27213549 Spectral and Spatial Detection Limits of Leafy

8/8/2019 27213549 Spectral and Spatial Detection Limits of Leafy

1/129


2/129

SPECTRAL AND SPATIAL DETECTION LIMITS OF LEAFY

SPURGE (EUPHORBIA ESULA L.): SENSOR COMPARISONS AND

MATCHED FILTERED BEHAVIOR

by

Jessica Mitchell

A thesis

submitted in partial fulfillment

of the requirements for the degree of

Master of Science in Geoscience


3/129


4/129

Acknowledgements

First I should like to say that I am most grateful for the encouragement and

understanding of my husband, Kris. Since kindergarten he has been reminding me how

important it is to take breaks and play outside. I would like to extend many special thanks

to my advisor, Dr. Nancy Glenn, who has provided endless guidance, patience, and

opportunity throughout graduate school. What fortune to have both an advisor and a

mentor. I would also like to especially thank Drs. Ames, Van Kirk, and Welhan for their

time, patience, and input. My thanks as well to Melissa Neiers, Michele Hughes, and

Diana Boyack, whose help and friendliness made for many pleasant trips to Pocatello.

This thesis research was funded by USDA Natural Resources Conservation

Service Conservation Innovation Grant No. 68-0211-6-124, Pacific Northwest Regional

Collaboratory, as part of a Pacific Northwest National Laboratory project which is funded

by NASA through Grant No. AGRNNX06AD43G, and NOAA OAR ESRL/ Physical

Sciences Division (PSD) Grant No. NA04OAR4600161. Special thanks to Tom

Stohlgren for kindly accommodating an Idaho delegate in the true spirit of non-native

species forecasting. Field data collection was made possible through generous advice and

assistance from Jeffrey Pettingill and staff at Bonneville County Weed and Pest Control,

Shane Jacobson (U.S. Forest Service, Dubois, Idaho), and Keith Bramwell (Clark

County). Thanks also to the many friends who kept me company or helped me out in the


5/129

Table of Contents

Photocopy Use and Authorization ................................................................................. iTitle Page ...................................................................................................................... iiCommittee Approval Page........................................................................................... iiiAcknowledgements...................................................................................................... ivTable of Contents.......................................................................................................... vList of Figures ............................................................................................................. vii

List of Tables .............................................................................................................. viiThesis Abstract........................................................................................................... viii

Thesis Abstract Idaho State University (2007)....................................................... viiiChapter 1: Introduction and Background...................................................................... 1

1.1 Statement of Purpose .................................................................................... 11.2 Study Area .................................................................................................... 3

1.3 Background ................................................................................................... 61.3.1 Leafy Spurge Invasion Characteristics ........................................................ 61.3.2 Remote Sensing of Vegetation and Sensor Comparison .............................. 81.3.3 Remote Sensing of Leafy Spurge ................................................................ 101.3.4 Mixture-Tuned Matched Filtering Classification ...................................... 13

Chapter 2: Materials and Methods.............................................................................. 172.1 Image Acquisitions ..................................................................................... 17

2.2 Field Data Collection .................................................................................. 182.3 Field Spectroscopy...................................................................................... 222.4 Mixture-Tuned Matched Filtering Classification ....................................... 232.4.1 Minimum Noise Fraction (MNF) ............................................................... 252.4.2 Pixel Purity Index (PPI)............................................................................. 272.4.3 Endmember Selection and n-D Visualizer................................................. 282.4.4 Leafy Spurge Presence / Absence Thresholds ........................................... 332.4.5 Georegistration .......................................................................................... 35

Chapter 3: Manuscript 1.............................................................................................. 37Abstract ................................................................................................................. 373.1 Introduction................................................................................................. 383.1.1 Sensor Comparison .................................................................................... 39

3 2 Previous Work 41


6/129

3.3.7 Accuracy Assessment................................................................................. 583.4 Results......................................................................................................... 60

3.5 Discussion and Conclusions ....................................................................... 643.6 References................................................................................................... 71

Chapter 4: Manuscript 2.............................................................................................. 76Abstract ................................................................................................................. 764.1 Introduction................................................................................................. 774.2 Technical Background ................................................................................ 794.3. Methods....................................................................................................... 82

4.3.1 Data Collection .......................................................................................... 824.3.2 Field Spectroscopy ..................................................................................... 864.3.3 Image Processing....................................................................................... 864.3.4 MF Score Analysis ..................................................................................... 90

4.4 Results......................................................................................................... 914.5 Discussion and Conclusions ....................................................................... 964.6 References................................................................................................... 98

Chapter 5: Conclusion............................................................................................... 1005.1 References................................................................................................. 102

Appendix A: Field Data Collection .......................................................................... 108Appendix B: HyMap Georegistration Error ............................................................. 115Appendix C: Leafy Spurge Presence / Absence Error Matrices............................... 116Appendix D: Preliminary Endmembers and Eigenvalue Plots for Landsat TM5Classifications........................................................................................................... 120


7/129

List of Figures

Figure 1 - Location of study area.................................................................................. 4Figure 2 - Spectral mixing and feasibility triangle ..................................................... 14Figure 3 - Field data collection scheme ...................................................................... 19Figure 4 - Canopy cover estimation techniques.......................................................... 21Figure 5 - Field spectroradiometer measurements of leafy spurge............................. 24Figure 6 - Individual and cumulative MNF band variance......................................... 26Figure 7 - Pixel Purity Index plots, Medicine Lodge ................................................ 29Figure 8 - Pixel Purity Index plot, Spencer................................................................. 30Figure 9 - Potential leafy spurge endmembers, Medicine Lodge ............................... 31Figure 10 - Potential leafy spurge endmembers, Spencer........................................... 32Figure 11 - Scatterplot of Infeasibility values versus MF scores................................ 34Figure 12 - Location of hyperspectral flightlines and reference samples................... 46Figure 13 - HyMap and Landsat processing flow....................................................... 53Figure 14 - Ground reference samples in the Medicine Lodge training area ............. 57

Figure 15 - Spectral profiles of endmembers, MTMF classifications ........................ 59Figure 16 - Comparison of true and spectrally resample HyMap calssifications....... 65Figure 17 - Spectral profiles, true and spectrally resampled HyMap endmembers.... 66Figure 18 - Spencer true HyMap classification .......................................................... 67Figure 19 - Medicine Lodge true HyMap classification............................................. 68Figure 20 - Hypersectral flightlines and reference data, Spencer............................... 84Figure 21 - Hyperspectal flightlines and reference data, Medicine Lodge................. 85Figure 22 - Field spectroradiometer measurements of leafy spurge........................... 87

Figure 23 - Infeasibility versus MF scatterplot calculated over training area ............ 89

List of Tables

Table 1 - Minimum DN thresholds............................................................................. 30Table 2 - Comparison of HyMap and Landsat classification performances............... 61Table 3 - Incremental Cover Evaluations ................................................................... 63

Table 4 - Linear regression summaries, ground cover versus MF scores................... 92Table 5 - MF score image statistics ............................................................................ 92Table 6 - MF scaling behavior summary .................................................................... 94


8/129

SPECTRAL AND SPATIAL DETECTION LIMITS OF LEAFYSPURGE (EUPHORBIA ESULA L.): SENSOR COMPARISONS AND

MATCHED FILTERED BEHAVIOR

Thesis Abstract Idaho State University (2007)

Two demonstration areas in southeast Idaho were used to extend research of remote

sensing of leafy spurge (a noxious weed in Idaho) in two directions: 1) coarse scale

detection for regional distribution mapping and monitoring, and 2) fine scale

detection for abundance mapping using matched filter scores. While a Landsat TM5

image classification indicated that the sensor is inadequate for regional distribution

mapping, a spectrally degraded hyperspectral dataset, with 6 bands similar to those of

the TM5, produced unexpectedly high results. These results suggest that a sensor with

comparable spatial and spectral resolutions but improved instrumentation may offer a

viable alternative for regional mapping of leafy spurge. Examination of the

relationship between matched filter estimates and ground estimates of leafy spurge

cover as well as the scaling behavior of matched filter scores both indicate that

matched filter scores consistently underestimate true abundance. Results also indicate

that matched filter estimates at the high resolution scale (3 m) are not equivalent to

estimates at a coarser scale (up to 25 m) even though abundance measurements in the

field were consistent at these scales. Investigation of MF scores suggests that it is not


9/129

Chapter 1: Introduction and Background

1.1 Statement of Purpose

The increasing rate and extent of harmful biological invasions are recognized

as major components of global environmental change and as mounting, expensive

national problems (Vitousek et al., 1997; Lodge et al., 2005). The overall cost of

biological invasions in the United States, in terms of damage, loss and control, was

estimated at $120 billion per year for approximately 50,000 species (Pimentel et al.,

2005). Locally, the State of Idaho spends at least $7- $10 million per year controlling

invasive species (Northwest Natural Resource Group, 2003).

