New England, Armidale, New South Wales 2351, Australia ‡Animal Genetics and Breeding Unit, University of New England, Armidale, New South Wales 2351, Australia (Received 2 March 2011; in final form 5 September 2011) The primary objective of this research was to determine the optimal hyperspectral wavelengths based on spectroscopy data over the spectral range of 450–2500 nm for the detection of the invasive species Lantana camara L. (lantana) from seven of its co-occurring species. A procedure based on statistical analysis of the reflectance and the first derivative reflectance (FDR) identified 86 and 18 bands, respectively, where lantana significantly differed from its co-occurring species. The effectiveness of the identified optimal bands was then evaluated using Hyperion imagery. The original Hyperion image with 155 bands gave an overall accuracy of 80% compared to 77% and 76% from the 86- and 18-band spectral subsets, respectively. A pairwise comparison of the three error matrices showed no significant difference in the accuracy achieved. The FDR analysis combined with the statistical analysis proved to be a useful procedure for data reduction by refining the discrimination to fewer optimal bands for lantana detection with no adverse impact on classification accuracy.
1. Introduction Invasive species are a major threat to the Earth’s biodiversity because they often dramatically affect the structure and function of ecosystems (Binggeli 1996). The impacts, both economic and environmental, of such species have been documented by several authors (Vitousek et al. 1996, Mack et al. 2000, Day et al. 2003, Henderson et al. 2006). One such species is lantana (Lantana camara L.), which is regarded as one of the world’s 10 worst weeds (Sharma et al. 2005). In Australia, lantana currently covers more than 4 million ha (Day et al. 2003) and costs the Australian grazing industry in excess of $121 million per annum in lost production and management costs (Johnson 2007). Lantana negatively affects more than 1300 native species including 279 plant and 93 animal species listed as rare or threatened in Australia (Johnson 2007). Management strategies include determining the extent and level of infestations, reporting new infestations and investigating the dynamics of the spread of lantana *Corresponding author. Email:
[email protected] The Animal Genetics and Breeding Unit is a joint venture between the University of New England and Industry and Investment NSW. International Journal of Remote Sensing ISSN 0143-1161 print/ISSN 1366-5901 online © 2012 Taylor & Francis http://www.tandf.co.uk/journals http://dx.doi.org/10.1080/01431161.2012.661093
Downloaded by [University of New England] at 14:27 01 March 2012
Hyperspectral bands for Lantana camara detection
5419
populations (ARMCANZ 2000). Such strategies require detailed and accurate maps of lantana distribution. Traditional methods of ground-based mapping tend to be prohibitively expensive in terms of time, personnel and cost, particularly if large and remote areas need to be investigated. Remote sensing is a useful tool for mapping and monitoring invasive species, and detection abilities have improved because sensor technology has become more advanced in terms of enhanced spatial resolution (Lass et al. 2005). Remotely sensed imagery can sample 100% of an area along with the possibility of rapid return intervals, which eliminates the problem of out-of-date data for rapidly spreading weeds (Andrew and Ustin 2006). This technology has been particularly useful in studying the location, extent of infestations and rate of spread of invasive plant species (Lass et al. 2002, Noujdina and Ustin 2008). The field of remote sensing has recently seen a steep increase in the number of spectral bands in acquired data, going from multispectral to hyperspectral. Multispectral systems record reflectance in a few bands of the electromagnetic spectrum (Jensen 2005), while hyperspectral sensors acquire data in hundreds of spectral bands (Goetz 2002). Hyperspectral data present certain advantages over multispectral data in the discrimination between plant species. This is because many species have characteristic features in their spectral signature, which occur in very narrow bandwidths and as such can only be ‘sensed’ by narrow-band sensors. The narrow bandwidths of hyperspectral sensors deliver more information about the fine spectral features of vegetation, thus permitting a range of more precise applications such as invasive species identification. However, the question arises whether all of the bands are really required for a particular application. Neighbouring bands in hyperspectral data are often strongly correlated and this may mean that they are providing similar information. Some studies have shown that for classification tasks, it is often sufficient to select only a dozen specific bands for adequate results (Serpico and Bruzzone 2001, Bruce et al. 2002). A practical approach would be to select the best bands for the user’s particular application to save time and costs associated with data processing. Studies that identify optimal hyperspectral bands for various applications thus serve a useful purpose. Although many researchers have used a variety of techniques for determining the optimal narrow bands for species discrimination, there is no specific methodology that is best suited to this task. Previous studies have used techniques such as multiple and stepwise discriminant analyses (Galvao et al. 2005, Hamada et al. 2007), principal component analysis (Thenkabail et al. 2004) as well as derivatives of reflectance spectra (Becker et al. 2005), with some studies focusing on optimal band selection for invasive species detection (Laba et al. 2005, Hamada et al. 2007). In the latter studies, researchers used either a statistical technique or a derivative analysis for the purposes of identifying the best bands for invasive species detection. The non-parametric Mann–Whitney U-test has been used by various researchers (Schmidt and Skidmore 2001, 2003, Artigas and Yang 2006) to identify the best wavebands for detection of their species of interest. Furthermore, first derivative analysis can be used to enhance differences between species because, according to Laba et al. (2005), such differences appear much more clearly in graphs of first-order derivatives. This type of analysis amplifies absorption features that might be very useful in discriminating an invasive species such as lantana from other species that occur in the same area because, as Kumar and Skidmore (1998) point out, ‘any differences observed in the first derivative curve are more likely to be due to leaf chemical composition, leaf structure or water content and do not depend on the absolute magnitude of
Downloaded by [University of New England] at 14:27 01 March 2012
5420
S. Taylor et al.
reflectance’. A combination of both first derivative analyses and the Mann–Whitney U-test may prove powerful in identifying the best narrow wavebands for lantana detection. In the future, satellites may carry specialist sensors that collect data for targeted applications such as invasive species mapping. On the other hand, hyperspectral sensors such as Hyperion may still be employed for the extraction of optimal narrow bands depending on the users’ needs (Thenkabail et al. 2002). Thus, prior knowledge of the optimal narrow bands for various applications is useful. These sorts of applications could contribute towards potential improvements in accuracies over broadband sensors as well as a reduction in costs related to image acquisition and processing (Thenkabail et al. 2004). Studies that aim to identify optimal bands for specific applications can contribute useful information towards sensor design, particularly in the light of the comment made by De Backer et al. (2005) that ‘sensors are being developed with limited degrees of freedom with respect to optical filter settings’. Therefore, prior knowledge of the required band settings may contribute useful information towards the design of dedicated sensors. However, a more immediate use of such information would be to assist land managers involved in invasive species mapping to select the best bands from hyperspectral sensors such as Hyperion for their purposes. The aim of this research was to determine optimal narrow bands for lantana discrimination from surrounding vegetation using spectrometer data. The study used a combination of two techniques, the Mann–Whitney U-test and the first derivative analysis, to identify the locations of unique spectral features for detecting lantana. Finally, the usefulness of these selected bands was evaluated using Hyperion data to determine the impact on classification accuracy.
