Remote Sensing of Environment 112 (2008) 4034–4047

Contents lists available at ScienceDirect

Remote Sensing of Environment

j ourna l homepage: www.e lsev ie r.com/ locate / rse

Identification of invasive vegetation using hyperspectral remote sensing in theCalifornia Delta ecosystem

Erin L. Hestir a,⁎, Shruti Khanna a, Margaret E. Andrew a, Maria J. Santos a, Joshua H. Viers b,Jonathan A. Greenberg a, Sepalika S. Rajapakse a, Susan L. Ustin a

a Department of Land, Air, and Water Resources, University of California, Davis, One, Shields Ave., Davis, CA 95616, United Statesb Department of Environmental Science and Policy, University of California, Davis, One Shields Ave., Davis, CA 95616, United States

⁎ Corresponding author. Tel.: +1 530 752 5092; fax: +E-mail address: elhestir@ucdavis.edu (E.L. Hestir).

0034-4257/$ – see front matter © 2008 Elsevier Inc. Aldoi:10.1016/j.rse.2008.01.022

A B S T R A C T

A R T I C L E I N F O

Article history:

Estuaries are among themost Received 6 April 2007Received in revised form 17 January 2008Accepted 26 January 2008

Keywords:Aquatic and wetland weedsLogistic regression modelSpectral mixture analysisSpectral Angle MapperHyperspectralHyMapPhenology

invadedecosystemson theplanet. Such invasionshave led inpart, to the formationof amassive $1 billion restoration effort in California's Sacramento–San Joaquin RiverDelta. However, invasions ofweedsinto riparian, floodplain, and aquatic habitats threaten the success of restoration efforts within the watershed andjeopardize economic activities. The doctrine of early detection and rapid response to invasions has been adopted byland and water resource managers, and remote sensing is the logical tool of choice for identification and detection.However meteorological, physical, and biological heterogeneity in this large system present unique challenges tosuccessfully detecting invasive weeds. We present three hyperspectral case studies which illustrate the challenges,and potential solutions, tomapping invasiveweeds inwetland systems: 1) Perennial pepperweedwasmapped overone portion of the Delta using a logistic regression model to predict weed occurrence. 2) Water hyacinth and3) submerged aquatic vegetation (SAV), primarily composed of Brazilian waterweed, were mapped over the entireDelta using a binary decision tree that incorporated spectral mixture analysis (SMA), spectral anglemapping (SAM),band indexes, and continuum removal products. Perennial pepperweed detection was moderately successful;phenological stage influenced detection rates. Water hyacinth was mapped with modest accuracies, and SAV wasmappedwithhighaccuracies. Perennial pepperweedandwaterhyacinthbothexhibited significant spectral variationrelated toplantphenology. Suchvariationmustbeaccounted for inorder tooptimallymapthese species, and thiswasdone for the water hyacinth case study. Submerged aquatic vegetation was not mapped to the species level due tocomplexnon-linearmixingproblemsbetween thewatercolumnand its constituents,whichwasbeyond thescopeofthe current study. We discuss our study in the context of providing guidelines for future remote sensing studies ofaquatic systems.

© 2008 Elsevier Inc. All rights reserved.

1. Introduction

Invasions of aquatic weeds into freshwater, estuarine, and floodplainhabitats can decrease biodiversity, threaten critical habitat, alter nut-rient cycles, and degrade water quality. An estimated USD $100 millionper year is spent on control and eradication programs targeting aquaticweeds in the United States (Pimentel et al., 2000). Systematic, com-prehensivemonitoringprograms are needed todetect invasions inorderto effectively control aquatic weeds. Traditional methods of monitoringweed infestations are costly, time consuming, and often require directcontact with the weeds which can result in further weed dispersal(Bossard et al., 2000). Additionally, aquatic ecosystems are often inac-cessible or logistically difficult for field-based monitoring methods.Remote sensing provides a synoptic solution for monitoring aquatic

1 530 754 5491.

l rights reserved.

weed infestations over large spatial areas (Ackleson and Klemas, 1987).To be successful, a remote sensing approach must be accurate, repeat-able over space and time, and account for the inherent spatial andenvironmental heterogeneity of a system.

Wepresent three case studies that illustrate how these problems canbe addressed to develop regional-scale monitoring of invasive aquaticandwetlandweeds in the Sacramento–San Joaquin Delta: the terrestrialriparianweed, perennial pepperweed (Lepidium latifolium); the floatingaquatic weed, water hyacinth (Eichhornia crassipes); and the submergedaquatic weed, Brazilianwaterweed (Egeria densa). Each case study high-lights the complexities of remote sensing of aquatic and wetland sys-tems. They also demonstrate a range of techniques that can be used, andoften integrated, to produce accurate maps for wetland and estuarineresourcemanagers. Ourexamples are froma study tomap these invasivespecies at high spatial resolution (3m) over a regional area of 2139 km2.The challenge of analyzing airbornehyperspectral remote sensingover alarge study region notwithstanding, we demonstrate that the techni-ques presented can be applied consistently to all flightlines in the study

4035E.L. Hestir et al. / Remote Sensing of Environment 112 (2008) 4034–4047

regionovermultipleyears, thus creatinganeffective and comprehensivemonitoring program.

In estuarine systems, meteorological, physical, and biological hetero-geneity present serious challenges that must be resolved to successfullymap and monitor distributions of aquatic vegetation species. Meteor-ological heterogeneity is one challenge to invasive species monitoring.Factors such as weather conditions, sun, and view angle determine thebidirectional reflectance distribution function (BRDF),which complicatesremote sensing of aquatic vegetation, especially submerged aquaticvegetation (SAV) such as Brazilianwaterweed. Sunglint is light specularlyreflected off the water surface which effectively impedes retrieval of auseable signal (Bostater et al., 2004; Mertes et al., 1993; Morel andBelanger, 2006). Variable sun angles and wind speeds contribute diffe-ring flightline effects that confound region-wide analyses. The spectraldetectionof SAV is also affectedbothby the apparent optical properties ofwater, such as surface reflectance and vertical diffuse attenuation, andinherent optical properties that do not depend on the ambient radiancedistribution in the water column (Mobley, 1994). Water depths changewith tidal stage and runoff; suspended anddissolvedmaterials vary overgeomorphological gradients, meteorological conditions, flow conditions,and land use practices. All of these effects influence remote sensing ofaquatic systems by limiting the detection of SAV as light attenuates withdepth, altering the water-leaving reflectance.

Another challenge to monitoring this system is its biological hetero-geneity. Plant phenology varies across the large gradients present in thissystem.Waterhyacinth andperennial pepperweed lifehistoriesmustbeaccounted for to capture key phenological attributes such as floweringand senescence. Different phenologic states, combined with differencesin leaf and canopy structure canbepresentover short distances, creatingintra-species variation that can result in overlapping spectral featuresbetween co-occurring species. Additionally, weed species may be pre-sent in subpixel mixtures even at high spatial scales. Hyperspectralimaging may provide sufficient information to overcome these chal-lenges, allowing the application of more complex spectral analyses andspectral unmixing techniques (Becker et al., 2007; Bostater et al., 2004;Hirano et al., 2003; Schmidt and Skidmore, 2003;Williams et al., 2003).

To summarize, detecting and monitoring invasive weed species inwetland and aquatic ecosystems at the region-wide scale is complicatedby considerable physical and environmental variability. The spatialheterogeneity of these systems requires moderate to high spatialresolution (b5m) imagery (Becker et al., 2007). High spectral resolutionimaging (N100 bands with narrow bandwidths) can be used to improvediscrimination of target species and resolve complex mixing problems.Flightline effects also contribute additional variability.We overcame thisvariability by incorporating multiple widely available techniques intonovel classification schemes for water hyacinth and SAV detection, andby using a logistic regression model to map perennial pepperweed.

2. Study site

The Sacramento–San Joaquin River Delta (hereafter the “Delta”) isformed by the confluence of the Sacramento and San Joaquin Rivers anddrains into the Pacific Ocean via the San Francisco Bay (Fig. 1). The Bay–Delta is the largest estuary in the western United States and its water-shed drains over 160,000 km2 of California. The hydrodynamic hetero-geneity of the Delta system ismanifest in awide range of salinities, tidalfluxes, water depths, and freshwater inflowswith extreme seasonal andinterannual variability (Jassby and Cloern, 2000). Saline waters flowupstream into the Delta on flood tides, and freshwater inflows from theSacramento, San Joaquin, and tributary rivers flow downstream at anaverage of 1700±300 m3 s−1 in the winter and 540±40 m3 s−1 in thesummer (Jassby and Cloern, 2000). The El Niño–Southern Oscillationinfluences precipitation patterns, resulting in extremely dry or wetyears, that respectively translate into very low flow (e.g. 230 m3 s−1 in1977) or high flow (e.g. 2700 m3 s−1 in 1983) water years (Jassby andCloern, 2000). This study focuses on the Central Delta (2139 km2),which

contains the confluence of the Sacramento and San Joaquin Rivers andthe network of small channels, rivers, and lakes formed by floodedagricultural tracts that connects them.