Invasive plant species can alter ecosystem functions and cause negative

economic impacts in a number of ways, including devaluation of land; reduction of

agricultural productivity and rangeland; loss of native habitat; decline of species

diversity; and alteration of fire regimes and soil dynamics (Olson, 1999). Leafy

spurge (Euphorbia esula L.) is an introduced plant listed as a noxious weed in parts of

Canada and the north central and western United States. Once established, leafy

spurge invasions can spread rapidly, causing particularly serious economic problems

on rangelands, where grazing capacity sharply declines (Hein & Miller, 1992). For

examples, large infestations of leafy spurge in the Dakotas, Montana, and Wyoming


10/129

Identifying new populations of leafy spurge is essential for long-term control

because chemical treatments can eradicate new populations and prevent further

infestation (Anderson et al., 2003). Hyperspectral remote sensors are capable of

detecting new patches of leafy spurge with canopy cover as low as 10%, and can

repeatedly detect leafy spurge with canopy cover of at least 40% in arid climates

(Parker Williams & Hunt, 2002; Glenn et al., 2005). In addition to the early detection

of new infestations, another critical component of long-term leafy spurge

management is identifying and regularly monitoring high risk core infestations at the

regional scale (i.e., 1:100,000). Frequent large scale mapping would enable the

prioritization of containment areas suitable for large-scale treatment applications, the

evaluation of treatment effectiveness, and the discernment of large-scale distribution

and abundance patterns (Parker Williams & Hunt, 2002; Underwood et al., 2007).

This study investigates two components of remote sensing of leafy spurge that

are needed to move toward regional population distribution and abundance

monitoring. Since previous hyperspectral studies of leafy spurge (ONeill et al., 2000;

Root et al., 2002; Parker Williams & Hunt, 2002, 2004; Dudek et al., 2004; Kokaly et

al., 2004; Glenn et al., 2005; Mundt et al., 2007) have established finer scale detection

limits, the first component of this study investigates coarser scale detection limits. To

gain better insight into coarser scale detection limits, the first research paper evaluates

how leafy spurge mapping capabilities vary under different resolution scenarios


11/129

are sufficient for regional distribution mapping and monitoring purposes. The second

research paper addresses matched filtering, which is a spectral mixture analysis

product that in theory estimates sub-pixel target abundance or cover. There are only a

few known studies that have focused on the use of matched filter scores to estimate

vegetation abundance (Parker Williams & Hunt, 2002; Mundt et al., 2007). This study

builds upon previous mixture-tune matched filtering classifications of leafy spurge

(Dudek et al., 2004; Parker Williams & Hunt, 2002, 2004; Glenn et al., 2005) by

addressing the need to determine if match filtering can be reliably used to produce

vegetation abundance maps.

1.2 Study Area

Research was conducted on approximately 7,700 ha of sagebrush steppe on

and in the vicinity of Medicine Lodge (-112 30W, 44 19N), and Spencer (-

112 10W, 44 21N), Idaho, USA (Figure 1). Both sites are located just south of the

Continental Divide, in the Centennial Mountains of Clark County, within 20 km of

the town of Dubois. Riparian and meadow zones associated with Beaver Creek in

Spencer and Medicine Lodge Creek in Medicine Lodge contain willows (Salix);

sedges (Carex); rushes (Juncus); grasses such as spreading bentgrass (Agrostis

stolonifera) and reed canary grass (Phalaris arundinacea); and forbs such as wild iris

(Iris). These bottomlands transition into hillslopes characterized by xeric knolls, rock


12/129

Figure 1. Location of study area, with hyperspectral flightlines shown.

- 4 -


13/129

serviceberry ( Amelanchier alnifolia), snowberry (Symphoricarpos), and chokecherry

(Prunus virginiana).

The Spencer area has a long history of leafy spurge invasion. The weed was

likely first introduced into the towns of Dubois and Spencer by way of the Union

Pacific Railroad, which was built in the late 1800s and located alongside Interstate

15, both of which span the length of the Spencer study site. The railroad and the

highway are two prominent sources of leafy spurge introduction in Spencer, and from

these sources leafy spurge has continued to spread by way of creeks, draws, irrigation

features, and recreational and livestock trails (S. Jacobson & K. Bramwell, pers.

comm.). Large, dense infestations (acres) of leafy spurge occur on the Spencer study

site. Biological treatments began in the early 1990s and two species of Apthona flea

beetles ( A. nigriscutis and A. lacertosa) are believed to have established stable

populations. Sheep have also grazed in enclosures in the Spencer area for the past 6 to

7 years, due to leafy spurge management facilitated by the U.S. Department of

Agriculture (USDA), Dubois Sheep Experiment Station. Chemical treatments within

the Spencer area are concentrated along roads, all-terrain vehicle trails, and the

railroad. Remote patches have also been treated in the past few years and noted in

Geographical Information System (GIS) databases maintained by the Continental

Divide Weed Management Area (CDWMA).

The Medicine Lodge and surrounding drainages (i.e., Rocky Creek, Middle


14/129

site are located in the lower Medicine Lodge and Rocky Creek drainages, where

infestations appear to have been introduced via water, spreading from creeks up

hillslopes by way of draws and drainages. General distribution patterns that were

observed at both the Medicine Lodge and Spencer sites include well-established

colonies of leafy spurge associated with rock outcrops and areas of concentrated

livestock use; and absence or low concentrations of leafy spurge associated with xeric

knolls along hillslopes. Despite similar general distribution patterns, leafy spurge was

present at high densities throughout the Medicine Lodge site, with ground cover

estimates averaging 60%. In contrast, infestations at the Spencer site were

characterized by a single expansive, core infestation, and infrequent occurrences of

low density infestations throughout the remainder of the site. Ground cover estimates

averaged 38% at the Spencer site. Both study sites contained areas where rugged or

remote terrain made access difficult Medicine Lodge more so than Spencer.

1.3 Background

1.3.1 Leafy Spurge Invasion Characteristics

An investigation into the origins and distribution of leafy spurge in North

America concluded that early introductions of leafy spurge along the east coast and

the central Great Plains were likely separate and that in the United States leafy spurge

has spread westward from the central Great Plains (Dunn, 1979, 1985). Leafy spurge


15/129

Leafy spurge grows in dense patches 30 to 80 cm tall. Leafy spurge can form

dense uniform patches, in part, because it exhibits alleopathy, a chemical plant

process that prevents other associated species from growing too close (Steenhagen &

Zindahl, 1979). The plant is identifiable by the milk latex substance it produces and

by terminal, heart-shaped bracts, which are yellow-green and often confused for

flowers when blooming. In Northern climates growth begins around late April and

flowering begins mid-June to early July. Full bloom is followed by a period of

dormancy, then a late season height increase in the fall, when most surrounding

vegetation is dormant (Belcher & Wilson, 1989; Lajeunesse et al., 1999; Whitson et

al., 2002). The species tends to dominate bottomlands; however, it is highly adaptable

and can invade ridges, forest patches, and open mountain slopes from toe-slope to

summit. Populations tend to do well on coarse textured soil, in dry habitats where

competition is limited, and in disturbed areas such as roadsides, abandoned lands,

pastures, rangelands, and recently burned sites (Selleck et al., 1962; Belcher &

Wilson, 1989; Lajeunesse et al., 1999).

The invasion mechanisms and reproductive characteristics of leafy spurge are

such that complete eradication is unlikely. Land managers report that seeds are

persistent and easily dispersed and transported by way of animals, mud, hay, and

water. Reproduction can occur through extensive seed production as well as

vegetative reproduction from both the crown and root buds (Bakke 1936; Bowes &


16/129

become a useful tool for efficiently mapping the distribution of some invasive plant

species, including leafy spurge, over large areas that would otherwise be difficult to

survey (Everitt et al., 1995a; Everitt & Yang, 2004; Lamb & Brown, 2001; Lewis et

al., 2000).