2. Materials and methods 2.1 Study site and data collection This study was conducted in north-eastern New South Wales (NSW), Australia (figure 1). Past logging has led to an altered forest structure and a reduction in canopy cover, which encouraged weed establishment, particularly lantana. The level of lantana infestation is high, especially in previously logged areas where it suppresses forest regeneration and succession. Large patches of lantana that are greater than a Hyperion 30 m × 30 m pixel exist. There are also large areas of rainforest with no gaps in the canopy where it does not occur. Field data were collected to coincide with image acquisition on 24 May 2005. This was carried out at various sample locations, and typical vegetation categories were identified and noted. After careful evaluation of the vegetation data from these sample sites, 72 sites were chosen based on the vegetation structure. Of these 72 sites, some consisted of large pure stands of lantana and some were free of lantana infestation. An estimated scale of low to high density was assigned to the level of lantana. Low density reflected no lantana or the presence of odd, small individual plants; medium density was an intermediate level of lantana, where healthy growth was visible but it was still low and had not grown upon itself; and high density represented extreme infestations of lantana often over 2 m in depth. Since the aim of this study was to identify wavelength regions that were optimum for lantana discrimination, only high-density areas were selected to ensure that pure lantana pixels were included. The area selected for each site was homogeneous within a 60 m × 60 m square, the equivalent of four Hyperion pixels. The data collected
Hyperspectral bands for Lantana camara detection
5421
Queensland
N
Toonumbar NP New South Wales
Richmond Range NP
Downloaded by [University of New England] at 14:27 01 March 2012
25 km
0
Surrounding National Parks
Northern Territory
Queensland
Western Australia South Australia
New South Wales Victoria
0
1600 km
Tasmania
Data collection sites
0
2 km
Figure 1. Map showing the location of study area and surrounding national parks. The extent of the Hyperion imagery is shown by the large rectangle and the black dots within it represent data collection sites.
for each square included the GPS coordinates, main vegetation species, percentage canopy cover and ground cover. Main vegetation species was the most common tree species, with a note made of lower storey species if relevant. Canopy cover and ground cover were estimated visually. These data were used for training and validation of the classifier. The categories were generalized into two classes: lantana and non-lantana. One additional class labelled pasture was created and regions of interest (ROIs) were defined by visual means from the imagery as it was a spectrally distinct class (shown as white sections on the Hyperion image in figure 1). Reflectance spectra of lantana and seven common co-occurring species were collected (table 1). They were described as the dominant species in the study area by the park rangers who work in this area of national park. Reflectance measurements were taken from five different plants of each species. Ten leaf samples were selected from each plant and the tree number that the sample was collected from was also recorded. Five spectral readings were recorded for each leaf, giving a total of 250 readings per species. The spectral signatures from all five trees of each species were compared and if a set of spectra from one tree did not match with the others or looked visually different, they were discarded. This was done to remove any outliers that may have resulted from noise or shaking during measurement. Table 1 shows the number of samples of each species used in subsequent analyses. Spectral measurements were made in the field immediately after leaf collection. Five different trees were used to collect the leaf samples for each species to eliminate the possibility of a particular tree being diseased. Leaves that showed visual signs of stress or were diseased were not used. Care was
5422
S. Taylor et al. Table 1. Tree species and sample size.
Common name
Downloaded by [University of New England] at 14:27 01 March 2012
Crofton weed Giant stinging tree Flooded gum Tallow-wood Grey gum Grey ironbark Lantana Scentless rosewood
Scientific name
Code
Family
No. of samples
Ageratina adenophora Dendrocnide excelsa Eucalyptus grandis Eucalyptus microcorys Eucalyptus punctata Eucalyptus siderophloia Lantana camara Synoum glandulosum
AA DE EG EM EP ES LC SG
Asteraceae Urticaceae Myrtaceae Myrtaceae Myrtaceae Myrtaceae Verbenaceae Meliaceae
245 248 244 196 236 245 217 230
also taken to use only mature leaves from each species since there were large variations between leaves of different ages of a single species. Indeed, spectral variability related to growth stage may be greater within species than between species (Okin et al. 2001). Once leaves are past their period of rapid growth and are mature, few differences occur until senescence (Kumar and Skidmore 1998). Therefore, for comparative purposes between species, it was prudent to select mature leaves. The reflectance spectra were collected with a full-range portable spectroradiometer manufactured by Analytical Spectral Devices (ASD Inc., Boulder, CO, USA). The instrument collected data over the range of 350–2500 nm. The full-width-halfmaximum (FWHM) spectral resolution of this spectroradiometer is 3 nm for the visible to near-infrared region (350–1000 nm) and 10 nm for the near- and shortwaveinfrared region (1000–2500 nm) (ASD 2002). The study area is quite hilly and the illumination can vary considerably from place to place. A leaf clip device with a builtin artificial light source was used for reflectance measurements to avoid errors caused by variations in solar illumination. The field of view of the fibre-optic cable was 25◦ ; the distance from the cable to the leaf surface was 0.33 cm and the area of the leaf that was measured was 1.16 cm2 . Optimization of the instrument was carried out approximately every 20 min using the white background standard provided with the leaf clip device. Spectral data were recorded as reflectance values using the software provided by ASD and then imported into a spreadsheet (Microsoft Office Excel 2007) for further analyses. The reflectance spectra were too noisy at the lower extreme of the spectral range of the spectrometer and thus only the spectral range between 450 and 2500 nm was analysed. All field spectra were resampled to match the wavelengths and bandpass of the Hyperion data for later comparison (Jupp and Datt 2004). 2.2 First derivative reflectance The first derivative reflectance (FDR) was calculated on the spectra and used for refining the discrimination analysis. This type of analysis can be used to differentiate potentially unique band locations of absorption and reflection features that might be very useful in discriminating between species. A first difference transformation of the reflectance spectrum calculates the slope values from the reflectance and can be derived from equation (1) (Dawson and Curran 1998): FDRλ(i) =
Rλ(j+1) − Rλ(j) , λ
(1)
Hyperspectral bands for Lantana camara detection
5423
where FDR is the first derivative reflectance at a wavelength i midpoint between wavebands j and j + 1, Rλ(j) is the reflectance at the j waveband, Rλ(j+1) is the reflectance at the j + 1 waveband and λ is the difference in wavelengths between j and j + 1.
Downloaded by [University of New England] at 14:27 01 March 2012
2.3 Statistical analysis The reflectance and the first derivative spectra from the eight species were statistically analysed to identify the regions of greatest difference between lantana and the other seven species. The reflectance at these wavelengths was specific to lantana and could potentially be used for identifying this invasive species. Prior to analysis, data were further checked and distributions were assessed using the JMP (SAS 2009) and SAS software (SAS 1990) (SAS, Cary, NC). Various researchers (Schmidt and Skidmore 2001, 2003, Artigas and Yang 2006) have used the non-parametric Mann–Whitney U-test for testing whether significant differences exist between the mean of the reflectance for each measured waveband. A Kruskal–Wallis test was first used (‘proc npar1way’ procedure in SAS) to assess whether there were differences between lantana and all other species in the data. After a significant Kruskal–Wallis test was returned, the Mann–Whitney pairwise comparisons (lantana vs. all the other species, in turn) were undertaken (also using ‘proc npar1way’ in SAS using the ‘Wilcoxon’ option to achieve the Mann–Whitney test). The null hypothesis of the Mann–Whitney test is that the comparison between lantana and the other species was not significant for each wavelength. This being the case, the null hypothesis was H0 : ηspecies 1(i) = ηspecies 2(i) ,
(2)
where species 1 and species 2 are compared at wavelength i. The alternative hypothesis was that the reflectance means were not equal, i.e. H1 : ηn(i) = ηn + 1(i) .