The Delta provides drinking water, agricultural water and land,recreational opportunities in the formof boating andfishing, and shippingaccess to the cities of Stockton and Sacramento. The Delta is a major hubfor freshwater conveyance systems in the state, providing drinking waterfor 25 million people and important irrigation resources for California's$32 billion agricultural industry (CDFA, 2006). The estuary contains ha-bitat for waterfowl along the Pacific flyway as well as threatened andendangered fish species such as Delta smelt, longfin smelt, and chinooksalmon. Unfortunately, the Deltamay have the largest number of invasivespecies of any estuary in the world (Cohen and Carlton, 1998) and is thefocusof amassive coordinatedecosystemrestorationprogram,withdirectexpenditures exceeding USD $1 billion beginning in themid-1990s (Lundet al., 2007). Invasions of aquatic weeds into the Delta negatively impactecosystem health, drinking and agricultural water quality, pumping,recreation, and shipping (Bossard et al., 2000), and may prevent thesuccess of native habitat restoration projects within the Delta (Simenstadet al., 2000). The three focal invasive species (perennial pepperweed,water hyacinth, and Brazilian waterweed; Fig. 2) are all considered pro-blems in the Delta, and have been granted high visibility and threat statusaccording to the California Invasive Plant Council (Cal-IPC) and the Cali-fornia Department of Food and Agriculture (CDFA).

3. Remotely sensed data

To map perennial pepperweed, water hyacinth and SAV (dominatedby Brazilian waterweed), we acquired 64 HyMap flightlines encom-passing the entire Delta (Fig. 1). HyMap is an airborne hyperspectralimager that collects 128 bands in the visible and near-infrared (VNIR;0.45–1.50 μm) through the shortwave infrared (SWIR; 1.50–2.5 µm), atbandwidths from 10 nm in the VNIR to 15–20 nm in the SWIR (Cockset al., 1998). The spatial resolution of the data is 3m,with a swathwidthof 1.5 km. The areawas flown in an east–west orientation, at an altitudeof approximately 1510 m. Imagery was collected during late morningand earlyafternoon low tides between June22and July 5, 2005, and June21 and June 26, 2006. The HyMap images were converted to apparentsurface reflectance using the HyCorr atmospheric correction software(Hyvista Corp., Sydney, Australia), amodified versionof theAtmosphericRemoval (ATREM) algorithm (Gao et al., 1993). We used an orthorecti-fication algorithm (Analytical Imaging and Geophysics, Boulder, CO) toorthorectify the imagery using the United States Geologic SurveyNational Elevation Set Digital Elevation Models and a set of 1-foot(covering 30 flightlines) and 1-meter (covering 34 flightlines) colororthophotos. Aminimum of 20 ground control points with a total RMSE≤1.0wereused foreachflightline. Visual inspection of image registrationaccuracy confirmed a registration error of ≃1 pixel. For a detailed reportof image data collection and preprocessing steps please see Ustin et al.(2006).

4. Ground reference data

4.1. Perennial pepperweed

Field observations were recorded between 2002 and 2006 witha comprehensive inventory of perennial pepperweed populationson selected Cosumnes River Preserve (CRP) lands (Viers et al., 2005).Mapped botanical surveys of perennial pepperweed were conductedduring flowering to aid in the identification of pepperweed patches,which were defined as areas containing a minimum of one indivi-dual and located at least 3 m from another patch. For each perennialpepperweed patch, GPS polygons delineating the patch, the number ofindividuals, areal percent cover, and patch area were recorded (1 m2

minimum). A total of 345 patches over 15.5 ha were mapped by theinventory.

Fig. 1.Map of the California Bay–Delta. Overlaid are the boundaries of the HyMap flightlines across which water hyacinth and submerged aquatic vegetation were mapped. An insethighlights the Cosumnes River Preserve, where perennial pepperweed was mapped.

4036 E.L. Hestir et al. / Remote Sensing of Environment 112 (2008) 4034–4047

4.2. Water hyacinth and SAV

Field data collection efforts for both water hyacinth and SAV wereconducted simultaneously with hyperspectral image acquisition. GPSlocations of large (≥3×3 m), homogenous patches of aquatic vegetation

(submerged and floating or emergent species), patch size, percent coverof target and co-occurring species, algal cover, and presence of inflo-rescences were recorded by survey crews in six boats. Eight emergent orfloating species and five submerged aquatic species were selected forfield identification. Emergent and floating species included water

Fig. 2. The three focal invasive species for monitoring in the Delta are a) perennial pepperweed, b) water hyacinth, c) Brazilian waterweed.

4037E.L. Hestir et al. / Remote Sensing of Environment 112 (2008) 4034–4047

hyacinth, pennywort (Hydrocotyle ranunculoides), water primrose(Ludwigia spp.), California tule (Schoenoplectus acutus), cattail (Typhaspp.), the wetland common reed (Phragmites australis), mosquitofern(Azolla spp.), and duckweed (Lemma minor). Submerged aquatic vege-tation species recorded included Brazilian waterweed, Eurasian water-milfoil (Myriophyllum spicatum), coon's tail (Ceratophyllum demersum),Carolina fanwort (Cabomba caroliniana), and pondweed (Potamogeton spp.,and Stuckenia spp.). We also collected GPS locations of water with noapparent algae or vegetation and assessed turbidity with a Secchi disk. Atotal of 2631points over 33,000hawere collected in2005, and3285pointsover the same areawere collected in 2006. Field data pointswere screeneda posteriori for accuracy in species identification, homogeneity inpatch sizeand percent cover, and geographic error associated with themovement ofboatswithwind and currents duringpoint collection. This resulted in 2381ground reference points for 2005, and 2962 for 2006.

Table 1Hyperspectral remote sensing tools used in case studies

Analysis tool Summary

Minimum noise fraction (MNF) Essentially a two-tiered principal components analysidata dimensionality for subsequent analyses (Green eeffective for mapping invasive weeds with HyMap dat

Band indexes Augment spectral differences between bands, highlighcharacteristics of vegetation (Sims and Gamon, 2002)

Spectral mixture analysis (SMA) Solves for the fraction of endmembers contained withand Cox, 1994; Adams et al., 1995). Has been successfuDelta since 2003.

Spectral Angle Mapper (SAM) Measures the similarity between image spectra and aspectra as vectors with a dimensionality equal to the(Kruse et al., 1993). Effective for identifying aquatic ve

Continuum removal Quantifies useful absorption features which can be usvegetation (Clark and Roush, 1984).

5. Case studies

The novel analysis approaches used in these case studies integratedmany well-established tools used in hyperspectral remote sensing thathave previously been applied to a variety of remote sensing studies,including invasive weed mapping (Table 1).

5.1. Cosumnes River floodplain perennial pepperweed case study

Perennial pepperweed is an herb of Eurasian origin that was intro-duced into the USA in the 1930s and is now found throughout California(Fig. 2a; Bossard et al., 2000). The weed grows in several environments,including freshwater, brackish to saline, and alkaline, and in awide rangeof habitats including riparian areas, wetlands, marshes, meadows, andfloodplains (Bossard et al., 2000; Renz and Blank, 2004; Young et al.,1997;

Relevant applications

s that whitens noise and reducest al., 1988). Demonstrated to bea.

Mundt et al. (2005), Underwood et al. (2003)

ting specific physiological.

Pontius et al. (2005), Schlerf et al. (2005),Underwood et al. (2003)

in a pixel (Smith et al., 1990; Foodylly applied to SAV detection in the

Underwood et al., (2006)

reference spectrum by treating thenumber of bands in the imagegetation.

Alberotanza (1999), Hirano et al. (2003),Merenyi et al. (2000)

ed to characterize and identify Huang et al. (2004), Kokaly et al. (2003),Underwood et al. (2003)

Table 2Confusion matrix and classification accuracy of the perennial pepperweed logistic regression model training and validation points

Ground reference

Perennial Pepperweed Pseudoabsence Total Producer's accuracy User's accuracy

Training Map classes Perennial pepperweed 458 130 588 66.1% 77.9%Unclassified 235 47,423 47,658 99.7% 99.5%Total 693 47,553 48,246 Perennial pepperweed Kappa=0.71

Validation Map classes Perennial pepperweed 138 44 182 63.0% 75.8%Unclassified 81 15,842 15,923 99.7% 99.5%Total 219 15,886 16,105 Perennial pepperweed Kappa=0.68

Table 3Classes of perennial pepperweed dominated pixels as determined by a K-meansclassification, along with the phenological interpretation; mean NDVI, NDWI, and CAIfor each group and proportion of each group classified as perennial pepperweeddominated by the regression model during training and validation

Group Phenology n NDVI NDWI CAI % classified

Training Test

1 Senescent 1 0.331 −0.093 247.0 0 02 Trees, shaded 77 0.864 0.071 6.4 1.7 03 Flowering 277 0.765 0.042 1.0 68.9 73.54 Fruiting 363 0.637 −0.016 36.3 76.3 84.95 Senescing 190 0.494 −0.062 102.9 59.6 64.6

4038 E.L. Hestir et al. / Remote Sensing of Environment 112 (2008) 4034–4047

Young et al.,1995). It is also often found in hypersaline conditions (Spenst,2006), although it is not an obligate halophyte. The natural history ofperennial pepperweed,which spreads via prolific seedproduction (Younget al., 1997) and highly friable root fragments (Young et al., 1998), enableshighly effective propagation and establishment in riparian and floodplainareas.