1.3.2 Remote Sensing of Vegetation and Sensor Comparison

Most multispectral remote sensors detect solar reflectance and absorption of

earth materials at a moderate spatial resolution by way of a few broad bands in the

visible and infrared (near, short, and thermal) portions of the electromagnetic

spectrum. Imaging spectrometry, or hyperspectral imaging, is a remote sensing

technology whereby many narrow bands collect surface reflectance information by

over-sampling throughout a near-continuous range of the visible, near infrared, and

short-wave infrared regions of the electromagnetic spectrum (Goetz et al., 1985).

Multispectral sensors such as Landsat Thematic Mapper, ASTER, and SPOT

are satellite-based and provide global coverage at regular time intervals. Standard

multispectral classification techniques have been developed to classify images into

broad categories (Shippert, 2003). Remotely sensed data with spatial resolutions of 15

to 30m pixels have been used for vegetation applications such as land/use land cover

classification, rangeland and forestry monitoring, and land degradation. Such sensors

are capable of differentiating individual invasive species in some cases where the


17/129

areas susceptible to invasion (Johnson, 1999; Morisette et al., 2006, Bradley &

Mustard, 2005).

Hyperspectral sensors such as HyMap and AVIRIS are airborne and cover

relatively small, narrow geographic areas at irregular intervals and with spatial

resolution typically ranging from 3 to 20 m. Compared to multispectral sensors,

hyperspectral sensors have higher spectral, spatial, and radiometric resolution than

multispectral data. Thus, they are more capable of distinguishing subtle spectral

responses among species and improving quantitative model estimations of canopy

structure and chemistry (Ustin et al., 2004). High spectral resolution also facilitates

the use of linear spectral mixture analysis classification techniques that estimate sub-

pixel abundance (Boardman, 1998; Aspinall et al., 2002). High spatial resolution

airborne data increases the probability of detecting smaller infestations, but such

images, when used for repeat monitoring, can present unique challenges in the way of

georegistration and geometric errors (e.g., image rotation, non-uniform pixel shifts;

Aspinall et al., 2002; Glenn et al., 2005). Additional challenges include successfully

requesting and coordinating image acquisition, the need for extensive image

processing techniques, and overall costs. Consequently, high resolution hyperspectral

imagery is less suitable for frequent vegetative monitoring. Satellite-based

hyperspectral imagery has the potential to overcome some of these challenges and has

been collected by the Hyperion sensor (30m pixels) onboard NASAs Earth


18/129

1.3.3 Remote Sensing of Leafy Spurge

During peak phenology, the yellow-green flower bracts of leafy spurge are

spectrally unique and can be distinguished from surrounding vegetation using remote

sensors because of higher reflectance in the visible region (0.5 to 0.7 m) and higher

reflectance values and different shapes in the chlorophyll absorption region (0.55 to

0.69 m; Everitt et al., 1995b; Anderson et al., 1996, 1999;Hunt et al., 2004; Parker

Williams & Hunt 2002, 2004). Chlorophyll and carotenoid concentrations in the

flower bracts explain the hue of the flower bracts, but not reflectance values, which

could be caused by higher fluorescent emission of blue and red wavelengths from the

leaves (Hunt et al., 2004).

Both hyperspectral and multispectral sensors have been used, with varying

degrees of success, to identify leafy spurge (Everitt et al., 1995b; ONeill et al., 2000;

Dudek et al., 2004; Parker Williams & Hunt, 2002, 2004; Hunt & Parker Williams,

2004; Glenn et al., 2005; Stitt et al., 2006). Several hyperspectral detection and

change detection studies of leafy spurge have been conducted in Theodore Roosevelt

National Park (TRNP) as part of a USDA Agricultural Research Station (ARS)

sponsored 5-year research and demonstration program that focused on the control of

leafy spurge (The Ecological Area-wide Management (TEAM) of leafy spurge;

TEAM, 2006; ONeill et al., 2000; Dudek et al, 2004; Kokaly et al., 2004; Root et al.,

2002, 2004; and Stitt et al., 2006). ONeill et al. (2000) applied a number of spectral


19/129

found that minimum noise fraction (MNF) transformed data yielded the most realistic

results, but georegistration errors precluded detailed accuracy assessments. Dudek et

al. (2004) also tested several image classification methods on AVIRIS imagery

obtained over TRNP in 1999 and 2001 for leafy spurge change detection.

Classification methods that were tested included SAM, linear spectral mixture

analysis (Adams et al., 1993), spectral feature fitting (Clark et al., 1990), and mixture-

tuned matched filtering (Boardman, 1998). The authors found that the mixture-tuned

matched filtering algorithm produced the most accurate results, with an overall

accuracy of 39% in 1999 and 47% in 2001. Further, despite high omission errors, in

part due to registration issues, a regional change map of leafy spurge distribution

from 1991 to 2001 was used to identify areas where successful treatments had

reduced leafy spurge populations and other areas where leafy spurge was expanding

and new control efforts would be needed.

In a related study, Kokaly et al. (2004) used a different hyperspectral sensor,

Compact Airborne Spectrographic Imager (CASI), to further investigate a decrease in

leafy spurge cover from 2000 to 2001. Root et al. (2002) used images obtained from

the Hyperion sensor on board the EO-1 satellite to classify leafy spurge. The

Hyperion imagery were acquired in the summer of 2001 over TRNP and the SAM

algorithm was used for classification, resulting in an overall accuracy of 63%. While

the classification method underestimated leafy spurge occurrence it demonstrated


20/129

Imager has seven bands that are spectrally and spatially comparable to Landsat (30m

pixels), as well as three panchromatic bands with a 10m spatial resolution. Stitt et al.

(2006) used Advanced Land Imager (ALI) (with seven bands that are spectrally and

spatially comparable to Landsat and a panchromatic band with a 10m spatial

resolution) and unsupervised classification methods to produce conservative accuracy

assessments in the range of 59% to 66%.

Parker Williams and Hunt (2002, 2004) used AVIRIS imagery and MTMF to

map leafy spurge near Devils Tower National Monument in northeastern Wyoming.

The classifications were 75% to 95% accurate at predicting occurrences of large, high

density leafy spurge infestations. Similar results were found by Glenn et al. (2005),

who found that HyMap imagery (3.5m pixels, 127 bands (0.45 to 2.48 m) were able

to detect leafy spurge cover as low as 10% within a 3.5m pixel and could repeatedly

detect infestations at 40% cover over the same area. Overall classification accuracies

for this study were above 84%.

A limited number of known studies have compared the use of hyperspectral

and multispectral imagery for leafy spurge detection (Hunt & Parker Williams, 2004;

Root et al., 2004; Stitt et al., 2006). Hunt and Parker Williams compared the use of

AVIRIS, Landsat 7, and SPOT for leafy spurge detection in prairie, riparian, and

woodland covertypes. Several vegetation indices and various classification methods

l d d h h d i d h f h h diff f


21/129

was applied to the imagery, overall classification accuracy increased from 68% to

74% for AVIRIS, while overall classification accuracies decreased with Landsat and

SPOT imagery.

Root et al. (2004) evaluated remote sensing of leafy spurge research from

1998 through 2003 and determined that hyperspectral detection of leafy spurge

yielded slightly higher overall classification accuracies (63-78%) than multispectral

classification accuracies (60-70%), although multispectral classification techniques

and accuracy assessment details are unpublished. This study included cost/benefit

analyses of satellite and aircraft-based sensors, which indicated that the use of

multispectral sensors, possibly combined with predictive modeling, may be the most

cost efficient means of mapping leafy spurge infestations at the regional scale.

1.3.4 Mixture-Tuned Matched Filtering Classification

Since hyperspectral remote sensing instruments sample at near-continuous

wavelength intervals, linear mixture analysis methods have been developed to exploit

the high dimensionality of the data. Linear spectral unmixing aims to un-mix pixels

into component materials where the relative area occupied by each material

represents abundance fractions that sum to 1 (Goetz et al., 1985; Boardman, 1995;

Okin et al., 2001; Aspinall, 2002). A mixed pixel may have a combination of

materials that add up to 100%, while a pure pixel ideally contains 100% of a single


22/129

14

Figure 2. Spectral mixing and feasibility triangle (adopted from Boardman, 1995).


23/129

falls within this triangle represents some combination of materials A, B, and C, with

greater degrees of mixing closest to the center of the triangle. Pixels outside of the

triangle represent infeasible pixels that would be considered false positives (e.g.,

artifacts of noise or instrument error).