(3)
The motivation for using the U-test was that it is non-parametric and does not assume a normal distribution of samples. The unequal number of samples per species (table 1) was assumed not to influence the test, as the number of samples was large (Lehmann 2006). The Mann–Whitney U-test was carried out on the raw reflectance data for all combinations of lantana with the other seven species and all bands from 450 to 2500 nm. After Mann–Whitney analysis, the p-values were corrected in SAS for multiple testing using Bonferroni corrections (‘proc multtest’ in SAS). The corrected p-values were deemed significant only if p = 0.001, to be as conservative as possible. 2.4 Image processing and classification Hyperion is a pushbroom sensor with two spectrometers and a single telescope (Pearlman et al. 2003, Ungar et al. 2003). It is carried by the NASA Earth Observing 1 (EO-1) satellite. Hyperion provides a high-resolution hyperspectral imager capable of resolving 242 spectral bands over 400–2500 nm with a spatial resolution of 30 m and a sampling interval of 10 nm. The instrument can image a 7.5 × 100 km area per image, and provides detailed spectral mapping across all channels with high radiometric accuracy. Although the level 1 radiometric (1R) product has a total of 242 bands, this study used only 155 ‘stable’ bands. Bands prone to atmospheric scattering (USGS
5424
S. Taylor et al. Table 2. 155 stable Hyperion bands. Region
Band number
Wavelength (nm)
VNIR SWIR
10−57 81−97 101−119 134−164 182−221
447.9−925.9 952.9−1114.3 1154.7−1336.2 1487.6−1790.2 1971.8−2365.2
Downloaded by [University of New England] at 14:27 01 March 2012
Note: VNIR, visible and near-infrared; SWIR, shortwave infrared.
2003) as well as those affected by major water vapour absorption areas and features due to O2 and CO2 (Datt et al. 2003) were removed from further analysis. Table 2 shows the final 155 bands chosen for further analysis. Atmospheric correction of the Hyperion scene was performed with ENVI’s (Research Systems Inc., Boulder, CO, USA) fast line-of-sight atmospheric analysis of spectral hypercubes (FLAASH). This procedure converts the data from radiance to apparent surface reflectance (Felde et al. 2003, Matthew et al. 2003) and is an important procedure for spectral analysis-based mapping methods (Galvao et al. 2009). The image was pre-processed prior to conversion using the following steps: repair of ‘bad’ pixel values, fixing of out-of-range data, fixing of outliers, de-smiling and de-striping (Datt et al. 2003). The researchers found that the pre-processing steps were the best noise management strategy and selection of 155 stable bands provided a simple but effective way of avoiding any residual noise that may remain after the processing. A Landsat thematic mapper (TM) rectified image was used as a base map for georectification. The image had been terrain corrected prior to delivery and provided systematic radiometric and geometric accuracy by incorporating ground control points (GCPs) while employing a digital elevation model (DEM) for topographic accuracy (USGS 2010). The adequate spatial resolution and rectification provided by the 30 m pixel size ensured that geographic fidelity was consistent throughout the image. Ten GCPs were selected with good dispersion throughout the image. A first-order polynomial transformation with nearest neighbour resampling was used to retain as much spectral fidelity as possible. The resulting image had an estimated total root mean square (RMS) error of 0.0795 pixels or about 2.4 m. A sub-scene covering the study site was extracted from the Hyperion image using the ENVI (2008) imageprocessing software package. This sub-scene was used in subsequent analyses. Figure 2 shows the spectrum extracted from the atmospherically corrected image as well as the field reflectance spectrum of lantana. Two spectral subsets of the original 155-band image were created using only the bands identified as being optimal for lantana detection from the statistical analysis of the reflectance and the first derivative data. These two spectral subsets as well as the original 155-band images were classified using the spectral angle mapper (SAM) algorithm. This algorithm determines the spectral similarity between two spectra by calculating the angle between them, treating them as vectors in a space with dimensionality equal to the number of bands (Kruse et al. 1993). The SAM is designed for use primarily with hyperspectral data, and has broad application in hyperspectral remote sensing of vegetation (Ustin et al. 2002, de Lange et al. 2004, Eckert and Kneubühler 2004, Lumme 2004). The reference spectra for the SAM can either be taken from laboratory or field measurements or be extracted directly from the image.
Downloaded by [University of New England] at 14:27 01 March 2012
Reflectance (%)
Hyperspectral bands for Lantana camara detection 60 55 50 45 40 35 30 25 20 15 10 5 0 400
L. camara field spectrum
800
1200 1600 Wavelength (nm)
5425
L. camara image spectrum
2000
2400
Figure 2. Extracted reflectance spectrum from atmospherically corrected Hyperion sample pixel containing lantana and the field reflectance spectrum of lantana leaf.
For the classification, three specific land-cover classes, pasture, lantana and nonlantana, were identified based on field knowledge. Pasture was chosen because it was identifiable in the image and therefore its spectrum could be easily extracted from the image for classification. Non-lantana was created because we were only interested in lantana as a target species; therefore, all other vegetation types were put into a single class labelled non-lantana. Twenty-nine sites that were identified from field knowledge as pure lantana and non-lantana (non-lantana consisted of a mixture of different species but no lantana) were selected for training the classifier. An additional 14 training sites were chosen for the spectrally distinct pasture class from the imagery. Spectral reflectance was extracted for these classes from the image by creating ROIs based on visual inspection of the Hyperion image and field data. Mean spectra were then extracted for each ROI from the reflectance data to act as endmembers for each class. These were then used as reference spectra in the SAM classifier. The SAM compares the angle between the reference spectrum vector and each pixel vector in n-dimensional space. Smaller angles represent closer matches to the reference spectrum. Pixels further away than the specified maximum angle threshold in radians are not classified. In this study, the spectral angle threshold was set to 0.3 radians for all classifications. The accuracy assessment data consisted of 57 verification sites for the three landcover classes. These data were used to derive an error matrix, and the kappa (κ) value was calculated. The kappa analysis resulted in a khat (Cohen’s kappa coefficient) statistic, which gave a measure of agreement based on the difference between the actual agreement of the classification (i.e. agreement between computer classification and reference data as shown by the major diagonal elements) and the chance agreement, which was indicated by the product of the row and column marginals of the error matrix (Congalton et al. 1983, Congalton 1991). Producer accuracy (measure of omission error) and user accuracy (measure of commission error) (Story and Congalton 1986) were also calculated for each category. Once the error matrix has been developed, further statistical analyses can be carried out to test if two independent error matrices are significantly different (Congalton and Green 2009). This statistical analysis enabled us to compare the effectiveness (overall accuracy) achieved by the three images.