The Cosumnes River Preserve (CRP) is a collection of parcels (18,000 ha)in the Delta that are managed as a combination of working and naturallandscapes by a consortium of state, federal, and non-profit organiza-tions. The preserve includes a lowland riverfloodplain that is the focus ofan ongoing study to characterize the successional trajectory followingreconnection of the Cosumnes River with its floodplain. Restoration ofthe riverfloodplainwas initiated byanaccidental levee breach (ca.1985),and subsequently by intentional breaching of levees during the late-1990s (Florsheim andMount, 2003). Since then, the river channel and itsfloodplain have undergone considerable geomorphic change (FlorsheimandMount, 2002, 2003) that has increased thehabitatheterogeneity dueto colonization of riparian and floodplain vegetation on sand deposits.However, invasion by non-native plant species including perennial pep-perweed threatens the long-term success of this restoration project.

Perennial pepperweed frequently reveals mixed, degraded spectradue to its sparse architecture, and the features that lend spectral dis-tinctness at the canopy scale (chiefly, high reflectance throughout thevisible imparted by characteristic white inflorescences (Andrew andUstin, 2006); resemble vegetation mixed with litter or soil at the pixelscale. We therefore limited our classification to pixels dominated byperennial pepperweed defined as those where perennial pepperweedcovers at least 75% of the total pixel area. Because of the wide range inenvironmental and hydrological conditions throughout the CRP, at thetime of image acquisition, perennial pepperweed exhibited vegetative,flowering, fruiting, and senescing phenologies. The large amount ofspectral variation caused by flightline effects, environmental variability,and multiple phenologic stages, in combination with other naturalenvironmental variations, precluded the use of sophisticated unmixingtechniques that have been successful at other sites (notably mixture-tuned matched filtering, Andrew and Ustin, submitted for publication).Instead, our mapping strategy used a two-tiered approach: first wegenerated a continuousmap of the probability of perennial pepperweedpresence using a logistic regression model, and second we set a thres-hold to classify all pixels with N75% perennial pepperweed cover asbeing “perennial pepperweed dominated”.

5.1.1. MethodsEight of the 64 2005HyMapflightlines of the Delta (acquired 23 June

2005 and 28–29 June 2005) contained the CRP. The portions of theseflightlines encompassing CRP were cropped and mosaicked together. Aminimum noise fraction (MNF) transformation was applied to thehyperspectralmosaic. Logistic regressionmodelswere developed in JMPIN (v. 5.0, SAS Institute, Cary, NC) to predict the per-pixel probability ofthe occurrence of perennial pepperweed. The regression was trainedand validated using a randomsample of pixelswithN75% cover from theCRP inventory (n=930) as well as a random sample of pseudo-absencepoints (n=63,439). Thepseudo-absencepointswere a randomsample ofpixels selected from the region most comprehensively surveyed for

perennial pepperweed. Known perennial pepperweed pixels wereexcluded from the random sample, and remaining pixels were assumedto indicate perennial pepperweed absence. 75% of the presence (n=693)and pseudo-absence points (n=47,553) were used to train the regres-sion. MNF bands containing negligible informationwere excluded fromanalysis.MNFbandswere visually inspected and those exhibiting severeflightline effects were discarded. The remaining bands were chosenthrough stepwise entry into the logistic regression model. A predictionformula usingMNF bands 2–11,15, and 17 optimized the discriminationof perennial pepperweed from other land cover types (evaluated usingR2). This model was assessed by regressing the estimated probability ofperennial pepperweed occurrence (p) against percent cover estimatesfrom the CRP inventory data. The value of p corresponding to 75% coverby perennial pepperweed was identified and pixels were classified as“perennial pepperweed dominated” if their predicted value exceededthis threshold.

Classification accuracy was assessed with the remaining perennialpepperweed presence (n=219) and pseudo-absence points (n=15,886).To assess the contribution of phenological variation to omission errors,we divided all perennial pepperweed dominated pixels into phenolo-gical classes using a K-means unsupervised classification on the sameMNF bands as used by the regressionmodel. Phenology of these classeswas determined through visual inspection and verified with three phy-siological indexes: the normalized difference vegetation index (NDVI;Tucker,1979), thenormalized differencewater index (NDWI; Gao,1996),and the cellulose absorption index (CAI; Nagler et al., 2000).

5.1.2. ResultsThe predicted value (p) provided a reasonably good indicator of the

percent cover of perennial pepperweed within a pixel (p=−0.084253+0.0072013[% cover perennial pepperweed], R2=0.41, n=2787). The finalclassification distinguished perennial pepperweed from pseudo-absencepixels (Table 2)with user's and producer's accuracies of perennial pepper-weed detection of 75.8% and 63.0% respectively, and Kappa coefficientrelative to the perennial pepperweed class of 0.68.

The K-means classifier found five distinct groups of perennialpepperweed dominated pixels (Table 3). These groups fell out along agradient of dryness/senescence and were interpreted as completelysenescent (n=1), shaded by or under a tree canopy (n=77), flowering(n=277), fruiting (n=363), and senescing (n=190). These interpretations

4039E.L. Hestir et al. / Remote Sensing of Environment 112 (2008) 4034–4047

were supported by the physiological indexes. NDVI, NDWI, and CAI allshowed significant effect of group (pb0.0001; ANOVA) and, for eachindex, all groups were significantly different from each other at thep=0.05 level (Tukey test, Sokal and Rohlf, 2000). The detection rate ofthese classes ranged from 0% (senescent) to 85% (fruiting), with fruitingand flowering phenologies showing the highest detection rate (Table 3)in both the training and test sets. The fruiting, flowering, and senescingphenology classes are themost relevant to a remote sensing study.Whenjust these three classes were used to evaluate the classifier performance,the producer's accuracy of the classifier increased to 76.6%.

5.2. California Delta water hyacinth case study

Introduced to the Sacramento River in 1904 by horticulturalists(Cohen and Carlton, 1995; Finlayson, 1983; Toft et al., 2003) waterhyacinth now obstructs navigable waterways, degrades water quality,foulswater pumps, blocks irrigation channels, andhas causedsignificantchanges to ecological assemblages throughout the Delta (Bossard et al.,2000; Toft et al., 2003). The California Department of Boating andWaterways is themanagement agency responsible for controllingwaterhyacinth in the Delta (Cohen and Carlton, 1995), which is done throughthe application of chemical herbicides and mechanical removal.

One of the fastest growing plant species in the world (N1 ton drymatter day−1 ha−1; Bossard et al., 2000), water hyacinth is a floatingaquatic plant forming dense stands that exhibit diverse leaf and canopymorphologies, dependent on patch location and mat density (Fig. 2b).Within a single mat of water hyacinth several different leaf morphol-

Fig. 3. Schematic of the decision tree used to identify w

ogies and phenologies can be found (Gopal, 1987). In the Delta, waterhyacinth often occurs alongside other floating emergent species likepennywort and water primrose.

As with perennial pepperweed, the spectral signature of waterhyacinth is highly variable as the plant can exhibit many contempora-neous phenologies including vegetative, flowering, stressed, andsenescent forms that result fromherbicide application and frost damageto overwinteringplants. The spectral variabilityofwaterhyacinth resultsin spectral overlap with the co-occurring floating plant species penny-wort and water primrose, which leads to mixed pixels with spectrallysimilar constituents. Spectral mixture analysis (SMA; Smith et al., 1990),Spectral AngleMapper (SAM; Kruse et al.,1993) and continuumremoval(Clark and Roush, 1984) have been demonstrated as an effective tech-nique for species-level vegetation detection. Before using these tech-niques for water hyacinth mapping, we explicitly eliminated othersources of confusion (e.g. terrestrial pixels, mixing with water, SAV, andother emergent plant species) from the analysis area.

5.2.1. MethodsThe strategy forwater hyacinth classificationwas to develop a binary

decision tree, where each node reduces the variance in the remainingpixels until amap of hyacinth presence and absencewas achieved, usingthe 2005HyMap data (64 flightlines). The 2005 field datawas separatedinto training (n=1027) and validation (n=1354) data for the decisiontree. This same methodology was also applied to the 2006 HyMap datausing 2006 field data separated into 158 training and 2804 validationpoints. Fig. 3 depicts a schematic of the water hyacinth decision tree.

ater hyacinth in 2005 across 64 HyMap flightlines.

Fig. 5. Trainingendmembers for the Spectral AngleMapper nodeof the 2005 decision tree.Spectra were extracted from the HyMap imagery from ground reference data points.