Mixture-tuned matched filtering (MTMF) is a spectral mixture analysis

technique that has been successfully used in previous studies to identify leafy spurge

in hyperspectral imagery (Parker Williams & Hunt 2002, 2004; Dudek et al., 2004;

Glenn et al., 2005) and is the classification method primarily used in the studies

presented herein. Mixture-tuned matched filtering, as implemented in Environment

for Visualizing Images (ENVI; ITT, 2007), involves a series of four processing steps:

1) a minimum noise fraction transformation (MNF) of the apparent reflectance data to

reduce spectral redundancy, 2) the isolation of spectrally pure pixels from the MNF

bands using a pixel purity indexing (PPI) procedure, 3) the selection of an

endmember pixel from the collection of potentially pure pixels using the n-

dimensional visualizer tool (N-DV), and 4) a MTMF partial unmixing algorithm to

classify the MNF transformed reflectance data (Kruse, 2003). When endmembers are

selected from pixels within the imagery, the MTMF classification is designed to

select one spectrally pure pixel as an endmember to unmix and estimate subpixel

target abundance in each image pixel.

Th MNF f d i ENVI i difi d i f G l


24/129


25/129


26/129

north to Stoddard Creek. Two additional flightlines (1.75 by 10 km each) were

located in the Medicine Lodge area, of which the first was oriented parallel and the

second perpendicular to the Medicine Lodge Creek drainage. The perpendicular

orientation captured portions of the plateaus and canyons formed by Rocky Creek,

Middle Creek, and Indian Creek drainages (Figure 1).

The HyMap instrument collects calibrated radiance data in 126 near-

contiguous spectral bands (0.45 to 2.48 m) that range in width from 15 m in the

visible and near infrared to 20 m in the shortwave infrared (Kruse et al., 2000). For

comparative purposes, a single Landsat 5 image was acquired over the study area on

June 13, 2006 (path 39, row 29). The Thematic Mapper (TM) on board the Landsat 5

satellite collects data in 7 relatively wide bands: Band 1 (blue, 0.45 -0.52 m), Band 2

(green, 0.52 - 0.60 m), Band 3 (red, 0.63-0.69 m), Band 4 (near-infrared, 0.76-0.90

m), Band 5 (mid-infrared, 1.55-1.75m), Band 6 (thermal infrared, 10.4-12.5m),

and Band 7 (mid-infrared, 2.08-2.35m). The thermal band has a spatial resolution of

120 x 120 m and the other 6 bands have a spatial resolution of 30 x30 m.

2.2 Field Data Collection

Beyond North America Weed Management Association (NAWMA) mapping

standards were used as a guide for field data collection (Stohlgren et al., 2005). The

sample design used a 7.32-m (24-ft) radius circle with 3 transects extending from the


27/129

Figure 3. Modified beyond NAWMA field data collection scheme (adopted fromStohlgren et al., 2005).

quadrats, each with an area of 1m2, were positioned to the right of the transects, at

intervals of 1.8 m (6 ft), 3.7 m (12 ft), and 5.5 m (18 ft) from the plot center. The

structure of the sampling plot is a slightly modified version of the Beyond NAWMA

plot in that nine quadrats were used instead of three in order to improve the accuracy

of abundance estimations. To calibrate ocular estimates of leafy spurge percent

canopy cover across a continuous interval, estimates for the first five plots included


28/129

methods, the initial estimate at the plot scale and the average estimations using a

point frame produced the closest results only one of the five calibration plots varied

by more than one percent. On the other hand, estimations at the five calibration plots

using the Daubenmire quadrat were consistently about 20% lower than estimates

obtained using the other two methods. A single observer proceeded to estimate cover

at plots with either high or low percent leafy spurge cover before moving on to plots

with leafy spurge cover in the mid-range. For presence plots, field data collection

typically included measurements and estimations at both the plot and quadrat level.

Initial ocular cover estimates of leafy spurge were made at the plot scale, then the

Daubenmire frame was also used to estimate leafy spurge percent canopy cover to the

nearest percent at each of the nine quadrats. Similar percent canopy cover estimations

were made for shrub, bare ground, and rock. Although the point frame estimation

technique was designed for sagebrush steppe ecosystems and is regarded as a more

objective estimation method than visual cover estimation (Bonham, 1989), the

Daubenmire frame was chosen as the cover estimation method. The Daubenmire

frame provides for ease of use and speed, and is more effective at locating rare

species (Meese & Tomich, 1992; Dethier et al., 1993), which was a field data

collection criterion. In the final analysis, regression plots (Appendix A) indicated that

there was strong agreement between the ocular cover estimation techniques at the plot

and quadrat level for both leafy spurge (r2 = 0.76) and shrub (r2 = 0.82). These


29/129

a


30/129

In addition to cover estimations, leafy spurge height was measured, individual

species within the plots were identified, and the total number of species recorded. A

walking species search was also conducted to identify any new plant species that

were not previously identified during the quadrat sampling. Additional ancillary

information was also collected, including slope, aspect, elevation, distance from road,

and distance from water. For absence plots, ocular estimates of shrub and bare ground

cover were typically recorded at the plot level. Individual species were also identified,

and the total number of species recorded. A sample field data collection form is also

provided in Appendix A,as well as a summary of field data collection results.

Sampling was initiated at the Spencer site on June 16, 2006, a few days prior

to full bloom and continued during and shortly after peak phenology, ending July 26,

2006. A total of 56 plots, 43 present and 13 absent, were sampled in Spencer.

Validation samples were collected in Medicine Lodge from July 26 to August 13,

2006, after peak phenology. A total of 55 plots, 43 present and 12 absent, were

sampled in Medicine Lodge.

2.3 Field Spectroscopy

To assess the spectral characteristics of leafy spurge ground cover data at

varying percent covers, a field spectroradiometer (Analytical Spectral Device (ASD),

Boulder CO) was used to measure the spectral signatures of leafy spurge at three


31/129

The instrument was calibrated prior to measurements at each location using dark and

white reflectance panels. A series of 15 readings were collected for each infestation

and representative signatures were selected for comparison (Figure 5). These field

measurements indicate that the magnitude of reflectance values is directly related to

the density of the infestation. The spectral signatures were collected two days after

image acquisition (June 30, 2006), at the same time of day that the imagery was

acquired, and under similar atmospheric conditions. Errors in the ASD prevented the

collection of spectral data concurrent with image acquisition.

2.4 Mixture-Tuned Matched Filtering Classification

The hyperspectral imagery that was acquired over the study area by HyVista

was delivered with radiometric and geometric corrections. Radiance values were

converted by the vendor to apparent reflectance using the HyCorr absolute

atmospheric correction modeling package. After a preliminary evaluation of the data,

bands 1 (0.4538 um), 63 (1.3893 um), 64 (1.4042 um), and 126 (2.4963 um) were

removed from both mosaics due to the obvious influence of noise and water

absorption. Similarly, the 2 hyperspectral flightlines collected over the Medicine

Lodge study site were processed as a single georeferenced mosaic and the 3

flightlines collected over the Spencer study site were processed as a single

georeferenced mosaic. Masks were then applied to the mosaic backgrounds to


32/129

-20

980

1980

2980

3980

4980

0.35 0.55 0.75 0.95 1.15 1.35 1.56 1.76 2.05 2.25 2.45

Wavelength (microns)

Re

flectance(x10,000)

98% leafy spurge cover



Figure 5. Field spectroradiometer measurements of leafy spurge at locations with 34%, 63%, and 98% cover. Atmosphericwindows excluded for clarity.

24


33/129

of the Medicine Lodge and Spencer study areas. The two study sites were classified

independently, as the use of a single endmember from the Medicine Lodge training

area, when applied to both sites, produced unrealistic classification results for the

Spencer site. In addition, nonparametric Kolmogorov-Smirnov (P = 0.037) and

Mann-Whitney (z-score = 2.832) tests were used to compare the distribution shapes

and population statistics for leafy spurge cover data collected at the Medicine Lodge

and Spencer sites. Both tests concluded that the Medicine Lodge and Spencer sites

were not statistically similar at a 95% confidence level.

2.4.1 Minimum Noise Fraction (MNF)

As mentioned previously (Section 1.3.4), the first processing step in the

mixture-tuned matched filter (MTMF) classification is to transform the apparent

reflectance bands into Minimum Noise Fraction (MNF) bands to reduce spectral

redundancy and noise in the data sets. To segregate noise from the data, cumulative

eigenvalue plots were used to partition MNF transformed bands associated with

larger eigenvalues and greater coherence from MNF bands dominated by noise

(Kruse, 2003). For the Medicine Lodge mosaic, the cumulative band variance plot

determined that the first 30 MNF bands explained 82% of the data (Figure 6a). Given

that visual inspection of the bands also demonstrated noticeable degradation after

band 30, the first 30 MNF bands were selected as input for pixel purity indexing the


34/129

Medicine Lodge

0

10

20

30

40

50

60

70

80

90

100

0 20 40 60 80 100 120

MNF Band

Variance(%)

a..