5426
S. Taylor et al.
3. Results
Figure 3 shows the plots of wavelength versus mean reflectance spectra for all species. The spectra have not been offset so that differences in reflectance can be easily observed. In general, all the curves exhibited differences in the magnitude of reflectance, but the overall shape of the curves was similar. The similarity in shape can be attributed to the relatively small number of variables such as chlorophyll and water content that influence the spectral properties of vegetation (Price 1994). However, there were many subtle differences in terms of absolute reflectance, depths of absorption features and the relative position of change in terms of the wavelength. Some species also exhibited crossovers in reflectance. Variation in the relative amounts of chlorophyll, water content and cell-to-air space ratio gives rise to the amplitude differences (Smith and Blackshaw 2003). In the visible part of the spectrum (400–700 nm), lantana showed similar reflectance to Ageratina adenophora, Eucalyptus grandis and Eucalyptus punctata. The reflectance of all species increased after 701 nm, with lantana rising to the third highest position at 1335 nm and then decreasing again at 1487 nm. Beyond 1487 nm, the reflectance of all species showed two minor peaks, with the reflectance values of lantana being among the highest of the eight species, along with Eucalyptus microcorys and Synoum glandulosum. 3.2 Statistical difference A comparison of lantana reflectance with each co-occurring species meant seven pairwise comparisons to identify regions of the spectrum showing the greatest difference between lantana and the other species. The Mann–Whitney U-tests were carried out for each wavelength from 450 to 2500 nm and the results are shown in figures 4(a)–(g). The mean reflectance curve of lantana and the species it was being compared with was overlaid on the plot to visualize the positions of the main features where lantana returned a statistically significant difference in reflectance from the other species (indicated by the shaded areas). Figures 4(a)–(g) show that there are large sections of the spectrum where lantana differs significantly from the other seven species. A. adenophora (figure 4(a)) reflectance was significantly different from
Reflectance (%)
Downloaded by [University of New England] at 14:27 01 March 2012
3.1 Visual difference
60 55 50 45 40 35 30 25 20 15 10 5 0 400
A. adenophora
D. excelsa
E. grandis
E. microcorys
E. punctata
E. siderophloia
L. camara
S. glandulosum
800
1200 1600 Wavelength (nm)
2000
2400
Figure 3. Mean reflectance spectra for all species (resampled to match the wavelengths and bandpass of the Hyperion data).
Hyperspectral bands for Lantana camara detection (a)
(e)
1600 1200 Wavelength (nm)
Reflectance (%)
(g)
2000
E. microcorys
1200 1600 Wavelength (nm)
800
2400
L. camara
2000
E. siderophloia
1200 1600 Wavelength (nm)
Significant (p = 0.001)
60 55 50 45 40 35 30 25 20 15 10 5 0 400
(b)
2400
L. camara
800
2000
S. glandulosum
1200 1600 Wavelength (nm)
2000
2400
Significant (p = 0.001)
50 45 40 35 30 25 20 15 10 5 0 400
(d)
Reflectance (%) 800
Significant (p = 0.001)
50 45 40 35 30 25 20 15 10 5 0 400
L. camara
800
800
800
L. camara
2000
E. punctata
1600 1200 Wavelength (nm)
Significant (p = 0.001)
50 45 40 35 30 25 20 15 10 5 0 400
E. grandis
1200 1600 Wavelength (nm)
Significant (p = 0.001)
50 45 40 35 30 25 20 15 10 5 0 400
(f)
Reflectance (%)
Reflectance (%)
800
Significant (p = 0.001)
60 55 50 45 40 35 30 25 20 15 10 5 0 400
Reflectance (%)
Downloaded by [University of New England] at 14:27 01 March 2012
(c)
A. adenophora
Reflectance (%)
Reflectance (%)
Significant (p = 0.001)
50 45 40 35 30 25 20 15 10 5 0 400
5427
L. camara
2000
D. excelsa
1200 1600 Wavelength (nm)
2400
2400
L. camara
2000
2400
L. camara
2400
Figure 4. Pairwise test results for the reflectance spectra. (a) L. camara vs. A. adenophora, (b) L. camara vs. E. grandis, (c) L. camara vs. E. microcorys, (d) L. camara vs. E. punctata, (e) L. camara vs. E. siderophloia, (f ) L. camara vs. D. excelsa and (g) L. camara vs. S. glandulosum. The grey-shaded regions indicate the wavelengths where reflectance differed significantly between species at p = 0.001 (Bonferroni adjusted).
lantana reflectance at all wavelengths except for narrow regions at 508, 620–640 and 1971 nm. Of all the eucalypt species (figures 4(b)–(e)), E. microcorys was most similar to lantana as the pairwise comparisons returned a non-significant difference at more wavelengths than the other eucalypt species. Dendrocnide excelsa (figure 4(f ))
5428
S. Taylor et al. Frequency of significance
L. camara
50
7
45
Reflectance (%)
5
35 30
4
25 3
20 15
2
Frequency of significance
6
40
10 1
Downloaded by [University of New England] at 14:27 01 March 2012
5 0 400
0 800
1200
1600
2000
2400
Wavelength (nm)
Figure 5. Frequency plot of statistically significant differences (using the Mann–Whitney U-test), between the reflectance for all possible pairs at every channel, overlaid with the mean reflectance curve of lantana. Grey-shaded areas and the right plot axis indicate wavelength regions where lantana is significantly different from the other species.
reflectance was similar to lantana reflectance in only two narrow regions, namely 457–488 and 650–681 nm. These two species were distinguishable at all other wavelengths. S. glandulosum (figure 4(g)) was most similar to lantana as the pairwise comparisons were non-significant at more wavelengths than the other species. A summary of all significant differences in the reflectance data between lantana and the other species is given in figure 5. There were a total of seven possible combinations; therefore, if lantana can be discriminated from the other species at a particular wavelength, then the frequency at that wavelength would be 7. The mean reflectance curve of lantana was overlaid on the plot to visualize the positions of the main features where discrimination was returned by all seven species. This figure shows that lantana can be discriminated from all of its seven co-occurring species in the visible/near-infrared region of wavelength 518–528, 569–610, 691 and 762–1325 nm. In the mid-infrared region, the highest number of discriminations was seen at 1547–1789 nm as well as at 1981 nm. This gave a total of 86 bands where lantana reflectance was significantly different from its co-occurring species. Figure 6 shows the first derivative of the mean reflectance spectra for each species. The derivative curves were offset to identify the regions of difference between species. The Mann–Whitney U-tests were also carried out for the first derivatives of reflectance and the results are shown in figure 7. The mean reflectance curve of lantana and the comparator species is overlaid on each plot. The first derivative data were analysed despite the fact that the raw reflectance data were significantly different at many regions of the spectrum. This procedure was carried out to identify unique absorption features exhibited by lantana that were not apparent in the reflectance data. The wavelength locations of such features could considerably narrow down the number of optimal bands for lantana detection to fewer than 86. Some differences between lantana and the other species are marked with a vertical line in figure 6. For example, the depth of the dip between 559 and 569 nm differs for all eight species. Another position of obvious difference was observed at around 590 nm, where most eucalypt species (except E. grandis) as well as lantana showed pronounced double peaks not seen in
5429 2304 nm
2233 nm
1759 nm
1729 nm
1154 nm
983 nm
590 nm
559 nm
First derivative (offset for clarity)
Hyperspectral bands for Lantana camara detection
SG LC ES EP EM EG DE
Downloaded by [University of New England] at 14:27 01 March 2012
AA
400
800
1200 1600 Wavelength (nm)
2000
2400
Figure 6. First derivative curves of the mean reflectance spectra for all species (species codes as shown in table 1). Vertical lines indicate wavelengths at which differences between the eight species can be seen.