4040 E.L. Hestir et al. / Remote Sensing of Environment 112 (2008) 4034–4047

In order to discriminate aquatic versus terrestrial land cover types,we first created a mask of the water boundaries (which includes sub-merged and emergent aquatic vegetation) bymanually adjusting the USBureau of Reclamation GIS layer of waterways. In ArcGIS (version 9.2,ESRI Inc, Redlands, CA) the line map of waterways was overlaid on acolor infrared composite of theHyMap images and the edge of thewaterboundary wasmatchedwith the image data. An unconstrained spectralmixture analysis (SMA)wasused to identifyandexclude any leveebanksthat may not have been removed by the watermask. The SMA wasperformed using spectral endmembers of aquatic vegetation, water, andsoil derived from the imagery using ground reference data (Fig. 4). Allpixels in which the SMA fraction for water was b10% or average albedobetween 1.5995 µmand 1.764 µmwas higher than 4%were identified asaquatic vegetation. The average reflectance of bands in the shortwaveinfrared wavelengths 1.5995 µm–1.764 µm (HyMap bands 78–91) wasused to discriminate floating emergent species from submerged aquaticvegetation, sparse tule, and turbidwater. Floatingemergent specieshavedistinctively high albedo in the SWIRwhereas submerged aquatic vege-tation, sparse tule and turbid water have muted albedo in the SWIR dueto strong infrared absorption by water. Thus, pixels with SWIR reflec-tance values less than 4% were classified as submerged aquatic vege-tation, sparse tule, or turbid water. All other pixels were classified asemergent aquatic vegetation. Note that although tule is an emergentplant, under moderate to low percent cover its reflectance values acrossthe spectrum are usually lower than those of floating emergent speciesas it has anerectophile canopystructure andoftenoccurs inmixedpixelswith water.

Inorder to separatedarkeremergentvegetation suchas tule andcattailwith high percent cover from bright floating emergent vegetation such aswaterhyacinth, pennywort, andprimrose, the average reflectance value inthe near-infrared 1.0451 µm–1.1186 µm (HyMap bands 42–47) was cal-culated. Pixels identified as emergent aquatic vegetation by the previousnode in the decision tree were further classified as “dark” emergentaquatic vegetation if their NIR reflectance value was less than 40%otherwise they were classified as “bright” floating emergent aquaticvegetation. This is anecessary stepasmanyof thepixels classifiedaswaterhyacinth “spectral class 2” (Fig. 5) are significantly darker in the near-infrared, and must be separated from brighter emergent aquatic vege-tation before being classified to the species level. We next set the NIRreflectance threshold even lower to separate tule from other emergentspecies. Dark emergent aquatic vegetation pixels with less than 24%reflectance in 1.0451 µm–1.1186 µmwere classified as tule.

Fig. 4. Endmembers of healthy vegetation, soil, andwater as used in the first spectralmixture an

The final nodes of the decision tree used both the Spectral AngleMapper (SAM) technique and a water absorption feature to separatewater hyacinth from co-occurring floating and emergent species in bothclasses of vegetation (bright and dark). Water hyacinth is a fleshy, suc-culent plant, having greater foliar water content than the co-occurringfloating species. To take advantage of this difference,weused continuumremoval (Clark and Roush, 1984) to quantify the water absorptionfeature locatedbetween0.907µmand1.0746µm(HyMapbands33–44).Lower continuum-removed values indicate greater water content.Spectral Angle Mapper is appropriate for species-level monitoring atthe regional scale as it is not sensitive to albedo differences across the 64flightlines; SAM can identify subtle differences in pixel spectra whilebeing insensitive to gain factors. SAM was applied to a subset of thevisible and NIR bands (0.45 μm to 0.9541 μm) to reduce noise contri-buted to the pixel spectra from mixing with water. The SAM techniqueused endmembers derived from the image data that were representa-tive of different phenological states and geographic locations through-out the dataset, as determined from the field data (Fig. 5). In the dark

alysis. Spectrawere extracted from theHyMap imagery fromground reference data points.

Table 4Condensed confusion matrix and classification accuracy of the 2005 water hyacinth decision tree

Ground reference

Water hyacinth Other emergent and floating vegetation Total Producer's accuracy (%) User's accuracy (%)

Map classes Water hyacinth 141 16 157 69.1 89.8Other emergent and floating vegetation 63 211 274 93.0 77.0Total 204 227 431

The Kappa coefficient relative to the water hyacinth class is 0.81.

4041E.L. Hestir et al. / Remote Sensing of Environment 112 (2008) 4034–4047

emergent class identified by low NIR average reflectance, pixels with acontinuum removal value less than 0.86, when coupled with the SAMwater hyacinth classification output were classified as water hyacinthspectral class 1. In the bright class, pixels were first classified using theSAMoutput to identifywaterhyacinth spectral class 2 and the remainingpixels from the continuum removal were classified as water hyacinthspectral class 1 if their value was less than 0.83, otherwise they wereclassified as pennywort or water primrose. These thresholds were de-termined empirically based on our field data.

Classification results for validation pixels were extracted from theclassified imagery using STARSPAN (http://starspan.casil.ucdavis.edu),an algorithm developed to provide fast, selective pixel extraction fromraster and vector data (Rueda et al., 2005). The accuracy of theemergent classification was assessed with a confusion matrix andkappa statistics were calculated (Lillesand et al., 2004).

5.2.2. ResultsIn 2005, water hyacinth infested 167 ha in the Delta. Out of the 204

field points of water hyacinth that were used for validation, 141 werecorrectly classified as water hyacinth and 63 were misclassified,resulting in a producer's accuracy of 69.1%. Sixteen pixels wereerroneously classified as water hyacinth, resulting in a user's accuracyof 89.8%. The estimated Kappa coefficient relative to the water hyacinthclass is 0.81 (Table 4), a value considered to show strong agreement(Congalton, 1996). We interpret water hyacinth spectral class 1 to beeither stressed or flowering water hyacinth due to its darker reflectancein the near-infrared (Hardisky et al., 1986) and a shallower waterabsorption feature at 0.907 µm–1.0746 µm (Peñuelas et al., 1993). Basedon field observations, these mats have sparser canopies and smallerleaves, resulting in greater confusion with other floating emergentspecies. When considering the detection rate of these two spectralclasses, the producer's accuracy of the healthy/robust water hyacinthclass increases to 93.3%, and the producer's accuracy of the other class is85.1%. Of the 16 commission pixels, 6 were mistakenly identified ashealthy/robust water hyacinth, and 10 were incorrectly classified asstressed or flowering water hyacinth.

5.2.3. 2006 Water Hyacinth classificationA decision tree containing these same variables but with different

thresholds determined from 2006 field data was applied to the 2006HyMap imagery. The areal infestation of water hyacinth increased in2006, where 303 hawere identified. Out of the 218 water hyacinth fieldpoints used for classification validation, 112 pixels were correctlyclassified as water hyacinth, and 106 were misclassified, producing auser's accuracy of 51.4%. Sixty-nine pixels were incorrectly identified aswater hyacinth, resulting in a producer's accuracy of 61.9% (Table 5). Theestimated Kappa coefficient relative to thewater hyacinth class is 0.49, a

Table 5Condensed confusion matrix and classification accuracy of the 2006 water hyacinth decisio

Ground reference

Water hyacinth Other emerg

Map classes Water hyacinth 112 106Other emergent and floating vegetation 106 592Total 218 661

The Kappa coefficient relative to the water hyacinth class is 0.49.

value considered to show moderate agreement (Congalton, 1996). Theproducer's accuracies of the healthy/robust water hyacinth and thestressed/flowering hyacinth are 86.5% and 44.9%, respectively; only 10comissed pixels were erroneously classified as healthy or robust waterhyacinth, whereas 59 commission pixels were misclassified as stressedor floweringwater hyacinth. Thus, the decision tree approach describedabove accounts for thevariability inherent in suchadynamic ecosystem;successfully identifying the target species over multiple flightlines, andwith minor modifications, over multiple years without requiringenormous training sets for each new year. Nearly half of the 2005 fielddata were needed to train the 2005 decision tree, whereas only about atenth was needed in 2006.

5.3. California Delta SAV case study

Brazilian waterweed is an SAV species that is the focus of consi-derable concern in the Delta. Delta aquatic habitats are generally con-sidered degraded (Meng and Moyle, 1995), but more specifically,Brazilianwaterweed infested areas hinder themovement of threatenedand endangered fishes such as anadromous salmonids, splittail, andDelta smelt (Brown, 2003), and it may prove to make restoration ofnative habitat nearly impossible where it is firmly established (Simen-stad et al., 2000). Furthermore, dense mats of Brazilian waterweedinterfere with recreational and commercial activities such as fishing,boating, swimming, and urban water and irrigation pumping (Bossardet al., 2000). The California Department of Boating and Waterways isagain the responsible agency for controlling Brazilian waterweed inorder to maintain navigable waterways in the Delta.