Spencer

0

10

20

30

40

50

60

70

80

90

100

0 20 40 60 80 100 120

Variance(%)

b..


35/129

noticeable degradation after band 40, the first 40 MNF bands were selected as input

for pixel purity indexing.

2.4.2 Pixel Purity Index (PPI)

Pixel purity indexing identifies the most spectrally pure pixels in an image in

an effort to narrow the selection of a single leafy spurge endmember pixel to use for

the MTMF classification. Potentially pure pixels are identified by randomly

generating scatter plots of MNF bands of varying dimensions (total of 30 in the case

of Medicine Lodge, and 40 in the case of Spencer) for a specified number of

iterations, at a specified threshold. The threshold is a factor of the variance within the

data set, where a greater threshold equates to a greater chance of including mixed

pixels. A threshold factor of 2 or 3 is recommended for MNF data (ITT, 2007). For

both mosaics, 20,000 iterations were selected, with a threshold of 3.5. When PPI

processing is complete, a plot of the number of potentially pure pixels found at each

iteration is generated. This plot typically resembles a step function that asymptotically

approaches a flat line when all potentially pure pixels are found (ITT, 2007). For the

Medicine Lodge study site, after iteration 17,751, the total pixel count jumped from

6,667 to 53,491, and by iteration 17, 776, the total pixel count is more than 1, 200,000

(Figure 7). If more than 17,754 iterations were specified, the plot changes from a

convex curve to a step function. Potential spectrally pure pixels that were found after


36/129

DN value must be specified, where the DN value in the PPI image is a brightness

value. Dark values (low DN) represent spectrally un-pure pixels and bright values

(high DN) represent spectrally pure pixels (ITT, 2007). A minimum DN threshold

value of 2 was used for the Medicine Lodge study site and a threshold DN of 1 was

used for the Spencer study site (Table 1). These thresholds are considered a good

trade-off between a group of too few pixels and a group of too many pixels from

which an endmember is selected.

2.4.3 Endmember Selection and n-D Visualizer

In the n-D Visualizer (N-DV), MNF pixel data are viewed on scatterplots that

can be rotated in MNF band space. The MNF bands are the plot axes, where the total

number of MNF bands equates to the total number of dimensions that can be

visualized. Purer pixels occur in clusters at the corners of the pixel cloud and can be

interactively selected, viewed, and compared to field data and reflectance image

locations (ITT, 2007). The N-DV tool was used to interactively explore different

classes within the Medicine Lodge mosaic (e.g., roads/high reflectance; rock;

saturated areas; soil/bare ground; and leafy spurge with canopy cover of

approximately 60% or more). Next, potentially pure pixels that geographically

coincided with areas of high percent leafy spurge cover on the ground (training areas)

were selected as potential endmembers for classifying the imagery. For each study


37/129

a..

b..


38/129

Figure 8. Pixel purity index plot for Spencer site at 10,000 iterations with a thresholdof 3.5.

Med. Lodge Spencer

DN value # Pure Pixels # Pure Pixels

>29 275 107

>25 311 116

>20 386 125

>15 493 140

>10 685 160

>5 1204 206

>2 2915 351

>1 11,713,152 749

>0 15,824,187 15,306,784


39/129

0

1000

2000

3000

4000

5000

6000

0.45 0.95 1.45 1.95 2.45


PercentReflectance(x10,0

00)

Figure 9. Potential leafy spurge endmembers for the Medicine Lodge study site. The spectral profile of the pixelselected to use for classification is depicted as ( ).

31


40/129

Figure 10. Potential leafy spurge endmembers for the Spencer study site. The spectral profile of the pixel selected to use forclassification is depicted as ( )

0

1000

2000

3000

4000

5000

6000

0.45 0.95 1.45 1.95 2.45


Reflectance(x10,0

00)

32


41/129


42/129

Figure 11.Scatterplot of Infeasibility values versus MF scores for Medicine Lodge. The shadedarea represents leafy spurge presence.

34


43/129

were then mathematically combined in ArcMap GIS such that values below the

threshold equated to presence and values above the threshold equated to absence.

2.4.5 Georegistration

Map classification accuracy can be influenced by radiometric and geometric

errors introduced during the data acquisition process (Jensen, 2005). The HyMap

sensor is typically calibrated in the lab prior to image acquisition to minimize the

influence of incorrect radiance measurements (Cocks et al., 1998). Atmospheric and

geometric corrections applied to the imagery during vendor prep-processing are

designed to further minimize introduced error. However, residual random geometric

error from sources such as the roll, yaw, and tilt of the aircraft, can influence

locational accuracy and should be evaluated.

To assess HyMap georegistration error, 10 differentially corrected GPS

ground control points were collected in the central Spencer flightline and nine in the

North-South Medicine Lodge flightline. Ground control points were collected for

these two flightlines because they contained the majority of ground reference

samples. Directional shifts occurred non-uniformly, with minimum errors of 0.29m

and 0.38m, maximum errors of 6.11m and 1.57m, and mean errors of 3.39m and

0.813m for the Medicine Lodge and Spencer flightlines respectively (Appendix B).

The validation plots (7.32m radius) were treated as polygons (168.25m2), and for the


44/129

Error matrices that compare classified map polygons or pixels to ground

reference samples are typically used in remote sensing to assess map accuracy and

calculate different error metrics, such as Users accuracy, Producers accuracy, and

overall accuracy (Congalton & Green, 1999; Congalton, 2004; Foody, 2004). The

Users accuracy represents the percentage of pixels that are correctly classified on the

ground and the Producers accuracy represents the percentage of a given class (e.g.,

absence or presence) that is correctly identified on the map. The overall accuracy is a

ratio of the number of correctly classified samples to the total number of samples. An

example error matrix with calculations for Users, Producers and overall accuracy is

presented in Appendix C, along with accuracy assessment results for all

classifications that were conducted in this study. The error matrix also characterizes

error in terms of errors of commission (false positives) and errors of omission (false

negatives). Errors of commission are samples that, in this study, did not contain leafy

spurge on the ground, but were classified as leafy spurge on the map. Likewise, errors

of omission are samples that contained leafy spurge on the ground, but were omitted

on the classified map.


45/129

Chapter 3: Manuscript 1

Leafy spurge (Euphorbia esula L.) classification performance usinghyperspectral (HyMap) and multispectral (Landsat Thematic Mapper 5)sensors

Abstract

Two demonstration sites in southeast Idaho were used to extend the scope of remote

sensing of leafy spurge research by shifting the focus from establishing finer scale

detection limits using high spectral and/or spatial resolution sensors toward

investigating coarser scale detection limits. Hyperspectral images were obtained to

produce baseline leafy spurge maps, from which spatially and/or spectrally degraded

images were subsequently derived for comparative purposes with Landsat TM5. The

baseline presence / absence maps had an overall accuracy of 67% at the Spencer

study site, and 85% at the Medicine Lodge study site. Unexpectedly high accuracy

results were produced from the images that were spectrally degraded to the

bandwidths of Landsat TM5, which suggests that high spectral resolution is not

critical to leafy spurge detection. However, a classification using a true Landsat TM5

image indicates that the sensor is inadequate for regional distribution monitoring. The

differences in results between the true and simulated images suggest that a sensor

with comparable resolutions but improved instrumentation (e.g. signal to noise) may


46/129

3.1 Introduction

The increasing rate and extent of harmful biological invasions are recognized

as major components of global environmental change and as mounting, expensive

national problems (Vitousek et al., 1997; Lodge et al., 2005). The overall cost of

biological invasions in the United States, in terms of damage, loss and control, is

estimated at $120 billion per year for approximately 50,000 species (Pimentel et al.,

2005). Invasive plant species can alter ecosystem functions and cause negative

economic impacts in a number of ways, including devaluation of land; reduction of

agricultural productivity and rangeland; loss of native habitat; decline of species

diversity; and alteration of fire regimes and soil dynamics (Olson, 1999). Leafy

spurge (Euphorbia esula L.) is an introduced plant listed as a noxious weed in parts of

Canada and the north central and western United States. Once established, leafy

spurge invasions can spread rapidly, causing particularly serious economic problems

on rangelands, where grazing capacity sharply declines (Hein & Miller, 1992). Leafy

spurge now infests approximately 2 million hectares of rangeland, pastures, hillsides

and riparian areas in North America, where the size of infested areas has been

doubling nearly every 10 years (Quimby & Wendel, 1997).