other species. The wavelength region of about 700–711 nm (red-edge) showed differences between all eight species and the depth of the absorption features at 983, 1154, 1729 and 2233 nm also varied. E. microcorys shows an absorption feature at 1689 nm, which is not shown by any of the other species. Lantana and S. glandulosum show features between 1759 and 1779 nm, which are not shown by any of the other species. At 2304 nm, lantana, D. excelsa and A. adenophora show a pronounced absorption feature. When pairwise test results for first derivative data (figure 7) were compared with pairwise test results for the raw reflectance data (figure 4), the regions of significant difference were fewer in the first derivative graphs than in the reflectance graphs. For example, the pairwise comparisons of the reflectance of lantana and A. adenophora (figure 4(a)) were significant over large wavelength ranges. In fact, except for three narrow regions, pairwise tests were significant for almost the whole spectrum. However, for the first derivative curve of the same species pair (figure 7(a)), significance was obtained at fewer points. The same observation could be made for all other comparisons (figures 7(b)–(g)). This indicates that most of the differences in the reflectance spectra were due to the magnitude of reflectance. The graphs in figures 4 and 7 correspond and therefore can be compared directly. Figure 8 summarizes all significant differences in the first derivative data between lantana and the other species. It was obtained in a similar way as described for figure 5. Compared to figure 5, this graph shows fewer wavelengths where lantana could be separated from the other seven species, specifically wavelengths in the visible/near-infrared region at 559–569, 701–711 and 732 nm. In the mid-infrared region, lantana could be discriminated at 1759–1779, 2031–2041, 2092–2142, 2273 and 2304 nm. This gave a total of 18 bands where lantana could be discriminated from cooccurring species. The wavelength locations of unique spectral features exhibited by lantana became more noticeable with first derivative analysis and some of these wavelengths coincided with the absorption features identified in figure 6 (559–569, 701, 1759–1779 and 2304 nm).
5430 (a)
1200 1600 Wavelength (nm)
800
60 55 50 45 40 35 30 25 20 15 10 5 0 400
(b)
2400
L. camara
2000
E. siderophloia
2400
L. camara
800
1600 1200 Wavelength (nm)
Significant (p = 0.001)
800
2000
S. glandulosum
1600 1200 Wavelength (nm)
2000
2400
Significant (p = 0.001)
50 45 40 35 30 25 20 15 10 5 0 400
(d)
800
800
800
2000
2000
D. excelsa
1200 1600 Wavelength (nm)
2400
L. camara
E. punctata
1600 1200 Wavelength (nm)
Significant (p = 0.001)
50 45 40 35 30 25 20 15 10 5 0 400
L. camara
E. grandis
1200 1600 Wavelength (nm)
Significant (p = 0.001)
50 45 40 35 30 25 20 15 10 5 0 400
(f)
Reflectance (%)
Reflectance (%)
(g)
2000
E. microcorys
1200 1600 Wavelength (nm)
Significant (p = 0.001)
50 45 40 35 30 25 20 15 10 5 0 400
L. camara
Reflectance (%)
Reflectance (%) Reflectance (%)
(e)
Reflectance (%)
800
Significant (p = 0.001)
60 55 50 45 40 35 30 25 20 15 10 5 0 400
A. adenophora
Reflectance (%)
Significant (p = 0.001)
50 45 40 35 30 25 20 15 10 5 0 400
(c)
Downloaded by [University of New England] at 14:27 01 March 2012
S. Taylor et al.
2400
L. camara
2000
2400
L. camara
2400
Figure 7. Pairwise test results for the first derivative of the reflectance spectra. (a) L. camara vs. A. adenophora, (b) L. camara vs. E. grandis, (c) L. camara vs. E. microcorys, (d) L. camara vs. E. punctata, (e) L. camara vs. E. siderophloia, (f ) L. camara vs. D. excelsa and (g) L. camara vs. S. glandulosum. The grey-shaded regions indicate the wavelengths where species differed significantly at p = 0.001 (Bonferroni adjusted).
3.3 Image classification The classification results are shown in table 3 together with the 155-band classified image. The original Hyperion image performed marginally better with an overall accuracy that was 3.3 and 3.9% higher than that provided by the 86- and 18-band spectral
Hyperspectral bands for Lantana camara detection
5431
L. camara
Frequency of significance
45
6
Reflectance (%)
40 35
5
30
4
25 20
3
15
2
10
1
5
Downloaded by [University of New England] at 14:27 01 March 2012
Frequency of significance
7
50
0 400
0 800
1200 1600 Wavelength (nm)
2000
2400
Figure 8. Frequency plot of statistically significant differences (using the Mann–Whitney U-test), between the first derivative of reflectance spectra for all possible pairs at every channel, overlaid with the mean reflectance curve of lantana. Grey-shaded areas and the right plot axis indicate wavelength regions where lantana is significantly different from the other species. Table 3. Overall accuracies, kappa coefficients and producer and user accuracies of the three classifications of the Hyperion image. Lantana Overall accuracy PA (%) Kappa (%) Classified 155-band Hyperion image
Original image (155 bands) 86-band subset (raw reflectance data) 18-band subset (first derivative data)
Pasture
Nonlantana
UA PA UA PA UA (%) (%) (%) (%) (%)
80.1
0.69
83.3 84.4 73.9 100 83.6 67.6
76.8
0.64
80.8 82.9 67.4 91.2 81.2 65.2
76.2
0.64
81.8 77.8 71.7 100 77.8 66.7
Notes: Pasture (red), forest (green) and lantana (blue). Black areas on the left of the image represent clouds and in the top half of the image represent areas that were in shadow due to an escarpment. These were masked out during classification. PA, producer accuracy; UA, user accuracy.
subsets, respectively. This image also yielded a producer accuracy that was 2.5% higher than that obtained by both the spectral subsets as well as a user accuracy that was 1.5 and 6.6% higher than that obtained by the 86- and 18-band spectral subsets, respectively. However, despite the slightly better performance shown by the original Hyperion image, a pairwise comparison of the three error matrices showed no significant difference in this accuracy.
