Balloon imagery verification of remotely sensed Phragmites australis expansion in an urban estuary of New Jersey, USA

Landscape and Urban Planning 95 (2010) 105–112

Contents lists available at ScienceDirect

Landscape and Urban Planning journal homepage: www.elsevier.com/locate/landurbplan

Balloon imagery verification of remotely sensed Phragmites australis expansion in an urban estuary of New Jersey, USA Francisco Artigas, Ildikó C. Pechmann ∗ New Jersey Meadowlands Commission, Meadowlands Environmental Research Institute, 1 DeKorte Park Plaza, Lyndhurst, NJ 07071, USA

a r t i c l e

i n f o

Article history: Received 7 November 2008 Received in revised form 9 November 2009 Accepted 4 December 2009 Available online 18 January 2010 Keywords: Balloon imagery Remote sensing Vegetation mapping Wetland monitoring Invasive species

a b s t r a c t The invasion of the exotic common reed (Phragmites australis) is increasingly displacing local native species from northeastern coastal estuaries. This study evaluates the accuracy of a remote sensing technique to map the distribution of common reed, monitor the rate of invasion and determine areas of natural resistance to invasion. The current invasion footprint of Phragmites in the Hackensack Meadowlands District in Northern New Jersey was determined using high spectral and spatial resolution hyperspectral imagery. A tethered balloon-based imaging device with limited coverage area was used to assess the accuracy of the hyperspectral imagery classification. The accuracy assessment based on true color balloon images revealed that the hyperspectral classification technique from images covering hundreds of hectares was 90% accurate in separating the dominant common reed-invaded areas from the native vegetation. Furthermore, linear spectral un-mixing techniques for sub-pixel classification revealed that for mixed areas where Phragmites covered 75% or more of a pixel, the classification was correct 96% of the time. The accuracy dropped to 52% for pixels that contained 25% or less of Phragmites cover, and was only 4% for pixels where invasive and native species cover was the same (50–50%). © 2010 Elsevier B.V. All rights reserved.

1. Introduction The invasion of Eastern Shore and Gulf Coast Marshes by the introduced exotic genotype of common reed (Phragmites australis (Cav.) Trin. Ex Stuedel is strongly associated with extensive monocultures and reduced overall biological diversity (Windham and Lathorp, 1999; Able et al., 2003). The ability of this plant to displace native species by altering the local hydrology has created a range of vegetation mixture types along biogeochemical gradients (Bart and Hartman, 2000; Chambers et al., 1998). Dense Phragmites monoculture stands are 3–4 m tall and represent the end point of the invasion. Prior to reaching that stage, Phragmites and native grasses co-exist in mixtures depending on the degree of invasion. The remaining pure native grass relicts usually exist completely surrounded by Phragmites mixtures. The fact that these relicts persist over time may be explained by sediment biogeochemical gradients, as the native species are known to be more tolerant to anoxic and high sulfide conditions than Phragmites (Artigas and Yang, 2005; Chambers et al., 2003). There is an increasing pressure to protect the remaining native species and its associated biodiversity (Ailstock et al., 2001). The

∗ Corresponding author. Present address: Department of Biology, Rutgers State University of New Jersey, 195 University Ave, Newark, NJ 07102, USA. Tel.: +1 201 460 4612/973 353 5385; fax: +1 201 460 2804/973 353 5518. E-mail address: [email protected] (I.C. Pechmann). 0169-2046/$ – see front matter © 2010 Elsevier B.V. All rights reserved. doi:10.1016/j.landurbplan.2009.12.007

Meadowlands District is part of the Atlantic flyway bird migration route that provides good sources of food water and cover and is in direct competition with pressure from developers seeking to develop light industrial warehouses and distribution centers near and with good access to Manhattan and New York City. Hence, there is a need for developing cost effective methodologies for accurately determining the health and extent of the remaining open areas specifically the spatial distribution of Phragmites and its associated mixtures types at the landscape level. High spatial and spectral resolution imaging spectrometers mounted on fixed wing airplanes and measuring light reflectance in many wavelengths have successfully been used to discriminate marsh vegetation types and even vigor classes of the same species as they respond to biogeochemical sediment gradients (Artigas and Yang, 2005). These airborne sensors are able to map hundreds of hectares in just a few hours and the only difficulty occurs when species having similar external architecture and internal tissue morphology reflect light in similar ways and are indistinguishable from one another (Townshend et al., 2000; Okin et al., 2001). Fortunately, studies have shown that the spectral response of marsh grass populations and assemblages have unique spectral characteristics with the exception of species belonging to the same genus (i.e. Spartina alterniflora – saltmarsh cordgrass and S. patens – saltmeadow cordgrass) (Artigas and Yang, 2005). Plant mixtures can be a problem, but there are techniques that help discriminate the importance (%cover) of each species in a pixel.