The invasion mechanisms and reproductive characteristics of leafy spurge are

such that complete eradication is unlikely. Land managers report that seeds are


47/129

widespread and cost-effective tools are needed to monitor changes in leafy spurge

distribution and abundance over time (Anderson et al., 2003). Remote sensing has

become a useful tool for efficiently mapping the distribution of some invasive plant

species, including leafy spurge, over large areas that would otherwise be difficult to

survey (Everitt et al., 1995b; Everitt & Yang, 2004; Lamb & Brown, 2001; Lewis et

al., 2000).

3.1.1 Sensor Comparison

Most multispectral remote sensors detect solar reflectance and absorption of

earth materials at a moderate spatial resolution by way of a few broad bands in the

visible and infrared (near, short, and thermal) portions of the electromagnetic

spectrum. Imaging spectrometry, or hyperspectral imaging, is a remote sensing

technology whereby many narrow bands collect surface reflectance information

throughout a near-contiguous range of the visible, near infrared, and short-wave

infrared portions of the electromagnetic spectrum (Goetz et al., 1985).

Multispectral sensors such as Landsat Thematic Mapper, ASTER, and SPOT

are satellite-based and provide global coverage at regular time intervals. Standard

multispectral classification techniques have been developed to classify images into

broad categories (Shippert, 2003). Remotely sensed data with spatial resolutions of 15

to 30m pixels have been used for vegetation applications such as land/use land cover


48/129

MODIS) may also improve the ability of predictive models to identify areas

susceptible to invasion (Johnson, 1999; Morisette et al., 2006, Bradley & Mustard,

2005).

Hyperspectral sensors such as HyMap and AVIRIS are airborne and cover

relatively small, narrow geographic areas at irregular time intervals and with spatial

resolution typically ranging from 3 to 20 m. Compared to multispectral sensors,

hyperspectral sensors have higher spectral, spatial, and radiometric resolution than

multispectral data. Thus, they are more capable of distinguishing subtle spectral

responses among species and improving quantitative model estimations of canopy

structure and chemistry (Ustin et al., 2004). High spectral resolution facilitates the use

of linear spectral mixture analysis classification techniques that estimate sub-pixel

abundance (Boardman, 1998; Aspinall et al., 2002). High spatial resolution airborne

data increases the probability of detecting smaller infestations, but such images, when

used for repeat monitoring, can present unique challenges in the way of

georegistration and geometric errors (e.g., image rotation, non-uniform pixel shifts;

Aspinall et al., 2002; Glenn et al., 2005). Additional challenges include successfully

requesting and coordinating image acquisition, the need for extensive image

processing techniques, and overall costs. As such, high resolution hyperspectral

imagery is less suitable for frequent vegetative monitoring. Satellite-based

hyperspectral imagery has the potential to overcome some of these challenges and has


49/129

3.2 Previous Work

During peak phenology, the yellow-green flower bracts of leafy spurge are

spectrally unique and can be distinguished from surrounding vegetation using remote

sensors because of higher reflectance in the visible region (0.5 to 0.7 m) and higher

reflectance values and different shapes in the chlorophyll absorption region (0.55 to

0.69 m; Everitt et al., 1995b; Anderson et al., 1996, 1999;Hunt et al., 2004; Parker

Williams & Hunt 2002, 2004). Chlorophyll and carotenoid concentrations in the

flower bracts explain the hue of the flower bracts, but not reflectance values, which

could be caused by higher fluorescent emission of blue and red wavelengths from the

leaves (Hunt et al., 2004).

Both hyperspectral and multispectral sensors have been used, with varying

degrees of success, to identify leafy spurge (Everitt et al., 1995b; ONeill et al., 2000;

Dudek et al., 2004; Parker Williams & Hunt, 2002, 2004; Hunt & Parker Williams,

2004; Glenn et al., 2005; Stitt et al., 2006). Several hyperspectral detection and

change detection studies of leafy spurge have been conducted in Theodore Roosevelt

National Park (TRNP) as part of a United States Department of Agriculture

Agricultural Research Station (USDA-ARS) sponsored 5-year research and

demonstration program that focused on the control of leafy spurge (The Ecological

Area-wide Management (TEAM) of leafy spurge; ONeill et al., 2000; Dudek et al,


50/129

minimum noise fraction (MNF) transformed data yielded the most realistic results,

but georegistration errors precluded detailed accuracy assessments. Using the same

dataset, Dudek et al. (2004) tested several image classification methods, including

SAM, linear spectral mixture analysis (Adams et al., 1993), spectral feature fitting

(Clark et al., 1990), and mixture-tuned matched filtering (MTMF; Boardman, 1998).

The authors found that the MTMF algorithm produced the most accurate results, with

an overall accuracy of 39% in 1999 and 47% in 2001. Further, despite high omission

errors, in part due to registration issues, a regional change map of leafy spurge

distribution from 1991 to 2001 was used to identify areas where successful treatments

had reduced leafy spurge populations and other areas where leafy spurge was

expanding and new control efforts would be needed.

In a related study, Kokaly et al. (2004) used a different hyperspectral sensor,

Compact Airborne Spectrographic Imager (CASI), to further investigate a decrease in

leafy spurge cover from 2000 to 2001. Root et al. (2002) used images obtained from

the Hyperion sensor on board the EO-1 satellite to classify leafy spurge. The

Hyperion imagery were acquired in the summer of 2001 over TRNP and the SAM

algorithm was used for classification, resulting in an overall accuracy of 63%. While

the classification method underestimated leafy spurge occurrence, it demonstrated

that the target could be discriminated when mixed with up to 65% of other vegetation

types The Advanced Land Imager (ALI) sensor also on board EO-1 acquired

(2006) d Ad d L d I (ALI) ( i h b d h ll d


51/129

(2006) used Advanced Land Imager (ALI) (with seven bands that are spectrally and

spatially comparable to Landsat and a panchromatic band with a 10m spatial

resolution) and unsupervised classification methods to produce conservative accuracy

assessments in the range of 59% to 66%.

Parker Williams and Hunt (2002, 2004) used AVIRIS imagery and MTMF to

map leafy spurge near Devils Tower National Monument in northeastern Wyoming.

The classifications were 75% to 95% accurate at predicting occurrences of large, high

density leafy spurge infestations. Similar results were found by Glenn et al. (2005),

who found that HyMap imagery (3.5m pixels, 127 bands (0.45 to 2.48 m) were able

to detect leafy spurge cover as low as 10% within a 3.5m pixel and could repeatedly

detect infestations at 40% cover over the same area. Overall classification accuracies

for this study were above 84%.

A limited number of known studies have compared the use of hyperspectral

and multispectral imagery for leafy spurge detection (Hunt & Parker Williams, 2004;

Root et al., 2004; Stitt et al., 2006). Hunt and Parker Williams compared the use of

AVIRIS, Landsat 7, and SPOT for leafy spurge detection in prairie, riparian, and

woodland covertypes. Several vegetation indices and various classification methods

were evaluated and the authors determined that, for the three different types of

imagery, a green: red reflectance ratio correlated best with cover measurements and

h b l i l i d l ifi i i di ( ll


52/129

spurge infestations (both current and historic) and monitoring regional distribution


53/129

spurge infestations (both current and historic) and monitoring regional distribution

and abundance patterns.

3.3 Methods

3.3.1 Study site

Research was conducted on approximately 7,700 ha of sagebrush steppe on

and in the vicinity of Medicine Lodge (-112 30W, 44 19N), and Spencer (-

112 10W, 44 21N), Idaho, USA (Figure 12). Both sites are located just south of the

Continental Divide, in the Centennial Mountains of Clark County, within 20 km of

the town of Dubois. Riparian and meadow zones associated with Beaver Creek in

Spencer and Medicine Lodge Creek in Medicine Lodge contain willows (Salix);

sedges (Carex); rushes (Juncus); grasses such as spreading bentgrass (Agrostis

stolonifera) and reed canary grass (Phalaris arundinacea); and forbs such as wild iris

(Iris). These bottomlands transition into hillslopes characterized by xeric knolls, rock

outcrops, and mesic draws. Sagebrush (Artemisia), bitterbrush (Purshia), rabbitbrush

(Chrysothamnus), and a variety of grasses (Elymus, Poa, Phleum, Bromus), and forbs

( Lupinus, Comandra, Achillea, Balsamorhiza, Aster, Potentilla) occur on drier sites,

while drainages are characterized by aspen (Populus tremuloides) and shrubs such as

serviceberry ( Amelanchier alnifolia), snowberry (Symphoricarpos), and chokecherry

(Prunus virginiana).