Brazilianwaterweed forms thick, stands of long, intertwining stemsand has small white flowers that float at the water surface (Fig. 2c),spreading via fragmentation (Bossard et al., 2000). Its stems can growaslong as 4.5 m, branch frequently, and although usually rooted to thesubstrate as deep as 7 m below the water surface (Parsons andCuthbertson,1992), it can also be found as a free floatingmat (Bossard etal., 2000). Brazilian waterweed is adapted to low-light conditions andturbidwater appears to favor growth, but it has also been found to thriveunder red light, suggesting an affinity for the water surface (Bossard etal., 2000).

The remote sensing signal of SAV at the water surface can becomedegraded due to attenuation of light with water depth, which varieswidely based onbathymetric variability and tidal variability (Underwoodet al., 2006), and increased concentrations of water constituents whichalso vary across the region. Additionally, non-linear spectral mixingoccurs as epiphytic algae often grow on Brazilian waterweed and otherspecies of SAV, and emergent algae mats form on top of SAV canopies,obscuring the spectral target. SMA has been successfully applied toBrazilian waterweed discrimination at local sites (single flightlines) in

n tree

ent and floating vegetation Total Producer's accuracy (%) User's accuracy (%)

181 51.4 61.9698 89.6 84.8879

Fig. 6. Schematic of the decision tree used to identify SAV in 2005 across 64 HyMap flightlines.

Fig. 7. Endmembers of emergent algae, turbid water, relatively “clear”water, and SAV asused in a “nested” second spectral mixture analysis to identify SAV. Spectra wereextracted from the HyMap imagery from ground reference data points.

4042 E.L. Hestir et al. / Remote Sensing of Environment 112 (2008) 4034–4047

theDelta since 2003 (Mulitsch andUstin, 2003;Underwood et al., 2006).After reducing other sources of variance from turbid water and very lowcover tule in an approach similar to the water hyacinth case study(Section 5.2), a combination of spectral unmixing and SAMwere used toovercome the mixing problem and to discriminate Brazilianwaterweeddominated SAV from other water constituents and isolate it from algaewhile taking flightline effects into account.

5.3.1. MethodsA decision tree approach was used to map SAV across the 64 2005

HyMap flightlines (Fig. 6). Training (n=1027) and validation data(n=1354) are described above. All pixels previously classified by theSMA and the SWIR reflectance value threshold (Section 5.2.1.) as SAV,turbid water, or tule were analyzed using this tree.

The classic “red edge” reflectance feature, indicative of chlorophyll invegetation (Curran et al., 1991; Fililla and Peñuelas, 1994; Gitelson andMerzlyak,1997)wasquantifiedwithHyMapbands 19 and16 (0.7313µmand 0.6736 µm) in a normalized difference band index which separatedturbid water pixels from vegetated pixels. Hoogenboom et al. (1998)found that turbid water bodies with relatively high total suspendedsolids (TSS) and low phytoplankton concentrations, such as the Delta,display high reflectance up to 0.72 µm. Thus it is relatively easy toseparate the red edge from visibly “bright” turbid water.

Although most emergent aquatic species were screened using theSMA and SWIR thresholds described in the water hyacinth decisiontree, it was necessary to isolate tule pixels of low percent cover thatwere classified as SAV because of the spectral dominance of water. The

average reflectance of the SWIR bands (1.5995 µm–1.764 µm) fromSection 4.2.1. was used, and the threshold was lowered so that pixelswith SWIR reflectance less than 3% were classified as submergedaquatic vegetation.

Table 6Condensed confusion matrix and classification accuracy of the 2005 SAV decision tree

Ground reference

SAV Water Emergent and floating vegetation Total Producer's accuracy (%) User's accuracy (%)

Map classes SAV 187 2 14 203 59.2 92.1Water 50 37 16 103 90.2 35.9Emergent and floating vegetation 79 2 349 430 92.1 81.2Total 316 41 379 736

The Kappa coefficient relative to the SAV class is 0.86.

4043E.L. Hestir et al. / Remote Sensing of Environment 112 (2008) 4034–4047

SMA is most successful when just a few endmembers that arespectrally distinct are used to train the SMA. When more similarendmembers are included, the RMSE of the estimated fractionsincreases (Roberts et al., 1993). To circumvent this problem, a secondSMAwas conducted on pixels with awater fraction of 10% or greater asdetermined by the first SMA (Section 5.2.1.) and the low SWIR albedo(5.3.1.). The second SMA was trained with four endmembers selectedfrom the imagery using ground reference points of clear water, turbidwater, SAV, and emergent algae. Although the endmembers appearspectrally similar (Fig. 7), they effectively encompassed the range ofspectral variation present in pixels analyzed, which was itself limited,resulting in acceptably accurate (based on validation data) SMAfractions. However, SMA and classic chlorophyll absorption character-istics could not fully identify SAV and discriminate it from otherconstituents of the water column. In order to further discriminatewater from submerged aquatic vegetation and to isolate SAV fromemergent algae on the water surface, SAM was conducted using thevisible and NIR wavelengths (as described in Section 5.2.1.). This stepwas included as the final node in the SAV decision tree.

At the time of this analysis, we were unable to identify individualSAV species accurately. Even in pixels with 100% SAV cover at thewater surface, pixel reflectance is still dominated by the very strongabsorption of water and the spectral variability between species isslight. Nevertheless, Brazilian waterweed has been successfullydiscriminated from co-occurring SAV species using SMA at verysmall-scale sites (Mulitsch, 2005). However, given the large scale ofthe entire Delta, and the extensive variation resulting from the manyconfounding factors discussed, we have not yet demonstrated that thesubtle spectral differences between SAV species are consistentlydetectable.

5.3.2. ResultsIn 2005 2246.5 ha of Delta waterways were infested with SAV. Out

of 316 validation points collected of submerged aquatic vegetation,

Fig. 8. Endmembers used to train the Spectral Angle Mapper used in both the 2006 water hground reference data points.

187 were correctly classified as SAV, 50 were misclassified as water,and 79 were incorrectly classified as emergent vegetation, resulting ina producer's accuracy of 59.2%; 16 field points were misclassified asSAV, resulting in a user's accuracy of 92.1%. The Kappa coefficientrelative to the SAV class is 0.86 (Table 6), indicating strong agreement(Congalton, 1996).

5.3.3. 2006 SAV decision treeAlthough wewere not able to differentiate SAV to the species level,

we were able to develop a method that, like the water hyacinthdecision tree, is repeatable across multiple flightlines and multipleyears for the same geographic location. A slightly modified decisiontree was developed for 2006 imagery that could detect three differentSAV classes based on the output of Spectral Angle Mapper. However,other than the additional SAM classes (see Fig. 8 for SAM end-members), the decision tree contained the same variables as the 2005tree. Areal infestation of SAV decreased to 1722.8 ha in 2006. Out ofthe 1218 SAV field points used for validation, 844 were correctlyidentified as SAV, 255 were misclassified as emergent vegetation, and119 were misclassified as water, resulting in a producer's accuracy of69.3%. Sixty-one field points were incorrectly classified at SAV,resulting in a user's accuracy of 93.3%. The Kappa coefficient relativeto the SAV class is 0.87 (Table 7), a value considered to show strongagreement (Congalton, 1996).

6. Discussion and conclusions

The approaches in these case studies provide useful solutions forovercoming the spectral heterogeneity inherent to aquatic andwetland ecosystems. We found that hyperspectral remote sensingcan be used to map invasive weeds in extensive dynamic ecosystemssuch as the Delta, and that multiple hyperspectral tools can becombined to accommodate high variability. Our case studies havedemonstrated that both traditional and novel approaches to species

yacinth and SAV decision trees. Spectra were extracted from the HyMap imagery from

Table 7Condensed confusion matrix and classification accuracy of the 2006 SAV decision tree

Ground reference

SAV Water Emergent and floating vegetation Total Producer's accuracy (%) User's accuracy (%)

Map classes SAV 844 12 49 905 69.3 93.3Water 255 337 57 649 96.3 51.9Emergent and floating vegetation 119 1 779 899 88.0 86.6Total 1218 350 885 2453

The Kappa coefficient relative to the SAV class is 0.87.

4044 E.L. Hestir et al. / Remote Sensing of Environment 112 (2008) 4034–4047

mapping and monitoring are effective for use with large datasets,composed of many flightlines with sometimes severe flightline effectsand multiple sources of variability. While we encourage the use of themethodology presented, we highlight the challenges involved so thatthey might serve as guidelines for future studies.

6.1. Meteorological and illumination geometry variability

A certain amount of spectral degradation can be attributed to spe-cular reflection from thewater surface receivedby the sensor, which candistort or block the features of interest. We used careful flight planningto avoid specular reflectance; we controlled for meteorological condi-tions, such as wind velocities and time of day (Lillesand et al., 2004;Masuda et al., 1988) during data acquisition. In the Delta, summermeteorological conditions are characterized by morning low cloudcover, followed by moderate to high winds in the late mornings andearly afternoons. The need for optimal weather conditions within anarrow range of sun angles forced imagery to be acquired over thecourse of twoweeks in 2005, which increased inter-flightline variabilitydue to differences in BRDF, ground conditions and environmental con-ditions. This variability had to be overcome with a decision tree ap-proach for water hyacinth and Brazilianwaterweed and limited the useof MNF, which is very sensitive to flightline effects, to the local-scaleperennial pepperweed analysis.However, itmaynot bepossible, or cost-effective, to control for sunglint by careful flight planning in all studies.Correction techniques, which minimize the effects of sunglint andinformation loss, are available to reduce this impact (Hedley et al., 2005;Hochberg et al., 2003) but were not used in this study.