106

F. Artigas, I.C. Pechmann / Landscape and Urban Planning 95 (2010) 105–112

The linear un-mixing technique first introduced by Adams and Smith (1986) in differentiating dust from bare rock surfaces is a method that can also be used for revealing the spectral characteristic of mixed vegetation pixels by taking into account the pure vegetation spectra that originate from the mixture types (Adams et al., 1995; Bastin, 1997; Wang et al., 2007a). In this study, we show that these sub-pixel techniques are useful in characterizing the mixture types at and near the invasion fronts but in order to extrapolate to larger areas require fine tuning using reliable and scale-appropriate ground reference areas. Schmidt et al. (2004) reviewed different vegetation remote sensing mapping efforts and concluded that the limited number of areas verified on the ground, along with the spectral averaging of non homogenous areas (sub-pixel mixtures) are the main reasons for the low accuracy of these vegetation maps. Belluco et al. (2006) concludes, however, that extensive sets of field observations resulted in higher accuracy vegetation maps derived from remote sensing. The extent of the accuracy that can be obtained from these classifications is uncertain and will mostly depend on the quality and quantity of the ground validation points. Traditional field verification requires reaching hard to access areas over unconsolidated marsh surfaces. Moreover, when the vegetation is 3 m or higher, it is difficult, if not impossible, to detect canopy differences from the ground. A way to approach this problem is to have a perspective from above the canopy.

Digital cameras mounted on a tethered helium-balloon are becoming more frequent in agricultural crop management; they are used to monitor genetic resources and to estimate water consumption (Jia et al., 2004; Jensen et al., 2007; Oberthür et al., 2007; Wang et al., 2007b). Digital cameras flown side by side with hyperspectral sensors opened a new way in addressing remote sensing challenges. In addition, as they are flown at a lower altitude they bridge the spatial gap between radiometric measurements collected near the surface and those acquired by other aircraft or satellites (Vierling et al., 2006; Chen and Vierling, 2006). Our objective was to use digital cameras from a tethered balloon to assess sub-pixel remote sensing classifications of marsh vegetation using a high spectral and spatial resolution spectrometer. We hypothesized that using high resolution balloon imagery for selecting training areas would improve the accuracy of the final vegetation map and provide a better tool for assessing Phragmites invasion. 2. Methods and materials 2.1. Study area The New Jersey Meadowlands district is located in northeastern New Jersey, three miles west from the Upper New York Bay (Fig. 1). Decades of draining, ditching, filling and installation of dikes and tide gates have altered the lower reaches of the Hackensack River

Fig. 1. Map of the New Jersey Meadowlands District showing the locations from where balloon images were captured to produce the training sites used to ground-truth the hyperspectral image: (1) Lyndhurst Marsh; (2 and 3) The Riverbend Wetland Preserve; (4 and 5) Fish Creek. Left upper coordinates: Lat.: N 40◦ 27 13.13 ; Long.: W 73◦ 41 0.82 .

F. Artigas, I.C. Pechmann / Landscape and Urban Planning 95 (2010) 105–112 Table 1 Detailed spectral specification of the 2004 AISA imagery. Band number

Band center (nm)

FWHMa

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34

454.66 474.83 498.11 512.08 526.05 532.26 540.02 550.88 560.19 580.37 589.68 601.29 619.55 639.47 651.09 664.37 671.01 679.31 690.93 701.36 713.33 738.98 744.11 749.24 766.34 778.31 797.12 806.52 811.65 816.78 822.77 845.00 862.10 884.33

4.66 4.66 4.66 4.66 4.66 4.66 4.66 4.66 4.66 4.66 4.66 4.98 4.98 4.98 4.98 4.98 4.98 4.98 4.98 5.13 5.13 5.13 5.13 5.13 5.13 5.13 5.13 3.42 3.42 3.42 5.13 5.13 5.13 5.13

a

Full width at half maximum of an emission line.

into highly developed mixed residential and industrial land uses interspersed among marsh grass fields, tidal creeks and mudflats. The areas of interest are the high and low saltmarsh plant communities. The high marsh community is mainly dominated by Spartina patens, Distichlis spicata (coastal salt grass) and pure patches of Juncus gerardii (black grass) in association with different degrees of invasion by the exotic genotype of P. australis (common reed) (personal communication, 2002). The low marsh is dominated mainly by S. alterniflora mixed with bare mudflats and surrounded by the common reed at higher elevations. 2.2. Image acquisition and processing 2.2.1. Hyperspectral imagery To characterize wetland vegetation, October of 2004 was chosen because while the plant biomass is still high, the flowers are already gone. The Airborne Imaging Spectrometer for Applications (AISA) sensor was flown at 2563 m above ground level (AGL) and configured to capture up-welling radiance in 34 discrete bands in the visible and near infrared (Table 1). A downwelling irradiance sensor on the aircraft was utilized to generate “at-platform” apparent reflectance. This attachment was called FODIS (Fiber Optic Downwelling Irradiance Sensor) that provided radiometric correction data for post-processing of surface reflectance. A surface of approximately 83 km2 was captured in 26 flight lines at a spatial resolution of 2.5 m. The sun angle was 37.2◦ above the horizon while the sun azimuth was 148.8◦ from map north. The image processing was carried out using ENVI 4.2 (Visual Information Solutions, 2004). Image atmospheric correction was accomplished using the Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) set for the Mid-Latitude summer atmospheric model with 2.92 g/cm2 average water vapor and 21 ◦ C