5432
S. Taylor et al.
Downloaded by [University of New England] at 14:27 01 March 2012
4. Discussion Although the reflectance spectra of the eight species (figure 3) look very similar, further analysis showed a striking statistical difference between lantana and almost all other seven species, perhaps with the exception of S. glandulosum. This difference covered virtually the whole spectrum from 450 to 2500 nm. This was most likely due to the species having very different physiognomy and physiology. The similarity between lantana and S. glandulosum reflectance highlights the problem with species identification using reflectance data as two very different species can have similar reflectance. In an effort to assist in vegetation discrimination, some researchers have applied the Mann–Whitney U-test to reflectance and continuum-removed data to help identify unique spectral features of different vegetation types (Schmidt and Skidmore 2003, Psomas et al. 2005, Artigas and Yang 2006). This study showed that with a combination of the Mann–Whitney U-test and first derivative analysis, wavelength locations of the unique spectral features displayed by lantana could be identified. The effect of first derivative analysis was to resolve overlapping spectra so that features that were distinctive to lantana became much more clearly separated from the other species (Demetriades-Shah et al. 1990). This analysis allowed us to extract wavelengths with the most pertinent information for lantana detection, thereby reducing the 86 bands identified from reflectance data to 18 bands centred at 559, 569, 701, 711, 732, 1759, 1769, 1779, 2031, 2041, 2092, 2102, 2112, 2122, 2132, 2142, 2273 and 2304 nm. The maximum frequency of significant differences between lantana and the other seven species (figure 8) was located at these wavelengths. The important spectral regions for optimal lantana discrimination are at the point of maximum reflectance in the visual region, the edge leading up to the near-infrared plateau as well as some portions of the mid-infrared region. Some of these findings are consistent with other studies (Spanglet et al. 1998, Thenkabail et al. 2002, Schmidt and Skidmore 2003, Splajt et al. 2003) since a number of optimal bands identified in this study appear to have widespread use in vegetation studies. One such region, termed the ‘red-edge’ (690–720 nm), has been identified by these researchers to be sensitive to slight differences in plant morphology and useful in discriminating between various plant community types. Daughtry and Walthall (1998), who studied the difference in reflectance between Canabis sativa and other plant species, suggested that differences in the slope of the red to near-infrared transition may be used to discriminate between species. The exact wavelength of the red-edge is reported as being potentially useful in separating plant species (Gong et al. 1997, Cochrane 2000). Two of the bands identified for lantana detection (701 and 711 nm) fall within this region. However, this study also highlighted bands that were specific to lantana, particularly in the mid-infrared region. The classification results show that similar accuracies to full Hyperion bands can be achieved for lantana mapping by using only a few optimal bands identified through field spectroscopy. There was no significant difference in the accuracy achieved by the full Hyperion bands and that by the spectral subset made up of 18 optimal bands. Similar results were found by Serpico and Bruzzone (2001), Bruce et al. (2002), Thenkabail et al. (2004) and De Backer et al. (2005), who found that for classification tasks, it was often sufficient to select only a few specific optimal bands to obtain results that were as good as the total number of bands. In this study, the 18-band subset obtained from the first derivative analysis of spectra provided classification results that were similar to the results provided by the full Hyperion image, particularly the producer and user accuracies for lantana. Based on these results, the 18-band data set is recommended for lantana mapping because it provided an overall accuracy that
Downloaded by [University of New England] at 14:27 01 March 2012
Hyperspectral bands for Lantana camara detection
5433
was similar to that of the 155-band data set but with approximately 88% reduction in data volume. Although overall accuracy, producer accuracy and user accuracy are each important to understand how well a classification performed (Congalton 1991), user accuracy is a good measure of the classification’s reliability and subsequent success with a land manager. Based on this, it could be argued that the best data set would be the 155-band image since it provided a user accuracy that was 6% higher than the accuracy achieved with the 18-band data set. However, the reduction in data volume of the latter would reduce computational time in image processing as well as reduce the costs of data acquisition, which would make it an appealing proposition for resource managers. Potential users of hyperspectral data should be aware that reducing the number of spectral bands for a specific application, for example, lantana mapping, may result in reduced capabilities for identifying and mapping other materials based on their specific spectral properties. 5. Conclusions This study identified optimal narrow bands for lantana detection using spectroscopy data and then evaluated the effectiveness of these bands using the Hyperion imagery. Most of the eight vegetation types had a characteristic signature. A better understanding was gained about those parts of the electromagnetic spectrum that offer the greatest information content for discriminating between lantana and co-occurring vegetation. This study found that the combination of the Mann–Whitney U-test and the first derivative analysis reduced the number of bands required for image classification with no adverse impacts on classification accuracy. The results contribute towards reduction in data redundancy, data volumes, and time and resources involved in hyperspectral image interpretation and analysis. The procedures used in this study could be extended to the mapping of other invasive species. The results from this study are particularly useful with programmable sensors such as compact airborne spectrographic imager (CASI), a portable imaging spectrometer that can be used on small aircraft and has the capability to be programmed in advance with the required number of bands and bandwidths for targeted applications. Studies such as this which attempt to identify optimal bands for a specific purpose are well suited for use with such sensors. With the move towards sensor design for specialized tasks, future generations of satellites are likely to carry specialized sensors targeting selected wavebands for particular applications. This research may assist resource managers who are involved in lantana management in selecting a sensor suitable for the problem at hand. Furthermore, the increased computational burden and image acquisition costs associated with the hyperspectral imagery restrict the use of hyperspectral sensors in tactical unmanned aerial vehicle (UAV) applications. There is renewed interest in placing hyperspectral devices on UAVs and studies such as this may contribute towards the development of smaller and lighter instruments. The results suggest the potential to separate invasive species from surrounding vegetation on the basis of their leaf spectral characteristics, particularly with hyperspectral remote sensing. However, these need to be considered in context especially since leaf spectral characteristics are known to vary with phenological stage. Therefore, it is unlikely that a single distinctive spectral signature exists for each plant species or plant group. Additionally, identified optimal bands may be specific to the terrain where the research was carried out. Lantana is a serious weed worldwide and has been targeted for control for over a century. It has a significant economic and environmental impact