Previous studies have shown that the contribution of SAV to thewater-leaving reflectance received by the sensor decreases as the depthof the water column between macrophytes and the sensor increases(Han and Rundquist, 2003), thus degrading the measured spectral sig-nature of the target species (Han, 2002). Some hyperspectral remotesensing analyses of benthic organisms have appliedwater depth correc-tions based on a homogenouswater column (Holden and LeDrew, 2001)ormore sophisticated techniques that use spectral libraries created fromradiative transfer computations (Hedley and Mumby, 2003; Louchardet al., 2003; Mobley et al., 2005). Our image acquisition protocol at-tempted to address some of this issue by further restricting the imageacquisition to low tide conditions.

6.2. Pixel constituent variability

In the three case studies presented, pixel composition presents acommon problem. The architecture and patchy distribution of perennialpepperweed reduce the strength and distinctiveness of its spectralsignature due to mixing with other vegetation, litter, and soil. Waterhyacinth is commonly found inmixedpatcheswithpennywort andwaterprimrose, and the SAV class is characterized by mixing among multiplesubmerged species and mixing with the water column itself. Multipleapproaches have been proposed to improve class separation includingimprovingendmember collection (Tompkins et al.,1997), using stochasticmixture models (Eismann and Hardie, 2004), and applying multipleendmember spectral mixture analysis (MESMA) that can account forphenological changes (Dennison and Roberts, 2003) or water quality.

SAV presents additional mixture problems due to the overlyingwater column. The Delta has high turbidity and there are gradientsinwater quality throughout the system caused by tidal flux and runoff.This added pixel constituent creates a most decidedly non-linear,three-dimensional mixing problem that reduces the signal of thetarget species; 20.9% of SAV field points were misclassified as water in2005. In addition, emergent algae, which have a distinct spectralsignature, occur epiphytically on the SAV canopy, making it necessaryto exclude pixels identified as emergent algae from the SAV class. Thisin turn results in an underestimate of SAV distribution. When presenton the water surface, supported by understory SAV, emergent algaeresemble emergent aquatic vegetation (Underwood et al., 2006). Ourresults corroborate this pattern: nearly 10% of SAV field points, whichwere generally covered in algae, were classified as emergent vegeta-tion. Our methods largely assume that planktonic algae play aninsignificant role in the turbidity of the Delta. However, there areregional locations of very high phytoplankton content that furtherconfound SAV detection. For example, chlorophyll concentrations inthe shallow eastern Delta sloughs and lower San Joaquin River arehigh and variable (Ball and Arthur, 1979). Although we were notable to successfully discriminate SAV species, several studies thathave been used to successfully resolve benthic substrates and speciesthrough a variety of water variables and depths (Kutser et al., 2003;Legleiter et al., 2004; Louchard et al., 2003; Louchard et al., 2002;Vahtmäe et al., 2006) suggest this may be possible upon furtherinvestigation. From our experience mapping SAV, we recommendfuture research to adequately identify sources of variability, charac-terize planktonic vs. SAV chlorophyll, and continue to investigatespecies-level spectral differences of SAV.

6.3. Biological variability

Species life history stages have different spectral characteristics.Whenever possible, data collection (both field and remotely sensed)must consider the life history of the target species (Fig. 9). The threespecies in our case studies all differ phenologically. Brazilian water-weed flowers early in the summer, at the time of data acquisition,however, these flowers are relatively small and sparse in the weedmat. Brazilian waterweed has two peaks in its growth, an initial oneduring the beginning of the summer, and a stronger peak once againduring late summer. An acquisition later in the summer may reduceomission errors as plants may be givenmore time to grow to thewatersurface, and may help improve classification to the species level.

Of the perennial pepperweed phenological classes identified bythe K-means classifier, the classes interpreted as flowering, fruiting,and senescing are relevant to a remote sensing study, and the classifierperformance improved when just these were considered (resulting ina producer's accuracy of 76.6%). Accuracy can be improved bymappingeach phenological state individually. As senescence of this sparseweed progresses, the similarity of its signal to surrounding litter andsoil increases. Andrew and Ustin (2006) were able to resolve bothflowering and fruiting phenologies of perennial pepperweed from co-occurring species with spectral data, and suggested that this weed ismost tractably mapped between flowering in late spring and sene-scence in late summer (Fig. 9).

Fig. 9. Case study invasive species annual life cycle and respective June spectra extracted from HyMap imagery using ground reference locations.

4045E.L.H

estiret

al./Rem

oteSensing

ofEnvironm

ent112

(2008)4034

–4047

4046 E.L. Hestir et al. / Remote Sensing of Environment 112 (2008) 4034–4047

Water hyacinth life stages are distinct in color and can occur simul-taneously in the samemat. Accounting for water hyacinth phenologicalstates in the decision tree allowed us to better map the weed. Weobserved two distinct spectral classes. Of the 67 field points classified as“stressed”waterhyacinth,10were identified in thefield aspennywortorwater primrose (error of commission=14.9%). The second spectral classis interpreted as vegetative or “healthy” water hyacinth, and is wellcharacterized by the spectral angle mapping algorithm, with an error ofcommission of just 6.7%. However, significant growth of water hyacinthin the Delta does not occur until August through September (Fig. 9),which may be limiting the success of water hyacinth detection. Wesuggest that timing data acquisition to coincide with the peak occur-rence of vegetative water hyacinth in the late summer will improve thedetection of this target weed.

6.4. Conclusions: applications of mapping and monitoring invasive aquaticweeds

Wedescribe the challenges ofmapping aquatic and riparianweeds inthis study as ecohydrological constraints, and although they were con-sidered in flight planning for this study, it is impossible to account for allspatial and temporal variability over such a large study area (2139 km2).From the results of our case studies, we present several conclusions thatmay serve as guidelines for future regional-scale wetland mappingstrategies: 1) incorporatingmultiplemethods in adecision tree approachcan account for variability in large datasets; 2) MNF-based approachesshould be used in smaller areas, especially when sensitivity to subtlespectral differences is required; 3) mapping accuracy is improved bypartitioning species variability into phenological stages.

A further complication to remote sensing of wetlands is that landmanagers must often request post facto analyses of image data that hasbeen acquired without consideration paid to the characteristics of thetarget speciesorwithfielddata thatwasnot acquiredwith remote sensingin mind (e.g. discounting phenological state or percent cover), as was thecase with the perennial pepperweed mapping study. This being said,rarely are managers in the position of having unlimited resources andadequate foresight: more typically, managers must adaptively respond toemerging crises, such as non-native plant invasions. Therefore, remotesensing practitioners must continue to develop methodologies that over-come thechallenges inherent in riparian,wetland, andfloodplain systems,and must do so in support of early detection and rapid response.

One of the most pressing issues for managers of complex, highlyinvaded ecosystems such as theDelta is to assess infestation cover and thespatial distributionof invasive species. In spite of the intrinsic variabilityofthe Delta, hyperspectral remote sensing provides a powerful set of toolsthat canbeapplied toproducemapsof speciesdistributions.Ourapproachprovides a comprehensive, timely, and repeatable product that assessesareal cover of very diverse invasive weeds. The use of hyperspectralimagery todetect andmap invasive aquatic and riparianweedshas shownthat the technology and its methodologies have reached a stage of ma-turity; it is being successfully applied to meet operational goals for weedmanagement. For instance, California state agencies are using the resultsdescribed herein to monitor the spatial and temporal dynamics of weedpopulations under their prescribed control programs.

Acknowledgements

This researchwas supported by the California Departmentof Boatingand Waterways Agreement 03-105-114, the California Bay–DeltaAuthority U-04-SC-005, and the California Bay–Delta Authority Ecolo-gical Restoration Program (ERP-01-NO1, ERP-02D-P66) for funding.

The authors gratefully acknowledge M. Carlock of California Depart-ment of Boating&Waterways; D. Kratville, J.R.C. Leavitt, P. Akers, and thefield crews of the California Department of Food & Agriculture; M. Lay,N. Noujdina, E. Tom, M. Whiting, Y.-B. Cheng, K. Olmstead, D. Darling,L. Chan, F. Anderson, C. Rueda, K. Keightley, W. Sprague, N. Doyle,

M. Trombetti, D. Riaño, C. Ramirez, E. Zhong, S. Reed, W. Stockard, andA. Cook for the assistance in the field. We also thank H. Manaker,R.McIlvaine, G. Scheer, and L. Ross for administrative support. JHVwouldlike to thank the following for thefield assistance: I.Hogle, R.Hutchinson,J. Cuixart, J. Bonilla, B. Harbert, L. Kashiwase, A. Holguin, W. Miller,A. Jacobs, E. Lee, & N. Jensen.