107

average surface air temperature. Water retrieval was performed for the 820 nm water feature. There was no aerosol retrieval and spectral polishing was not performed. Pre-processing was completed by rectifying the atmospherically corrected flight lines using a set of orthophotos of the State of New Jersey taken February through April, 2002 at a scale of 1:2400 (1 = 200 ) with a 0.33-m pixel resolution. ENVI’s map module was used for the orthorectification, which did not allow for higher RMS (root mean square) error than 2.5 m (same as the resolution of the image). The Normalized Difference Vegetation Index (Tucker, 1979) derived from band 12 (600–605 nm) and band 28 (805–810 nm) was calculated for the entire dataset. By selecting an empirical NDVI threshold value of 0.3, the impervious surfaces, pure water and mud pixels were masked out from the imagery leaving only vegetated surfaces for classification. 2.2.2. Aerial imagery During the summer of 2006, high resolution aerial images of the marsh community were captured with a seven mega pixel digital camera (Sony DSC-V3) mounted on a tethered helium-balloon platform suspended 150 m in the air. The camera was equipped with a 0.7 wide-angle lens and set to a shutter speed of 600, ISO of 400 and programmed to shoot every 15 s. The camera was moved either by boat or on foot. One location was recorded on the same day at the same time. Each photo covered an area of approximately 120 m × 120 m. The resulting digital images were mosaiced using Adobe Photoshop® (v.7.0) and saved in a tag image file format (TIFF) for further processing. The mosaics were geo-referenced based on the same set of orthophotos used for rectifying the AISA image. A nearest neighbor resampling method and warping procedure gave a maximum root mean square error (RMS) of 0.33 m which resulted in a 0.12-m spatial resolution. Geo-referenced mosaics from balloon imagery were produced for five distinct locations showing different degrees of high marsh invasion by the common reed (Fig. 2). 2.3. Classification and linear un-mixing The initial AISA imagery captured 34 bands in the visible and near infrared (NIR) spectra of light (452–886 nm). Some of the bands contain a great deal of noise and barely contribute information to separating the targeted plant assemblages, so the Minimum Noise Fraction (MNF) rotation (Hirano et al., 2003) was used to isolate bands with the highest signal to noise ratio. Although only the first 15 bands were found useful, the low-order MNF components contain more noise, thus they still have a signal constituent. Statistics obtained from the rotation were used to back-transform the images, i.e. smooth the spectral information using ENVI’s Forward MNF Rotation function (Green et al., 1988). In hyperspectral remote sensing, vegetation classifications can be created by matching image spectra with known reference spectra (Kruse et al., 1993). A commonly used algorithm to determine the similarity of spectra in multidimensional space (multi-band images) is the Spectral Angle Mapper (SAM) method. This method determines the similarity between two spectra by calculating the angle between them when treated as vectors in a space with dimensionality equal to the number of bands (Yuhas and Goetz, 1993). In our case, reference spectra were obtained by on-screen digitizing distinct plant assemblage areas from the aerial balloon imagery and later importing these in ENVI as regions of interest. A reference spectrum was calculated for each class by averaging all the pixel values falling within the boundaries of the area of interest (also known as training areas). Finally, seven reference spectra were saved as a spectral library and used in the classification procedure. These seven spectra were: (1) back marsh, a stunted form of S. alterniflora; (2) high marsh, S. patens and D. spicata at higher elevation; (3) mixtures, mixture of common reed and high marsh; (4 and 5) common reed, less and

108

F. Artigas, I.C. Pechmann / Landscape and Urban Planning 95 (2010) 105–112

Fig. 2. Example of balloon imagery (Fish Creek) showing the boundaries of selected training sites: (1) Vigorous Phragmites australis; (2) pure high marsh vegetation; (3 and 4) mixtures (75% Phragmites cover and 75% high marsh cover, respectively). Left upper coordinates: Lat.: N 40◦ 30 1.65 ; Long.: W 73◦ 45 15.05 .

more vigorous forms of P. australis (small and big flowers). Distinguishing vigorous and stunted forms of reed was based on field experience. The stunted form of common reed likely growing on more reductive sediment than the vigorous form, presents a shorter phenotype with smaller internodal length, more dense but shorter leaves and shorter inflorescence (unpublished data); (6) low marsh, a tall vigorous form of S. alterniflora associated with mud at lower elevations; (7) black grass, small and pure stands of J. gerardii (Fig. 3

displays the spectra of major vegetation types). The result of the classification procedure was a classified image indicating pixels spectrally similar to the corresponding set of seven end-member spectra. We treated the mixture class (Class 3), showing different degrees of Phragmites invasion, as a single class regardless of the cover ratio between the high marsh grasses and Phragmites. To determine the amount of the invasive common reed in each pixel, the areas identified as mixtures were masked out of the original

Fig. 3. Spectra of major vegetation types retrieved from the 2004 AISA imagery based on corresponding pixels of balloon imagery. Big Phrag: P. australis (common reed) with big inflorescence (vigorous form); Small Phrag: the stunted form with a small inflorescence. Low marsh: tall form of Spartina alterniflora (saltmarsh cordgrass); back marsh: stunted form. High marsh: community of Spartina patens and Distichlis spicata; mixture: combinations of common reed and high marsh species.