5434
S. Taylor et al.
Downloaded by [University of New England] at 14:27 01 March 2012
and is difficult to control. Optimal band selection for lantana detection has not been previously established. The methodology used in this study shows potential for the identification of optimal bands for large-scale lantana mapping at a national or even global level. These optimal bands were identified from data collected in late March. More research is needed to test the applicability of these bands across a variety of terrain as well as varying environmental conditions. Acknowledgements The authors are grateful to Steven King from the Department of Environment, Climate Change and Water for providing access to the study area and support in field work as well as his advice on species selection for this study. They also thank Cate Macgregor for technical and fieldwork assistance and Bisun Datt for his invaluable advice on the use of FLAASH for atmospheric correction of the Hyperion image. References AGRICULTURE AND RESOURCE MANAGEMENT COUNCIL OF AUSTRALIA AND NEW ZEALAND (ARMCANZ), 2000, Weeds of National Significance: Lantana (Lantana camara) Strategic Plan (Launceston: Australian & New Zealand Environment & Conservation Council and Forestry Ministers). ANALYTICAL SPECTRAL DEVICES, 2002, FieldSpec Pro User’s Guide (Boulder, CO: ASD). ANDREW, M.E. and USTIN, S., 2006, Spectral and physiological uniqueness of perennial pepperweed (Lepidium latifolium). Weed Science, 54, pp. 1051–1062. ARTIGAS, F.J. and YANG, J., 2006, Spectral discrimination of marsh vegetation types in the New Jersey meadowlands, USA. Wetlands, 26, pp. 271–277. BECKER, B.L., LUSCH, D.P. and QI, J., 2005, Identifying optimal spectral bands from in situ measurements of Great Lakes coastal wetlands using second-derivative analysis. Remote Sensing of Environment, 97, pp. 238–248. BINGGELI, P., 1996, A taxonomic, biogeographical and ecological overview of invasive woody plants. Journal of Vegetation Science, 7, pp. 121–124. BRUCE, L.M., KOGER, C.H. and LI, J., 2002, Dimensionality reduction of hyperspectral data using discrete wavelet transform feature extraction. IEEE Transactions on Geoscience and Remote Sensing, 40, pp. 2331–2338. COCHRANE, M.A., 2000, Using vegetation reflectance variability for species level classification of hyperspectral data. International Journal of Remote Sensing, 21, pp. 2075–2087. CONGALTON, R., 1991, A review of assessing the accuracy of classifications of remotely sensed data. Remote Sensing of Environment, 37, pp. 35–46. CONGALTON, R.G. and GREEN, K., 2009, Assessing the Accuracy of Remotely Sensed Data: Principles and Practices, pp. 1–183 (London: CRC Press). CONGALTON, R.G., ODERWALD, R.G. and MEAD, R.A., 1983, Assessing Landsat classification accuracy using discrete multivariate analysis statistical techniques. Photogrammetric Engineering and Remote Sensing, 49, pp. 1671–1678. DATT, B., MCVICAR, T.R., VAN NIEL, T.G., JUPP, D.L.B. and PEARLMAN, J.S., 2003, Preprocessing EO-1 Hyperion hyperspectral data to support the application of agricultural indexes. IEEE Transactions on Geoscience and Remote Sensing, 41, pp. 1246–1259. DAUGHTRY, C.S.T. and WALTHALL, C.L., 1998, Spectral discrimination of Canabis sativa L. leaves and canopies. Remote Sensing of Environment, 64, pp. 192–201. DAWSON, T.P. and CURRAN, J.P., 1998, A new technique for interpolating the reflectance red edge position. International Journal of Remote Sensing, 19, pp. 2133–2139. DAY, M.D., WILEY, C.J., PLAYFORD, J. and ZALUCKI, M.P., 2003, Lantana: Current Management Status and Future Prospects (Canberra: Australian Centre for International Agricultural Research (ACIAR) Monograph).
Downloaded by [University of New England] at 14:27 01 March 2012
Hyperspectral bands for Lantana camara detection
5435
DE BACKER, S., KEMPENEERS, P., DEBRUYN, W. and SCHEUNDERS, P., 2005, Band selection for hyperspectral remote sensing. IEEE Geoscience and Remote Sensing Letters, 2, pp. 319–323. DE L ANGE , R., VAN T IL , M. and D URY , S., 2004, The use of hyperspectral data in coastal zone vegetation monitoring. European Association of Remote Sensing Laboratories (EARSeL) eProceedings, 3, pp. 143–153. DEMETRIADES-SHAH, T., STEVEN, M.D. and CLARK, J.A., 1990, High resolution derivative spectra in remote sensing. Remote Sensing of Environment, 33, pp. 55–64. ECKERT, S. and KNEUBÜHLER, M., 2004, Application of Hyperion data to agricultural land classification and vegetation properties estimation in Switzerland. In Proceedings of the Twentieth ISPRS Congress, 12–23 July 2004, Istanbul, Turkey, Vol. 35, Part B7. ENVI, 2008, ENVI Software, Version 4.5 (Boulder, CO: Research Systems Inc.). FELDE, G.W., ANDERSON, G.P., ADLER-GOLDEN, S.M., MATTHEW, N.W. and BERK, A., 2003, Analysis of Hyperion data with the FLAASH atmospheric correction algorithm. In Proceedings of the International Geoscience and Remote Sensing Symposium (IGARSS), 21–25 July 2003, Toulouse, France (New York: IEEE), pp. 90−92. GALVAO, L.S., FORMAGGIO, A.R. and TISOT, D.A., 2005, Discrimination of sugarcane varieties in Southeastern Brazil with EO-1 Hyperion data. Remote Sensing of Environment, 94, pp. 523–534. GALVAO, L.S., ROBERTS, D.A., FORMAGGIO, A.R., NUMATA, I. and BREUNIG, F.M., 2009, View angle effects on the discrimination of soybean varieties and on the relationships between vegetation indices and yield using off-nadir Hyperion data. Remote Sensing of Environment, 113, pp. 846–856. GOETZ, S.J., 2002, Recent advances in remote sensing of biophysical variables: an overview of the special issue. Remote Sensing of Environment, 79, pp. 145–146. GONG, P., PU, R. and YU, B., 1997, Conifer species recognition: an exploratory analysis of in situ hyperspectral data. Remote Sensing of Environment, 62, pp. 189–200. HAMADA, Y., STOW, D.A., COULTER, L.L., JAFOLLA, J.C. and HENDRICKS, L.F., 2007, Detecting Tamarisk species (Tamarix spp.) in riparian habitats of Southern California using high spatial resolution hyperspectral imagery. Remote Sensing of Environment, 109, pp. 237–248. HENDERSON, S., DAWSON, T.P. and WHITTAKER, R.J., 2006, Progress in invasive plants research. Progress in Physical Geography, 30, pp. 25–46. JENSEN, J.R., 2005, Introductory Digital Image Processing: A Remote Sensing Perspective, pp. 1–526 (Upper Saddle River, NJ: Pearson Prentice Hall). JMP Version 8, 2009, SAS Institute, Inc., Cary, NC. JOHNSON, K., 2007, Review of Progress towards the Lantana camara Strategic Plan (Brisbane: Department of Primary Industries and Fisheries, Biosecurity Queensland). JUPP, D.L.B. and DATT, B., 2004, Hyperion Data Processing Workshop: Hands-on Processing Instructions, p. 40 (Canberra: CSIRO Earth Observation Centre). KRUSE, F.A., LEFKOFF, A.B., BOARDMAN, J.W., HEIDEBRECHT, K.B., SHAPIRO, A.T., BARLOON, P.J. and GOETZ, A.F.H., 1993, The Spectral Image Processing System (SIPS) – interactive visualization and analysis of imaging spectrometer data. Remote Sensing of Environment, 44, pp. 145–163. KUMAR, L. and SKIDMORE, A.K., 1998, Use of derivative spectroscopy to identify regions of differences between some Australian eucalypt species. In Proceedings of the Ninth Australasian Remote Sensing and Photogrammetric Conference, 20–24 July 1998, Sydney, Australia (Perth: Remote Sensing and Photogrammetry Association of Australasia), unpaginated CD-ROM. LABA, M., TSAI, F., OGURCAK, D., SMITH, S. and RICHMOND, M.E., 2005, Field determination of optimal dates for the discrimination of invasive wetland plant species using derivative spectral analysis. Photogrammetric Engineering and Remote Sensing, 71, pp. 603–611.