References

Ackleson, S. G., & Klemas, V. (1987). Remote sensing of submerged aquatic vegetation inlower Chesapeake bay: A comparison of Landsat MSS to TM imagery. RemoteSensing of Environment, 22, 235−248.

Adams, J. B., Sabol, D. E., Kapos, V., Almeida Filho, R., Roberts, D. A., Smith, M. O., &Gillespie, A. R. (1995). Classification of multispectral images based on fractions ofendmembers: Application to land-cover change in the Brazilian Amazon. RemoteSensing of Environment, 52, 137−154.

Alberotanza, L. (1999). Hyperspectral aerial images. A valuable tool for submergedvegetation recognition in the Orbetello Lagoons, Italy. International Journal of RemoteSensing, 20, 523−533.

Andrew, M. E., & Ustin, S. L. (2006). Spectral and physiological uniqueness of perennialpepperweed (Lepidium latifolium). Weed Science, 54, 1051−1062.

Andrew, M. E., & Ustin, S. L. (submitted for publication). The role of environmentalcontext in mapping Lepidium latifolium with hyperspectral image data. RemoteSensing of Environment.

Ball, M. D., & Arthur, J. F. (1979). Planktonic chlorophyll dynamics in the Northern SanFrancisco Bay and Delta. InT. J. Conomos (Ed.), San Francisco Bay: The urbanized estuary(pp. 265−285). San Francisco, CA: Pacific Division of the American Association for theAdvancement of Science c/o California Academy of Sciences.

Becker, B. L., Lusch, D. P., & Qi, J. (2007). A classification-based assessment of the optimalspectral and spatial resolutions for Great Lakes coastal wetland imagery. RemoteSensing of Environment, 108, 111−120.

Bossard, C. C., Randall, J. M., & Hoshovsky, M. C. (Eds.). (2000). Invasive plants of California'swildlands Berkeley: University of California Press.

Bostater, C. R., Ghir, T., Bassetti, L., Hall, C., Reyeier, E., Lowers, R., Holloway-Adkins, K., &Virnstein, R. (2004). Hyperspectral remote sensing protocol development forsubmerged aquatic vegetation in shallow waters. Remote sensing of the ocean andsea ice 2003, proceedings of SPIE-volume 5233, vol. 5233. (pp. 199−215).

Brown, L. R. (2003). Will tidal wetland restoration enhance populations of native fishes?San Francisco Estuary & Watershed Science, 1, 1−42.

California Department of Food and Agriculture, C. (2006). California agricultural resourcedirectory (pp. 178). Sacramento: California Department of Food and Agriculture.

Clark, R. N., & Roush, T. L. (1984). Reflectance spectroscopy — Quantitative analysistechniques for remote sensing applications. Journal of Geophysical Research, 89,6329−6340.

Cocks, T., Jenssen, R., Stewart, A., Wilson, I., & Shields, T. (1998). The HyMap airbornehyperspectral sensor: The system, calibration, and performance. Proceedings of the1st EARSEL workshop on imaging spectroscopy Zurich, 5.

Cohen, A. N., & Carlton, J. T. (1995).Nonindigenous aquatic species in a United States estuary:A case study of the biological invasions of the San Francisco Bay Delta In (pp. 238): UnitedStates Fish and Wildlife Service National Sea Grant College Program.

Cohen, A. N., & Carlton, J. T. (1998). Accelerating invasion rate in a highly invadedestuary. science, 279, 555−558.

Congalton, R. G. (1996). Accuracy assessment: A critical component of land cover. In T. H.T. J. M. Scott & F. W. Davis (Eds.), Gap analysis: A landscape approach to biodiversityplanning (pp. 119−131). Bethesda, Maryland: American Society for Photogrammetryand Remote Sensing.

Curran, P. J., Dungan, J. L., Macler, B. A., & Plummer, S. E. (1991). The effect of a red leafpigment on the relationship between red edge and chlorophyll concentration. Re-mote Sensing of Environment, 35, 69−76.

Dennison, P. E., & Roberts, D. A. (2003). The effects of vegetation phenologyon endmemberselection and species mapping in southern California chaparral. Remote Sensing ofEnvironment, 87, 295−309.

Eismann, M. T., & Hardie, R. C. (2004). Stochastic spectral unmixing with enhancedendmember class separation. Applied Optics, 43, 6596−6608.

Fililla, I., & Peñuelas, J. (1994). The red edge position and shape as indicators of plantchlorophyll content, biomass and hydric status. International Journal of RemoteSensing, 15, 1459−1470.

Finlayson, B. J. (1983).Water hyacinth: Threat to the delta?Outdoor California. (pp.10−14).Florsheim, J. L., & Mount, J. F. (2002). Restoration of floodplain topography by sand-splay

complex formation in response to intentional levee breaches, Lower CosumnesRiver, California. Geomorphology, 44, 67−94.

Florsheim, J. L., & Mount, J. F. (2003). Changes in lowland floodplain sedimentationprocesses: Pre-disturbance to post-rehabilitation, Cosumnes River, CA.Geomorphology,56, 305−323.

Foody, G. M., & Cox, D. P. (1994). Sub-pixel land cover composition estimation using alinear mixture model and fuzzy membership functions. International Journal ofRemote Sensing, 15, 619−631.

Gao, B. -C. (1996). NDWI — A normalized difference water index for remote sensing ofvegetation liquid water from space. Remote Sensing of Environment, 58, 257−266.

Gao, B. -C., Heidebrecht, K. B., & Goetz, A. F. H. (1993). Derivation of scaled surfacereflectances from AVIRIS data. Remote Sensing of Environment, 44, 165−178.

Gitelson, A. A., & Merzlyak, M. N. (1997). Remote estimation of chlorophyll content inhigher plant leaves. International Journal of Remote Sensing, 18, 2691−2697.

4047E.L. Hestir et al. / Remote Sensing of Environment 112 (2008) 4034–4047

Gopal, B. (1987). Water hyacinth. New York: Elsevier Science Pub. Co.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.Geoscience and Remote Sensing, IEEE Transactions on, 26, 65−74.

Han, L. (2002). Spectral reflectance of Thalassia testudinumwith varying depths. IGARSS '02(pp. 2123–2125 vol.2124).

Han, L., & Rundquist, D. C. (2003). The spectral responses of Ceratophyllum demersum atvarying depths in an experimental tank. International Journal of Remote Sensing, 24,859−864.

Hardisky, M. A., Gross, M. F., & Klemas, V. (1986). Remote sensing of coastal wetlands.BioScience, 36, 453−460.

Hedley, J. D., Harborne, A. R., & Mumby, P. J. (2005). Technical note: Simple and robustremoval of sun glint for mapping shallow-water benthos. International Journal ofRemote Sensing, 26, 2107−2112.

Hedley, J. D., & Mumby, P. D. (2003). A remote sensing method for resolving depth andsubpixel composition of aquatic benthos. Limnology and Oceanography, 48, 480−488.

Hirano, A., Madden, M., & Welch, R. (2003). Hyperspectral image data for mappingwetland vegetation. Wetlands, 23, 436−448.

Hochberg, E. J., Andrefouet, S., & Tyler, M. R. (2003). Sea surface correction of high spatialresolution Ikonos images to improve bottommapping in near-shore environments.Geoscience and Remote Sensing, IEEE Transactions on, 41, 1724−1729.

Holden, H., & LeDrew, E. (2001). Effects of thewater column on hyperspectral reflectanceof submerged coral reef features. Bulletin of Marine Science, 69, 685−699.

Hoogenboom, H. J., Dekker, A. G., & Althuis, I. A. (1998). Simulation of AVIRIS sensitivity fordetecting chlorophyll over coastal and inland waters. Remote Sensing of Environment,65, 333−340.

Huang, Z., Turner, B. J., Dury, S. J., Wallis, I. R., & Foley, W. J. (2004). Estimating foliagenitrogen concentration from HYMAP data using continuum removal analysis. RemoteSensing of Environment, 93, 18−29.

Jassby, A. D., & Cloern, J. E. (2000). Organic matter sources and rehabilitation of theSacramento–San Joaquin Delta (California, USA). Aquatic Conservation: Marine andFreshwater Ecosystems, 10, 323−352.

Kokaly, R. F., Despain, D. G., Clark, R. N., & Livo, K. E. (2003). Mapping vegetation inYellowstone National Park using spectral feature analysis of AVIRIS data. RemoteSensing of Environment, 84, 437−456.

Kruse, F. A., Lefkoff, A. B., Boardman, J. W., Heidebrecht, K. B., Shapiro, A. T., Barloon, P. J., &Goetz, A. F. H. (1993). The spectral image processing system (SIPS)-interactivevisualization and analysis of imaging spectrometer data. Remote Sensing of Environ-ment, 44, 145−163.

Kutser, T., Dekker, A. G., & Skirving, W. (2003). Modeling spectral discrimination of GreatBarrier Reef benthic communities by remote sensing instruments. Limnology andOceanography, 48, 497−510.

Legleiter, C., Roberts, D. A., Marcus, W. A., & Fornstad, M. A. (2004). Passive opticalremote sensing of river channel morphology and in-stream habitat: Physical basisand feasibility. Remote Sensing of Environment, 93, 493−510.