F. Artigas, I.C. Pechmann / Landscape and Urban Planning 95 (2010) 105–112

image and further processed separately using the linear un-mixing procedure described below. The common reed class was also further separated into more and less vigorous types depending on the size of the inflorescence. Spectral mixing occurs when materials having different spectral properties are present in a single image pixel. Singer (1981) found that in cases where the mixing is macroscopic, i.e. mixture types are visible to the naked eye, the mixing occurs in a linear fashion. Hence, linear un-mixing techniques assume that each end-member forming the initial data matrix (hyperspectral image), when linearly combined among them, is able to match all possible spectra in the image (Kruse, 1994). The Linear Spectral Un-mixing Model (LSM) (Settle and Drake, 1993) was used to determine the fractional cover of common reed within pixels classified as mixed vegetation. The linear spectral un-mixing was applied using the ENVITM 4.2 linear spectral un-mixing algorithm embedded in the software. The reflectance at each pixel of the image was assumed to be a linear combination of the reflectance of each material (or end-member) present within the pixel. For example, if 25% of a pixel contains the curve of stunted Phragmites and 75% of the pixel contains spectral curve of high marsh, the spectrum for that pixel is a weighted average of 0.25 times the spectrum of stunted Phragmites plus 0.75 times the spectrum of high marsh. So given the resulting spectrum (the input data) and the end-member spectra, the linear un-mixing is solving for the abundance values of each end-member for every pixel (Canty, 2006). The method presented in this paper applies an unconstrained technique, in which physically impossible negative covers are accepted and no constraint is imposed to force the sum of cover fractions to sum to unity (Elmore et al., 2000). The output of the LSM procedure is a grey scale abundance image where the pixel values indicate the importance of each end-member. In other words, the brighter the pixel, the higher the abundance of a certain end-member in that pixel.

109

Table 2 Ground control areas used in the accuracy assessment procedure. Big Phragmites: P. australis (common reed) with big inflorescence (vigorous form) and Small Phragmites (the stunted form) with a small inflorescence. Low marsh: tall form of S. alterniflora (saltmarsh cordgrass); back marsh: stunted form. High marsh: community of Spartina patens and Distichlis spicata; mixture: combinations of common reed and high marsh species; black grass: pure Juncus gerardii stands. Cover type

Number of ground reference areas

Number of pixels

Area (m2 )

Big Phragmites Small Phragmites Back marsh Low marsh High marsh Mixture Black grass

24 11 5 11 13 9 2

272 161 81 52 181 97 20

1700 1006 506 325 1131 606 125

Total

75

864

4899

7 6 5

35 10 35

219 62.5 219

18

80

508.5

Mixtures >75% Phragmites 50–50% high marsh and Phragmites <25% Phragmites Total

and up to 1700 m2 for the more common Phragmites cover types. Table 2 shows the type and size of the control areas used to assess the hyperspectral image classification results. The overall classification results are shown in Fig. 4. The accuracy assessment using balloon ground control areas resulted in a 90% overall accuracy with a kappa coefficient of 0.87. An earlier study conducted on the same area that used areas identified on the field by GPS as training areas for classification achieved 65% overall accuracy with kappa coefficient equals to 0.65 (Artigas and Yang, 2005). The relatively low overall accuracy was due to the similarity

2.4. Accuracy assessment of classification results The classification accuracy of the resulting vegetation maps was tested against reference areas representing the above six classes as well as areas showing different levels of Phragmites invasion and length of inflorescence captured in balloon images. The reference areas were randomly selected pixels, which had been earlier identified using balloon imagery and then imported into the accuracy assessment procedure as reference points. A confusion matrix (Jensen, 1986) that compares the location and class of each ground reference pixel with the corresponding location and class in the classification image was used to show the accuracy of the vegetation maps. The procedure reports the overall accuracy, the producer’s and user accuracies and kappa coefficient. The overall accuracy is calculated by summing the number of pixels classified correctly and dividing by the total number of pixels. The producer’s accuracy indicates the probability of a reference point being correctly classified (columns of the confusion matrix). Users’ accuracy on the other hand is a measure of the reliability of an output map generated from a classification scheme (rows of the confusion matrix). The kappa coefficient (k) is yet another way to measure classification accuracy. It describes the proportion of correctly classified validation sites after random agreements were removed (Rosenfield and Fitzpatrick-Lins, 1986). 3. Results 3.1. Results of image classification The control areas extracted from balloon images and representing homogenous cover for the classes and mixture types ranged in size from 125 m2 for the rarer communities (e.g. black grass),

Fig. 4. Spectral Angle Mapper classification results of the Riverbend Wetland Preserve (28 ha) showing the spatial distribution of the dominant plant cover types which, according to the accuracy assessment procedure, are 90% accurate. Left upper coordinates: Lat.: N 40◦ 45 1.65 ; Long.: W 74◦ 5 40.48 .