54/129

Figure 12. Location of hyperspectral flightlines and reference samples.

46

Pacific Railroad, which was built in the late 1800s and located alongside Interstate


55/129

Pacific Railroad, which was built in the late 1800 s and located alongside Interstate

15, both of which span the length of the Spencer study site. The railroad and the

highway are two prominent sources of leafy spurge introduction in Spencer, and from

these sources leafy spurge has continued to spread by way of creeks, draws, irrigation

features, and recreational and livestock trails (S. Jacobson & K. Bramwell, pers.

comm.). Dense infestations of leafy spurge as large as 0.75km2 occur in the Spencer

area. Biological treatments were begun in the early 1990s and two species of Apthona

flea beetles (A. nigriscutis and A. lacertosa) show evidence of stable population

levels. Sheep have also been grazed on enclosures in the Spencer area for the past 6

to7 years, which has been facilitated by Dubois Sheep Experiment Station

involvement. Chemical treatments within the Spencer area are concentrated along

roads, all-terrain vehicle trails, and the railroad. Remote patches have also been

treated in the past few years and noted in Geographical Information System (GIS)

databases maintained by the Continental Divide Weed Management Area (CDWMA).

The Medicine Lodge and surrounding drainages (i.e., Rocky Creek, Middle

Creek, and Indian Creek) have a somewhat shorter invasion history than Spencer

because the railroad is further away, although Medicine Lodge infestations were

exacerbated by fire in 2003. The heavier concentrations of leafy spurge are located in

the lower Medicine Lodge and Rocky Creek drainages, where infestations appear to

have been introduced via water spreading from creeks up hillslopes by way of draws

low concentrations of leafy spurge associated with xeric knolls along hillslopes.


56/129

y p g g p

Despite similar general distribution patterns, leafy spurge is present at high densities

throughout the Medicine Lodge site, with ground cover estimates averaging 60%

(described below). In contrast, infestations at the Spencer site are characterized by a

single expansive, core infestation, and infrequent occurrences of low density

infestations throughout the remainder of the site. Ground cover estimates averaged

38% at the Spencer site. Both study sites contained areas where rugged or remote

terrain made access difficult Medicine Lodge more so than Spencer.

3.3.2 Image acquisition

In this study, the high spectral and spatial resolutions of hyperspectral imagery

were necessary for obtaining baseline data, from which spectrally and/or spectrally

degraded images could be derived for comparative purposes. The relatively coarser

spectral and spatial resolutions of Landsat imagery were necessary to assess the

validity of the degraded image results and the suitability of using widely available

multispectral satellite imagery for regional distribution mapping.

Hyperspectral imagery were collected over the study area using the HyMap

sensor (operated by HyVista, Inc.) mounted on an aircraft flying about 1000m above

the ground to obtain 3.2 by 3.2 m pixel resolution. The HyMap sensor collected 5

flightlines of data on June 28, 2006, which was an optimal date for capturing leafy


57/129

multispectral scales the plots are treated as polygons at the hyperspectral scale and


58/129

as individual pixels at the multispectral scale.A Trimble GeoXT model GPS receiver

(Trimble Navigation Limited, Westminster CO) was used to collect geographic

locations of plots and infestation boundaries, which were then differentially corrected.

Sampling was initiated at the Spencer site on June 16, 2006, a few days prior to full

bloom and continued during and shortly after peak phenology, ending July 26, 2006.

A total of 56 plots, 43 with leafy spurge present, and 13 with leafy spurge absent were

sampled in Spencer. Validation samples were collected in Medicine Lodge from July

26 to August 13, 2006, after peak phenology. A total of 55 plots, 43 with leafy spurge

present, and 12 with leafy spurge absent were sampled in Medicine Lodge.

3.3.4 Image Pre-processing

All image pre-processing and processing, unless otherwise stated, was

performed using the Environment for Visualizing Images (ENVI) version 4.3

software (ITT Visual Information Solutions, Boulder, CO). Hyperspectral radiance

values were converted by the vendor to apparent reflectance using the HyCorr

absolute atmospheric correction modeling package. The multispectral imagery was

converted to apparent reflectance with ENVIs atmospheric correction package

FLAASH. The absolute atmospheric corrections produced scaled surface reflectance

values that account for scattering and absorption of solar radiation by the earths


59/129

um), and 126 (2.4963 um) were removed from both mosaics due to the obvious


60/129

influence of noise and water absorption. Masks were applied to the mosaic

backgrounds to minimize the influence of the high number of background pixels on

image classification results. The two study sites were classified independently, as the

use of a single endmember from the Medicine Lodge training area, when applied to

both sites, produced unrealistic classification results for the Spencer site. In addition,

nonparametric Kolmogorov-Smirnov (P = 0.037) and Mann-Whitney (z-score =

2.832) tests were used to compare the distribution shapes and population statistics for

leafy spurge cover data collected at the Medicine Lodge and Spencer sites. Both tests

concluded that the ground reference data sampled at the Medicine Lodge and Spencer

sites were not statistically similar at a 95% confidence level.

To explore changes in leafy spurge detection performance at coarsened

spectral and spatial resolutions, different combinations of spectral and spatial

resampling were applied to the Medicine Lodge HyMap mosaic (121 bands, 3.2m

pixels) to produce 3 additional images: a spatially resampled image, a spectrally

resampled image, and a spectrally and spatially resampled image (Figure 13). The

spatially resampled image was generated using a nearest neighbor resampling method

to simulate the Landsat TM5 spatial resolution (28.5m pixels) while retaining

hyperspectral resolution (121 bands). The spectrally resampled image was generated

using a filter function to simulate the 6 relatively wide non-thermal bands of the


61/129

Figure 13. Flow diagram depicting classification methods and spatial and spectral resampling processing parameters for theMedicine Lodge hyperspectral imagery (a), the Spencer hyperspectral imagery (b), and the Landsat TM5 imagery (c).

53

c.

SPENCER

True HyMap

121 Bands, 3.3m pixels

MEDICINE LODGETrue HyMap


MTMF Classification

MTMF Classification

True Landsat6 Bands

30m pixels

True Landsat6 Bands

30m pixels

True Landsat6 Bands

30m pixels

MTMF Classificationtuned to

Medicine Lodge Area

SAM Classification

Entire Landsat Scene


Spencer Area

SPATIALRESAMPLE

SPATIAL&

SPECTRAL

RESAMPLE

SPECTRALRESAMPLE

HyMap

Spatially

Resampled

121 Bands

28.5m pixels

HyMap

6 Bands

28.5m pixels

SIMULATED

LANDSAT TM

HyMap

Spectrally

Resampled

6 Bands

3.2m pixels

MTMF Classifications

SPENCER

True HyMap


MEDICINE LODGETrue HyMap


MTMF Classification

SAM Classification

Entire Landsat Scene


Medicine Lodge Area


Spencer Area

True Landsat6 Bands

30m pixels

True Landsat6 Bands

30m pixels

True Landsat6 Bands

30m pixels

HyMap

Spectrally

Resampled

6 Bands

3.2m pixels

SPATIALRESAMPLE

SPATIAL&

SPECTRAL

RESAMPLE

SPECTRALRESAMPLE

MTMF Classification

HyMap

Spatially

Resampled

121 Bands

28.5m pixels

HyMap

6 Bands

28.5m pixels

SIMULATED

LANDSAT TM

MTMF Classifications

b.

a.

28.5m. A similar approach was used in a related study to explore the detection


62/129

potential of 3 invasive species within the context of 6 vegetation communities

(Underwood et al., 2007). The Medicine Lodge data was chosen over the Spencer

data to be resampled because preliminary results indicated that it performed better at

detecting leafy spurge.