Lillesand, T. M., Kiefer, R. W., & Chipman, J. W. (2004). Remote sensing and image inter-pretation. New York: JohnWiley & Sons, Inc.

Louchard, E. M., Reid, R., Stephens, C., Davis, C., Leathers, R., Downes, T., & Maffione, R.(2002). Derivative analysis of absorption features in hyperspectral remote sensingdata of carbonate sediments. Optics Express, 10, 1573−1584.

Louchard, E. M., Reid, R. P., Stephens, F. C., Davis, C. O., Leathers, R. A., & Downes, T. V.(2003). Optical remote sensing of benthic habitats and bathymetry in coastalenvironments at Lee Stocking Island, Bahamas: A comparative spectral classifica-tion approach. Limnology and Oceanography, 48, 511−521.

Lund, J., Hanak, E., Fleenor, W., Howitt, R., Mount, J., & Moyle, P. (2007). Envisioningfutures for the Sacramento–San Joquin Delta (pp. 324). San Francisco: Public PolicyInstitute of California.

Masuda, K., Takashima, T., & Takayama, Y. (1988). Emissivity of pure and sea waters forthe model sea surface in the infrared window regions. Remote Sensing ofEnvironment, 24, 313−329.

Meng, L., & Moyle, P. B. (1995). Status of splittail in the Sacramento– San JoaquinEstuary. Transactions of the American Fisheries Society, 124, 538−549.

Merenyi, E., Farrand, W. H., Stevens, L. E., Melis, T. S., & Chhibber, K. (2000). Studying thepotential for monitoring Colorado River ecosystem resources below Glen CanyonDam using low-altitude AVIRIS data. 9th AVIRIS Earth science and applicationsworkshop. Pasadena, CA.

Mertes, L. A. K., Smith, M. O., & Adams, J. B. (1993). Estimating suspended sedimentconcentrations in surface waters of the Amazon River wetlands from Landsatimages. Remote Sensing of Environment, 43, 281−301.

Mobley, C. D. (1994). Light and Water: Radiative Transfer in Natural Waters. New York:Academic Press.

Mobley, C. D., Sundman, L. K., Davis, C. O., Bowles, J. H., Downes, T. V., Leathers, R. A.,Montes, M. J., Bissett, W. P., Kohler, D. D. R., Reid, R. P., Louchard, E. M., & Gleason, A.(2005). Interpretation of hyperspectral remote-sensing imagery by spectrummatching and look-up tables. Applied Optics, 44, 3576−3592.

Morel, A., & Belanger, S. (2006). Improved detection of turbid waters from ocean colorsensors information. Remote Sensing of Environment, 102, 237−249.

Mulitsch, M. J. (2005). Remote sensing of California's estuaries: Monitoring climatechange and invasive species. In: Graduate group of ecology. Davis: University ofCalifornia.

Mundt, J. T., Glenn, N. F., Weber, K. T., Prather, T. S., Lass, L. W., & Pettingill, J. (2005).Discrimination of hoary cress and determination of its detection limits via hyper-spectral image processing and accuracy assessment techniques. Remote Sensing ofEnvironment, 96, 509−517.

Nagler, P. L., Daughtry, C. S. T., & Goward, S. N. (2000). Plant litter and soil reflectance.Remote Sensing of Environment, 71, 207−215.

Parsons, W. T., & Cuthbertson, E. G. (1992). Noxious Weeds of Australia Melbourne:Indata Press, pp. 455–456.

Peñuelas, J., Gamon, J. A., Griffin, K. L., & Field, C. B. (1993). Assessing community type,plant biomass, pigment composition, and photosynthetic efficiency of aquaticvegetation from spectral reflectance. Remote Sensing of Environment, 46, 110−118.

Pimentel, D., Lach, L., Zuniga, R., & Morrison, D. (2000). Environmental and economiccosts of nonindigenous species in the United States. BioScience, 50, 53−65.

Pontius, J., Hallett, R., & Martin, M. (2005). Using AVIRIS to assess hemlock abundanceand early decline in the Catskills, New York. Remote Sensing of Environment, 97,163−173.

Renz, M. J., & Blank, R. R. (2004). Influence of perennial pepperweed (Lepidium latifolium)biologyandplant–soil relationships onmanagement and restoration.Weed Technology,18, 1359−1363.

Roberts, D. A., Smith, M. O., & Adams, J. B. (1993). Green vegetation, nonphotosyntheticvegetation, and soils in AVIRIS data. Remote Sensing of Environment, 44, 255−269.

Rueda, C. A., Greenberg, J. A., & Ustin, S. L. (2005). StarSpan: A tool for fast selective pixelextraction from remotely sensed data. Davis, CA: Center for Spatial Technologies andRemote Sensing.

Schlerf, M., Atzberger, C., & Hill, J. (2005). Remote sensing of forest biophysical variablesusing HyMap imaging spectrometer data. Remote Sensing of Environment, 95, 177−194.

Schmidt, K. S., & Skidmore, A. K. (2003). Spectral discrimination of vegetation types in acoastal wetland. Remote Sensing of Environment, 85, 92−108.

Simenstad, C., Toft, J., Higgins,H., Cordell, J., Orr,M.,Williams, P., Grimaldo, L., Hymanson, Z.,& Reed, D. (2000). Sacramento–San Joaquin Delta Breached Levee Wetland Study(BREACH) (pp. 51). Seattle: University of Washington, School of Fisheries.

Sims, D. A., & Gamon, J. A. (2002). Relationships between leaf pigment content andspectral reflectance across a wide range of species, leaf structures and develop-mental stages. Remote Sensing of Environment, 81, 337−354.

Smith, M. O., Ustin, S. L., Adams, J. B., & Gillespie, A. R. (1990). Vegetation in deserts: I. Aregional measure of abundance from multispectral images. Remote Sensing ofEnvironment, 31, 1−26.

Sokal, R. R., & Rohlf, F. J. (2000). Biometry. New York: Freeman and Company.Spenst, R.O. (2006). The biology and ecology of Lepidium latifolium L., in the San

Francisco estuary and their implications for eradication of this invasive weed. In,Graduate Group of Ecology (p. 88 pp.). Davis: University of California, Davis.

Toft, J. D., Simenstad, C. A., Cordell, J. R., & Grimaldo, L. F. (2003). The effects of introducedwater hyacinth on habitat structure, invertebrate assemblages, andfish diets. Estuaries,26, 746−758.

Tompkins, S., Mustard, J. F., Pieters, C. M., & Forsyth, D. W. (1997). Optimization ofendmembers for spectral mixture analysis. Remote Sensing of Environment, 59,472−489.

Tucker, C. J. (1979). Red and photographic infrared linear combinations for monitoringvegetation. Remote Sensing of Environment, 8, 127−150.

Underwood, E., Mulitsch, M., Greenberg, J., Whiting, M., Ustin, S., & Kefauver, S. (2006).Mapping invasive aquatic vegetation in the Sacramento–San Joaquin Delta usinghyperspectral imagery. Environmental Monitoring and Assessment, 121, 47−64.

Underwood, E., Ustin, S., & DiPietro, D. (2003). Mapping nonnative plants using hyper-spectral imagery. Remote Sensing of Environment, 86, 150−161.

Ustin, S. L., Greenberg, J. A., Hestir, E. L., Khanna, S., Santos,M. J., Andrew,M. E.,Whiting,M.,Rajapakse, S., & Lay, M. (2006). Mapping invasive plant species in the Sacramento–SanJoaquin Delta using hyperspectral imagery (pp. 92). Sacramento, CA: Department ofBoating and Waterway: Aquatic Weed Control.

Vahtmäe, E., Kutser, T., Martin, G., & Kotta, J. (2006). Feasibility of hyperspectral remotesensing for mapping benthic macroalgal cover in turbid coastal waters — a BalticSea case study. Remote Sensing of Environment, 101, 342−351.

Viers, J. H., Hogle, I. B., DiPietro, D., Arora, S., Gubaydullin, M., & Quinn, J. F. (2005).Geodatabase application for invasive plant tracking and coordinated habitatrestoration. ESRI international user's conference. San Diego, CA.

Williams, D. J., Rybicki, N. B., Lombana, A. V., O'Brien, T. M., & Gomez, R. B. (2003).Preliminary investigation of submerged aquatic vegetation mapping using hyper-spectral remote sensing. Environmental Monitoring and Assessment, 81, 383−392.

Young, J. A., Palmquist, D. E., & Blank, R. R. (1998). The ecology and control of perennialpepperweed (Lepidium latifolium L.). Weed Technology, 12, 402−405.

Young, J. A., Palmquist, D. E., & Wotring, S. O. (1997). The invasive nature of Lepidiumlatifolium: A review. In J. H. Brock, M. Wade, P. Pysek, & D. Green (Eds.), Plantinvasions: Studies from North America and Europe Leiden: Backhuys.

Young, J. A., Turner, C. E., & James, L. F. (1995). Perennial pepperweed. Rangelands, 17,121−123.