110

F. Artigas, I.C. Pechmann / Landscape and Urban Planning 95 (2010) 105–112

Table 3 Summary of the classification accuracy assessment showing the Producer’s accuracy as the percentage of a particular ground class that was correctly classified. The User’s accuracy shows what percentage of a class corresponds to the ground-truthed class. The overall accuracy shows the extent to which the classification procedure correctly classified all classes. Big Phragmites: P. australis (common reed) with big inflorescence (vigorous form) and Small Phragmites (the stunted form) with a small inflorescence. Low marsh: tall form of S. alterniflora (saltmarsh cordgrass); back marsh: stunted form; high marsh: community of S. patens and D. spicata; mixture: combinations of common reed and high marsh species. Black grass: pure J. gerardii stands. Phragmites

Back marsh

Low marsh

High marsh

Mixture

Total

User’s accuracy

Big Phragmites, Small Phragmites Back Marsh Low Marsh High Marsh Mixture Black Grass

401 3 0 21 7 0

5 72 4 0 0 0

4 0 48 0 0 0

4 0 0 160 17 0

4 0 0 9 84 0

0 0 0 0 0 20

418 75 52 190 108 21

96 96 92 84 78 95

Total

433

81

52

181

97

20

864

93

89

92

88

87

100

Producer’s accuracy Kappa coefficient: 0.87

Black grass

Overall accuracy: 90

of the low marsh and high marsh spectra which are dominated by species of the same genus (S. alterniflora and S. patens). Results of the current study show that the classification method was more accurate at separating the relatively pure vegetation types compared to mixed areas. User’s accuracy, i.e. what percent of the class corresponds to the verified balloon ground reference class, was between 90% and 96% for the relatively pure classes and 78% for the mixture classes (Table 3). The user’s accuracy for Phragmites was 96%. The chances of the two Phragmites types (yellow and orange areas in Fig. 4) being dominated by almost pure Phragmites is 96%. Similarly, user’s accuracy for stunted S. alterniflora and J. gerardii was greater than 90%. A slightly lower accuracy was obtained for high marsh communities dominated by S. patens and D. spicata (84%). The confusion matrix indicates that almost 9% of the high marsh area contained a mixture of high marsh and Phragmites. In other words, almost 9% of the areas classified as strictly high marsh most likely contain common reed. Results also show that when the Phragmites class was further separated into two sub-classes depicting the vigorous (with longer inflorescence) and stunted form (with shorter inflorescence), the method was more reliable at describing the more vigorous stands (86%) compared to the stunted form (68%). 3.2. Results of linear spectral un-mixing The LSM was performed only on those 12% of pixels that were found to be in the “mixed” category in the first step of the classification. The output of the LSM procedure was a raster map of proportions where the value of the pixel indicates the fractional cover of Phragmites in the mixed pixel. Fig. 5 illustrates that the patches that contained mixtures and the most advanced stages of invasion (brighter pixels) were always associated with slightly higher elevations. Table 4 shows the accuracy assessment for distinct patches of mixed vegetation with various degrees of invasion by the common reed. The classification method originally showed percent cover of the Phragmites to the other classes, however for validation purposes we needed to truncate the results into three classes, i.e. 25%, 50% and 75% reed cover. The final map distinguishes mixed pixels with greater than 75% Phragmites cover with 100% accuracy. The accuracy was only 4% when Phragmites cover was around 50% and increased to 52% when Phragmites cover was less than 25% of the mixed pixel. 4. Discussion The role of remote sensing in mapping coastal and tidal wetland ecosystems has become crucial to assist in the effort to preserve and sustain these valuable communities. Vegetation mapping of tidal

Fig. 5. The results of the Linear Spectral Un-mixing (LSM) procedure performed only on pixels that were identified as mixtures of reed and high marsh. The brighter pixels represent higher abundance of P. australis. Left upper coordinates: Lat.: N 40◦ 45 1.65 ; Long.: W 74◦ 5 40.48 .

wetlands primarily driven by the changing water level and salinity on the other hand are known to be challenging (Silvestri et al., 2003). This is especially true when it comes to mapping and characterizing the distribution of the invasive genotype of P. australis and the associated plant diversity. The success of these classification methods relies on how well the sensors capture and discriminate the variations in canopy texture along the natural gradients (Cracknell, 1999) and how well the classifications hold beyond the areas of the control points. Bajjouk et al. (1999) emphasize in their coastal study, that the main obstacle is obtaining accurate locations for the training data in the image, since as spatial variation in the tidal area is high, a shift of one pixel may induce a significant error. This study used a linear shifting model to ensure that field location of training areas matched with the image. A reason remote sensing data are not widely adopted for management purposes at fine scales is because it is not readily available yet and because of the uncertainty of the classifications and their reliability when extrapolating to larger areas. The traditional methods for collection of ground reference data are labor intensive and the amount of validation is limited in terms of the area cover. Using low cost balloon imagery to verify and assess remote sensing classifications allows for the