Downloaded by [University of New England] at 14:27 01 March 2012
5436
S. Taylor et al.
LASS, L.W., PRATHER, T.S., GLENN, N.F., WEBER, K.T., MUNDT, J.T. and PETTINGILL, J., 2005, A review of remote sensing of invasive weeds and example of the early detection of spotted knapweed (Centaurea maculosa) and babysbreath (Gypsophila paniculata) with a hyperspectral sensor. Weed Science, 53, pp. 242–251. LASS, L.W., THILL, D.C., SHAFII, B. and PRATHER, T.S., 2002, Detecting spotted knapweed (Centaurea maculosa) with hyperspectral remote sensing technology. Weed Technology, 16, pp. 426–432. LEHMANN, E.L., 2006, Nonparametrics: Statistical Methods Based on Ranks, pp. 1–464 (New York: Springer-Verlag). LUMME, J.H., 2004, Classification of vegetation and soil using imaging spectrometer data. In Proceedings of the Twentieth ISPRS Congress, 12–23 July 2004, Istanbul, Turkey (The Netherlands: ISPRS), Vol. 35, Part B7. MACK, R., SIMBERLOFF, D., LONSDALE, M., EVANS, H., CLOUT, M. and BAZZAZ, F., 2000, Biotic invasions: cause, epidemiology, global consequences, control. Ecological Applications, 10, pp. 689–710. MATTHEW, M.W., ADLER-GOLDEN, S.M., BERK, A., FELDE, G., ANDERSON, G.P., GORODETZKY, D., PASWATERS, S. and SHIPPERT, M., 2003, Atmospheric correction of spectral imagery: evaluation of the FLAASH algorithm with AVIRIS data. In: SPIE Proceedings, Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery IX , 2003, Orlando, FL (Washington, DC: SPIE), pp. 157–163. MICROSOFT OFFICE EXCEL, 2007, Microsoft Corporation, Redmond, WA, USA. NOUJDINA, N.V. and USTIN, S.L., 2008, Mapping Downy Brome (Bromus tectorum) using multivariate AVIRIS data. Weed Science, 56, pp. 173–179. OKIN, G.S., MURRAY, B., ROBERTS, D.A. and OKIN, W.J., 2001, Practical limits on hyperspectral vegetation discrimination in arid and semiarid environments. Remote Sensing of Environment, 77, pp. 212–225. PEARLMAN, J.S., BARRY, P.S., SEGAL, C.C., SHEPANSKI, J., BEISO, D. and CARMAN, S.L., 2003, Hyperion, a space-based imaging spectrometer. IEEE Transactions on Geoscience and Remote Sensing, 41, pp. 1160–1173. PRICE, J.C., 1994, How unique are spectral signatures? Remote Sensing of Environment, 49, pp. 181–186. PSOMAS, A., ZIMMERMAN, N.E., KNEUBÜHLER, M., KELLENBERGER, T. and ITTEN, K., 2005, Seasonal variability in spectral reflectance for discriminating grasslands along a drymesic gradient in Switzerland. In Proceedings of 4th EARSeL Workshop on Imaging Spectroscopy: New Quality in Environmental Studies, 2005, B. Zagajewski, M. Sobczak and M. Wrzesien´ (Eds.), Warsaw, Poland (Poland: EARSeL and Warsaw University), pp. 655–666. SAS INSTITUTE, INC., 1990, SAS/STAT User’s Guide, Version 6, 4th ed., Vol. 1 (Cary, NC: SAS). SCHMIDT, K.S. and SKIDMORE, A.K., 2001, Exploring spectral discrimination of grass species in African rangelands. International Journal of Remote Sensing, 22, pp. 3421–3434. SCHMIDT, K.S. and SKIDMORE, A.K., 2003, Spectral discrimination of vegetation types in a coastal wetland. Remote Sensing of Environment, 85, pp. 92–108. SERPICO, S.B. and BRUZZONE, L., 2001, A new search algorithm for feature selection in hyperspectral remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 39, pp. 1360–1367. SHARMA, G.P., RAGHUBANSHI, A.S. and SINGH, J.S., 2005, Lantana invasion: an overview. Weed Biology and Management, 5, pp. 157–165. SMITH, A.M. and BLACKSHAW, R.E., 2003, Weed–crop discrimination using remote sensing: a detached leaf experiment. Weed Technology, 17, pp. 811–820. SPANGLET, H.J., USTIN, S.L. and REJMANKOVA, E., 1998, Spectral reflectance characteristics of California subalpine marsh plant communities. Wetlands, 18, pp. 307–319.
Downloaded by [University of New England] at 14:27 01 March 2012
Hyperspectral bands for Lantana camara detection
5437
SPLAJT, T., FERRIER, G. and FROSTICK, L.E., 2003, Monitoring of landfill leachate dispersion using reflectance spectroscopy and ground penetrating radar. Environmental Science and Technology, 37, pp. 4293–4298. STORY, M. and CONGALTON, R., 1986, Accuracy assessment: a user’s perspective. Photogrammetric Engineering and Remote Sensing, 52, pp. 397–399. THENKABAIL, P.S., ENCLONA, E.A., ASHTON, M.S. and VAN DER MEER, B., 2004, Accuracy assessments of hyperspectral waveband performance for vegetation analysis applications. Remote Sensing of Environment, 91, pp. 354–376. THENKABAIL, P.S., SMITH, R.B. and DE PAUW, E.D., 2002, Evaluation of narrowband and broadband vegetation indices for determining optimal hyperspectral wavebands for agricultural crop characterization. Photogrammetric Engineering and Remote Sensing, 68, pp. 607–621. UNGAR, S., PEARLMAN, J., MENDENHALL, J. and REUTER, D., 2003, Overview of the Earth Observing One (EO-1) Mission. IEEE Transactions on Geoscience and Remote Sensing, 41, pp. 1149–1159. USGS, 2003, EO-1 User’s Guide. Available online at: http://eol.usgs.gov (accessed 15 April 2010). USGS, 2010, Thematic Mapper (TM) Product Description. Available online at: http://eros.usgs. gov/#/Find_Data/Products_and_Data_Available/TM (accessed 29 July 2009). USTIN, S.L., DIPIETRO, D., OLMSTEAD, K., UNDERWOOD, E. and SCHEER, G.J., 2002, Hyperspectral remote sensing for invasive species detection and mapping. In Proceedings of the IEEE International Geoscience and Remote Sensing Symposium (IGARSS’02), 24–28 June 2002, Toronto, ON (New York: IEEE), pp. 1658–1660. VITOUSEK, P.M., D’ANTONIO, C.M., LOOPE, L.L. and WESTBROOKS, R., 1996, Biological invasions as global environmental change. American Scientist, 84, pp. 218–228.