3.3.6 Image Classification

Mixture-tuned matched filtering (MTMF) is a spectral mixture analysis

technique that has been successfully used in previous studies to identify leafy spurge

in hyperspectral imagery (Parker Williams & Hunt, 2002, 2004; Dudek et al., 2004;

Glenn et al., 2005, Mundt et al., 2007). Mixture-tuned matched filtering, as

implemented in ENVI, involves a series of 4 processing steps: 1) a minimum noise

fraction transformation (MNF) of the apparent reflectance data to reduce spectral

redundancy, 2) the isolation of spectrally pure pixels from the MNF bands using a

pixel purity indexing (PPI) procedure, 3) the selection of an endmember pixel from

the collection of potentially pure pixels using the n-dimensional visualizer tool (N-

DV), and 4) a MTMF partial unmixing algorithm to classify the MNF transformed

reflectance data (Boardman, 1998). A mixed pixel may have a combination of

materials that add up to 100%, while a pure pixel ideally contains 100% of a single

material. When endmembers are selected from pixels within the imagery, the MTMF

abundance, where a value of 1 indicates a perfect match, and pixels below the


63/129

background distribution indicate no match. Infeasibility values provide estimates of

how closely the pixels approximate the endmember pixel, or the likelihood that the

classified pixel is a false positive. A high infeasibility value indicates a high

likelihood of misclassification (ITT, 2007). False positives can be reduced by

interactively selecting an infeasibility threshold such that pixels with low MF scores

must have low infeasibility scores and pixels with higher MF scores are allowed

higher infeasibility values. For weed management purposes, and the purposes of this

study, over-prediction is favored, as there is a greater risk associated with overlooking

a new infestation than with falsely identifying leafy spurge presence.

Mixture-tuned matched filter classifications were applied to the Spencer true

HyMap mosaic, the Medicine Lodge true HyMap mosaic, the 3 resampled Medicine

Lodge HyMap mosaics, and the true Landsat TM5 image (Figure 13). First, forward

MNF transformations were applied to the reflectance bands of each dataset. Noise

statistics estimated during these transformations used shift differencing over the

complete scenes rather than over homogenous subsets (dark current measurements

from the sensors were not available) because an associated study found that full scene

estimations result in less spectral confusion when distinguishing target vegetation

from surrounding vegetation types (Mundt et al., 2007). For the Spencer true HyMap

mosaic the first 40 MNF bands explained 83% percent of the data Given that visual

inspection of the bands demonstrated noticeable degradation after band 30. Therefore,


64/129

the first 30 MNF bands were selected as input for pixel purity indexing. The first 30

MNF bands were also retained for the spatially degraded Medicine Lodge HyMap

mosaic. For images with 6 simulated or true Landsat TM5 bands (i.e., spectrally

resampled Medicine Lodge HyMap mosaic, spectrally and spatially resampled

Medicine Lodge HyMap mosaic, and the true Landsat image), all six bands were

retained for subsequent classification steps.

Pixel purity indexing of the first 40 MNF bands in the Spencer imagery

identified 6 potentially pure pixels that coincided with a known infestation of leafy

spurge from field analysis. A single ground reference plot was associated with this

infestation and subsequently defined as a training area for the classifications and

removed from the pool of Spencer validation samples. Pixel purity indexing of the

first 30 MNF bands for the Medicine Lodge true HyMap mosaic identified one

potentially pure pixel that coincided with a known occurrence of leafy spurge (Plot

54; Figure 14). This pixel was identified as a classification endmember and Plot 54

was subsequently treated as a training area and removed from the pool of Medicine

Lodge validation samples. The endmember identified for the MTMF classification of

the Medicine Lodge true HyMap mosaic was also used for the MTMF classification

of the corresponding spectrally degraded image. For the spatially degraded HyMap

image pixel purity indexing failed to identify any potentially pure pixels that


65/129

Figure 14. Ground reference samples in the general Medicine Lodge training area. Classification endmembers were derivedfrom pixels in the vicinity of Plots 54 and 65.

57

identified Medicine Lodge endmember (presented in Figure 14) was selected for


66/129

MTMF classification of thespatially degraded image. Two potential endmembers

were identified for the spatially and spectrally resampled Medicine Lodge HyMap

mosaic: a user derived endmember in the vicinity of Plot 54, and a PPI derived

endmember in the vicinity of Plot 65 (Figure 14). In situations where more than one

potentially pure pixel was identified, the final endmember was selected on the basis

of optimal performance. Figure 15 depicts spectral profiles for pixels that were

selected as final MTMF classification endmembers.

A series of classification methods were applied to the entire true Landsat TM5

image, including MTMF, SAM, maximum likelihood, and minimum distance. These

classification methods were applied to all six bands, both untransformed and MNF

transformed. The MTMF classification method was applied to the Medicine Lodge

area of the Landsat image using an endmember pixel from the Medicine Lodge

training area and the MTMF classification method was also applied to the Spencer

area of the Landsat image using an endmember pixel from the Spencer training area.

A total of five image-derived endmembers from the Medicine Lodge and Spencer

training areas were used in the SAM classification. MNF band variance and

endmember spectral profiles are presented in Appendix D.

3.3.7 Accuracy Assessment


67/129

and Green, 1999; Congalton, 2004; Foody, 2004). The error matrices were also used


68/129

to compute a kappa coefficient of agreement for each classification. The kappa

statistic is a measure of how well the classified map agrees with the validation

samples compared to chance agreement. An incremental cover technique was applied

to each classified Medicine Lodge image to quantify the minimum percent cover of

leafy spurge necessary to detect leafy spurge under various resolution scenarios using

the Producers accuracy (Mundt et al., 2006). This method assumes that pixels

containing high percent target cover are more likely to classify correctly than pixels

containing low percent target cover. The Producers accuracy is calculated by

dividing the number of samples classified as leafy spurge by the total number of

presence (field) samples. Consequently, the Producers accuracy for the presence

category should increase as infestations with lower percent cover are successively

removed. Changes in Producers accuracy are evaluated in cumulative 10%

increments. An acceptable Producers accuracy for this project was determined by

land managers to be 70% given the risk associated with committing an error of

omission.

3.4 Results

The true HyMap Medicine Lodge classification had an overall accuracy of

85% and a high kappa value (0.70) indicating that agreement in the error matrix was


69/129

User's Accuracy (Present) 100% 89% 90% 72% 96%

User's Accuracy (Absent) 60% 35% 40% 22% 30%

Produer's Accuracy (Present) 80% 59% 64% 31% 30%

Produer's Accuracy (Absent) 100% 75% 77% 62% 96%

Overall Accuracy 85% 62% 67% 38% 46%

Kappa 0.70 0.36 0.43 0.16 0.33

Entire Landsat TM5

image using SAM and

MNF transformed

input bands

Spencer area

Landsat TM5Classification

Med. Lodge

HyMap

Med. Lodge area

Landsat TM5

Spencer

HyMap

Table 2. Comparison of HyMap and Landsat classification performances for the Medicine Lodge and Spencer study sites.

61

outperformed MTMF classifications of the Spencer HyMap and Landsat TM5

imagery (Table 2) The MTMF classification of the Landsat image tuned to the


70/129

imagery (Table 2).The MTMF classification of the Landsat image tuned to the

Medicine Lodge area produced a high overall accuracy (62%) but was accompanied

by a low kappa value (0.36). For classifications of the entire Landsat scene (i.e.,

minimum distance, maximum likelihood, and SAM using both untransformed and

MNF transformed bands as input), the only classification method realistic enough to

quantify with an accuracy assessment was the SAM classification using MNF

transformed bands. This classification method produced a low overall accuracy of

47% and a low kappa value (0.33) indicated poor agreement. The high Users

accuracy for the present category (96%) and the high Producers accuracy for the

absent category (96%) are artifacts of the relatively low number of absence validation

samples.

The Medicine Lodge simulated Landsat TM5 classification outperformed the

true Landsat TM5 classification of the Medicine Lodge area (Table 3). The simulated

classification produced acceptable results for detecting leafy spurge infestations with

cover greater than approximately 40%, while the true classification failed to produce

acceptable results for detecting leafy spurge infestations with cover up to

approximately 80%. Incremental cover evaluations of the Medicine Lodge spectrally

resampled HyMap mosaic produced Producers accuracies and overall accuracies that

were higher than the true HyMap classification The simulated Landsat (spectral and


71/129

Medicine Lodge - MTMF ClassificationsSample Size 53 51 49 44 41 37 34 29 24 15

Leafy Spurge Cover Classes 0-100% 10-100% 20-100% 30-100% 40-100% 50-100% 60-100% 70-100% 80-100% 90-100%

HyMap (121 Bands, 3.2m pixels)

Producer's Accuracy 80% 85% 89% 88% 93% 96% 100% 100% 100% 100%

Overall Accuracy 85% 88%

27213549 Spectral and Spatial Detection Limits of Leafy

Documents

Transcript of 27213549 Spectral and Spatial Detection Limits of Leafy