F. Artigas, I.C. Pechmann / Landscape and Urban Planning 95 (2010) 105–112

111

Table 4 Summary of the classification accuracy assessment of the un-mixing procedure showing the percentage of a particular ground class that was correctly classified (producer’s accuracy). The user’s accuracy shows what percentage of a class corresponds to the ground-truthed class and the overall accuracy shows the extent to which the classification procedure correctly classified all classes. AISA class

>25% Phragmites

Total

User’s accuracy

<75% Phrag 50–50% mixture >25% Phrag

11 11 13

>75% Phragmites

50–50% mixture 0 1 9

0 11 24

11 23 46

100% 4% 52%

Total

35

10

35

80

Producer’s accuracy Kappa coefficient: 0.46

31%

10%

67%

inclusion of smaller scale variations in texture that are then captured as individual spectral curves and result in more reliable and precise classifications. A system that allows for verification of large areas (hectares) and can produce vegetation classifications that can safely be extrapolated to the entire estuary in a cost effective way are in demand. We show that aerial balloon images are significant for measuring the accuracy of wetland vegetation maps created from airborne hyperspectral imagery, although vegetation interpretation by onscreen digitizing might introduce observer bias. This approach proved to be superior to the one used by Artigas and Yang (2005) that involved creating training sites using GPS coordinates from the field along with vegetation cover descriptions notes. Our approach demonstrated that pure Phragmites stands in the fall season are easily separated from other vegetation types with a high level of accuracy (∼90%). Similarly, distinct homogenous communities such as high, low and back marsh are also easily separated from each other with 90% overall accuracy. The most vulnerable community to common reed invasion is the high marsh community which is truly a mixture of two distinct species (S. patens and D. spicata). The spectral inhomogeneity of the high marsh habitat is apparent by the lower accuracy of the classification (84%) compared to the other more truly pure types (>90% accurate). Our results also confirm that during the fall season, vegetation associations within the Meadowlands Estuary are spectrally distinct from one another (Artigas and Yang, 2005). Wang et al. (2007a) achieved similar overall accuracy (∼91%) and kappa coefficient (0.87) using a Community based Neural Network Classifier, however they also acknowledged that detailed and careful field work is needed to achieve such high accuracy. To our best knowledge however, LSM techniques have not been used for determining relative abundance of vegetation in mixed pixels at coastal wetland areas. Silvestri et al. (2003) used LSM to link vegetation distribution to soil topography in saltmarsh vegetation communities and Shanmugam et al. (2006) found LSM more reliable over traditional classification methods, however they did not consider linking the relative abundance of vegetation communities to the result of un-mixing. We demonstrated a procedure to determine the proportion of common reed in 2.5 m × 2.5 m mixed pixels. Using spectra associated with mixed vegetation patches we were able to further separate mixed areas with 90% accuracy Those areas were later entered in the linear spectral un-mixing procedure and as a result we obtained 100% accuracy when the mixed pixels contained 75% or greater cover of the common reed. When the high marsh species in a pixel were dominant, the classification model slightly overestimated the reed cover with an accuracy of 52%. However, the method was unable to clearly distinguish with any level of accuracy the 50–50 mixture class (accuracy 4%). Andrew and Ustin (2008) discusses that un-mixing techniques are competent only when spectra of pure vegetation types are entered in the un-mixing procedure. In this study since computing pixel purity index (Boardman et al., 1995) did not bring the desired accuracy,

Overall accuracy: 45%

spectra entered in the LSM procedure were averages of several pixels covered by the same vegetation community. Entering averaged spectra into the LSM procedure might be the reason for the low accuracy where the abundance of the two spectra was equal. Overall, the LSM method was most accurate at determining areas with late and early stages of common reed invasion and less accurate at determining the intermediate stages. We were able to assess the results of our classification and are confident that the SAM classifications results for the relatively pure stands and mixtures (regardless of level of mixing) are accurate and can be safely extrapolated to greater areas within the estuary. Based on these results, we may use this method to track vegetation changes due to human impacts at the site level (e.g. ditching, filling etc.), and changes due to local shifts in weather and sea level rise. The mixture types remain elusive when percent cover of species in the mixture is similar (50:50). We believe that this was the result of our training sites failing to capture and summarize a greater variety of end-members needed to explain the mixtures in the 50:50 range. In terms of identifying invasion trends and invasion fronts, knowing areas with early and late stages of invasion may be sufficient enough; investing time and resources to find more end-members to explain the 50:50 mixtures may not justify the effort. The main reason remote sensing data are not adopted for management purposes at fine scales is because of the uncertainty of the classification performed on multi- or hyperspectral imagery. The collection of ground reference data is labor intensive, the amount of validation is usually limited in terms of the area cover. Using low cost systems such as balloon imagery to assess remote sensing classifications allows for the inclusion of smaller scale variations in texture that are then captured as individual spectral curves and result in classifications that are more reliable. A system that allows for verification of large areas (hectares) and can support hyperspectral image classifications that are more reliable by extrapolating spectral models from verified areas into the larger extension of an entire estuary system is essential for further land use and management decisions. Cost efficiency is another reason why balloon imagery is more favorable over traditional field validation. Meanwhile it would cost in the order of $15,000 to perform the typical wetland monitoring for a 30-acre area site, the estimate for doing the same work using aerial balloon imagery is in the order of $4000. This includes the cost of the camera and balloon (1500), two days of field verification (photos on the ground) $1000 and three days of image processing and classification work $1500. Finally, the question may arise, if balloon imagery can provide such detailed information on species distribution, is it still necessary to input the extra effort and expense of collecting hyperspectral imagery? Although collection of balloon imagery is considerably cheaper than contracting a hyperspectral flight, the latter can capture up to 5000 ha a day, if the weather conditions are favorable. Furthermore, image acquisition with a balloon-mounted camera is highly sensitive to the varying wind conditions. Under changing wind conditions, it is troublesome to acquire reliable,

112

F. Artigas, I.C. Pechmann / Landscape and Urban Planning 95 (2010) 105–112

planimetrically stable imagery. Hence the cost/time ratio still favors hyperspectral imagery. Acknowledgements The technical assistance of Dr. Jian-sheng Yang is gratefully acknowledged. The authors wish to thank Robert Saverino for his editorial work. This project was funded by the New Jersey Department of Environmental Protection. References Able, K.W., Hagan, S.M., Brown, S.A., 2003. Mechanisms of marsh habitat alteration due to Phragmites: response of young-of-the-year Mummichog (Fundulus heteroclitus) to treatment for Phragmites removal. Estuaries Coasts 26 (2), 484–494. Adams, J.B., Smith, M.O., 1986. Spectral mixture modeling: a new analysis of rock and soil types at the Viking lander 1 Site. J. Geophys. Res. 91, 8098–8112. Adams, J.B., Sabol, D.E., Kapos, V., Filho, R.A., Roberts, D.A., Smith, M.O., 1995. Classification of multispectral images based on fractions of endmembers: application to land-cover change in the Brazilian Amazon. Remote Sens. Environ. 52, 137–154. Ailstock, M.S., Norman, C.M., Bushman, P.J., 2001. Common reed Phragmites australis: control and effects upon biodiversity in freshwater nontidal wetlands. Restor. Ecol. 9 (1), 49–59. Andrew, M.E., Ustin, S.L., 2008. The role of environmental context in mapping invasive plants with hyperspectral image data. Remote Sens. Environ. 112, 4301–4317. Artigas, F.J., Yang, J.S., 2005. Hyperspectral remote sensing of marsh species and plant vigour gradient in the New Jersey Meadowlands. Int. J. Remote Sens. 26, 5209–5220. Bajjouk, T., Populus, J., Guillaumont, B., 1999. Quantification of subpixel cover fractions using principal component analysis and a linear programming method: application to the coastal zone of Roscoff (France). Remote Sens. Environ. 64, 53–165. Bart, D., Hartman, J.M., 2000. Environmental determinants of Phragmites australis expansion in a New Jersey salt marsh: an experimental approach. Oikos 89, 59–69. Bastin, L., 1997. Comparison of fuzzy c-means classification, linear mixture modeling and MLC probabilities as tools for unmixing coarse pixels. Int. J. Remote Sens. 18, 3629–3648. Belluco, E., Camuffoo, M., Ferrari, S., Modenese, L., Silvestri, S., Marani, A., 2006. Mapping salt-marsh vegetation by multispectral and hyperspectral remote sensing. Remote Sens. Environ. 105, 54–67. Boardman, J.W., Kruse, F.A., Green, R.O., 1995. Mapping target signatures via partial unmixing of AVIRIS data. In: Summaries, Fifth JPL Airborne Earth Science Workshop, vol. 1, pp. 23–26 (JPL Publication 95-1). Canty, M.J., 2006. Image Analysis, Classification and Change Detection in Remote Sensing, with Algorithms for ENVI/IDL. Taylor & Francis/CRC Press, 348 pp. Chambers, R.M., Mozdzer, T.J., Ambrose, J.C., 1998. Effects of salinity and sulfide on the distribution of Phragmites australis and Spartina alterniflora in a tidal marsh. Aquat. Bot. 62, 161–169. Chambers, R.M., Osgood, D.T., Bart, D.J., Montalto, F., 2003. Phragmites australis invasion and expansion in tidal wetlands: interactions among salinity. Sulfide and hydrology. Estuaries Coasts 26 (2B), 398–406. Chen, X., Vierling, L.A., 2006. Spectral mixture analyses of hyperspectral data acquired using a tethered balloon. Remote Sens. Environ. 103, 338–350. Cracknell, A.P., 1999. Remote sensing techniques in estuaries and coastal zones—an update. Int. J. Remote Sens. 19., 485–496. Elmore, A.J., Mustard, J.F., Manning, S.J., Lobell, D.B., 2000. Quantifying vegetation change in semi-arid environments. Precision and accuracy of spectral mixture analysis and the normalized difference vegetation index. Remote Sens. Environ. 73, 87–102.

Green, A.A., Berman, M., Switzer, P., Craig, M.D., 1988. A transformation for ordering multispectral data in terms of image quality with implications for noise removal. IEEE Trans. Geosci. Remote Sens. 26, 65–74. Hirano, A., Madden, M., Welch, R., 2003. Hyperspectral image data for mapping vegetation. Wetlands 23, 436–448. Jensen, J.R., 1986. Introductory Digital Image Processing. Prentice-Hall, Englewood Cliffs, NJ, p. 379. Jensen, T., Apan, A., Young, F., Zeller, L., 2007. Detecting the attributes of a wheat crop using digital imagery acquired from a low-altitude platform. Comput. Electron. Agric. 59, 66–77. Jia, L., Buerkertb, A., Chena, X., Roemheldc, V., Zhanga, F., 2004. Low-altitude aerial photography for optimum N fertilization of winter wheat on the North China Plain. Field Crop Res. 89, 389–395. Kruse, F.A., Lefkoff, A.B., Boardman, J.B., Heidebrecht, K.B., Shapiro, A.T., Barloon, P.J., et al., 1993. The spectral image processing system (SIPS)—interactive visualization analysis of imaging spectrometer data. Remote Sens. Environ. 44, 145–163. Kruse, F.A., 1994. Imaging spectrometer data analysis—a tutorial. In: Proceedings of the International Symposium on Spectral Sensing Research (ISSSR), San Diego, CA, 10–15 June 1994, pp. 44–54. ˜ Oberthür, T., Cocka, J., Andersson, M.S., Naranjo, R.N., Castaneda, D., Blair, M., 2007. Acquisition of low altitude digital imagery for local monitoring and management of genetic resources. Comput. Electron. Agric. 58, 60–77. Okin, G.S., Roberts, D.A., Murray, B., Okin, W.J., 2001. Practical limits on hyperspectral vegetation discrimination in arid and semiarid environments. Remote Sens. Environ. 77, 212–225. Rosenfield, G.H., Fitzpatrick-Lins, K., 1986. A coefficient of agreement as a measure of thematic classification accuracy. Photogramm. Eng. Remote Sens. 52, 223–227. Schmidt, K.S., Sikamore, A.K., Kloosterman, E.H., van Oasten, H.H., Kumar, L., Janssen, J.A.M., 2004. Mapping coastal vegetation using an expert system and hyperspectral imagery. Photogramm. Eng. Remote Sens. 70, 703–715. Settle, J.J., Drake, N.A., 1993. Linear mixing and the estimation of ground cover proportions. Int. J. Remote Sens. 14, 1159–1177. Shanmugam, P., Ahn, Y.H., Sanjeevi, S., 2006. A comparison of the classification of wetland characteristics by linear spectral mixture modelling and traditional hard classifiers on multispectral remotely sensed imagery in southern India. Ecol. Model. 194, 379–394. Singer, R.B., 1981. Near-infrared spectral reflectance of mineral mixtures: systematic combinations of pyroxenes, olivine and iron oxides. J. Geophys. Res. 86, 7967–7982. Silvestri, S., Marani, M., Marani, A., 2003. Hyperspectral remote sensing of salt marsh vegetation, morphology and soil topography. Phys. Chem. Earth 28, 15–25. Townshend, J.R.G., Huang, C., Kalluri, S.N.V., Defries, R.S., Liang, S., Yang, K., 2000. Beware of per-pixel characterization of land cover. Int. J. Remote Sens. 21, 839–843. Tucker, C.J., 1979. Red and photographic infrared linear combination for monitoring vegetation. Remote Sens. Environ. 8, 127–150. Vierling, L.A., Fersdahl, M., Chen, X., Li, Z., Zimmerman, P.R., 2006. The Short Wave Aerostat-Mounted Imager (SWAMI): a novel platform for acquiring remotely sensed data from a tethered balloon. Remote Sens. Environ. 103, 255–264. Wang, C., Menenti, M., Stoll, M.-P., Belluco, E., Marani, M., 2007a. Mapping mixed vegetation communities in salt marshes using airborne spectral data. Remote Sens. Environ. 107, 559–570. Wang, J., Sammis, T.W., Andales, A.A., Simmons, L.J., Gutschick, V.P., Miller, D.R., 2007b. Crop coefficients of open-canopy pecan orchards. Agric. Water Manage. 88, 253–262. Windham, L., Lathorp, R.G., 1999. Effects of Phragmites australis (common reed) invasion on aboveground biomass and soil properties in brackish tidal marsh of the Mullica River, New Jersey. Estuaries 22, 927–935. Yuhas, R.H., Goetz, A.F.H., 1993. Comparison of airborne (AVIRIS) and spaceborne (TM) imagery data for discriminating among semi-arid landscapes endmembers. In: Proceedings, Ninth Thematic Conference on Geologic Remote Sensing, Environmental Research Institute of Michigan, Ann Arbor, MI, USA, pp. 503–511.