Towards an operational root disease mapping methodology through lidar integrated imaging spectroscopy

(1)

by Geoffrey Quinn BA, Brock University, 2005

A Thesis Submitted in Partial Fulfillment of the Requirements for the Degree of

MASTER OF SCIENCE in the Department of Geography

_{Geoffrey Quinn, 2010} University of Victoria

(2)

Supervisory Committee

Towards an operational root disease mapping methodology through lidar integrated imaging spectroscopy

by Geoffrey Quinn BA, Brock University, 2005

Supervisory Committee

Knut O. Niemann (Department of Geography)

Supervisor

Dennis E. Jelinski (Department of Geography)

Departmental Member

David G. Goodenough (Department of Computer Science)

(3)

Abstract

Supervisory Committee

K. Olaf Niemann (Department of Geography) Supervisor

Dennis E. Jelinski (Department of Geography) Departmental Member

David G. Goodenough (Department of Computer Science) Outside Member

Root disease is a serious concern for the softwood timber industry. This thesis reports on the development of a root disease detection procedure that applies lidar data integrated with imaging spectrometer data. Photosynthetic pigments are frequently cited as one of the most responsive indicators of vegetation stress. This study estimated pigment content from needle and canopy reflectance and characterized the sensitivity of these pigments to a fungal-mediated stress. Samples were collected from the Greater Victoria Watershed District on Vancouver Island, BC, Canada. Lab reflectance measurements were made and pigments were extracted. Reflectance spectra were transformed into derivative spectra and a continuum removal band depth analysis was conducted. Reflectance metrics were generated and used in modeling pigment content. Chlorophyll-a was found to be significantly affected by the disease in the needle level portion of this study. The predictive power of reflectance attributes were assessed and yielded strong coefficients of determination (R2>0.80). Samples exhibiting stress responses affected by root disease were discriminated. It was determined that younger trees were more severely affected by the root pathogen than mature colonized trees. In the canopy level component of the study, chlorophyll-a was estimated through the application of partial least squares regression and achieved an R2 value of 0.82. Continuum removal metrics, which proved to be good estimators at the needle level, were found to be insufficient at the canopy level. Through the use of hyperspectral forest chemistry products, potential root disease sites can be identified.

(4)

List of Tables

Table 1. One-way analysis of variance for site effects on average tree foliar attributes.. 32

Table 2. Maximum correlation and spectral location... 50

Table 3. Regression comparison of continuum removal attributes ... 56

Table 4. Principal component regression of the red-well region (556-740nm) ... 60

Table 5. Partial least squares regression... 63

Table 6. AISA sensor characteristics ... 73

(7)

List of Figures

Figure 1. a) Tree 31 with mycelium. b) Scanning electron micrograph of hyphae ... 11

Figure 2. Irradiance by canopy position... 12

Figure 3. Spectrometer and apparatus... 15

Figure 4. Lab apparatus for pigment extraction ... 19

Figure 5. Average and standard deviation of healthy and colonized spectra... 27

Figure 6. Pigment content distributions for Douglas-fir samples. ... 28

Figure 7. Chlorophyll-a content by site and by canopy position. ... 29

Figure 8. Average tree pigment distributions... 33

Figure 9. Health classes suggest infected trees have greater variation ... 37

Figure 10. Variation in chlorophyll-a concentration between study sites... 38

Figure 11. Within tree variation of chlorophyll-a per site ... 39

Figure 12. Health classes suggest that infected trees have greater variation ... 40

Figure 13. Health class pigment distributions for young Douglas-fir... 42

Figure 14. Correlation spectra of pigments expressed on a fresh weight basis ... 44

Figure 15. Correlation spectra of pigments expressed on a leaf area basis ... 45

Figure 16. Correlation spectra of moisture content per leaf area and fresh weight ... 47

Figure 17. Derivative spectra of average chlorophyll-a classes... 53

Figure 18. Chlorophyll-a class continuum removed spectra... 55

Figure 19. Component space identifies sample 13w as distinct ... 59

Figure 20. Component space resolves the chlorophyll-a signal... 61

Figure 21. Red edge position in relation to PLS coefficient values... 65

Figure 22. ASD spectra of bright concrete, dark asphalt and standard deviations ... 77

Figure 23. Linear across track illumination effect in AISA data ... 80

Figure 24. Smile analysis in the AISA hyperspectral dataset. ... 81

Figure 25. Radiance spectra and atmospheric transmission ... 83

Figure 26. Ground spectra and ATCOR surface reflectance for atmospheric correction.85 Figure 27. Validation spectra for comparison of three approaches to correction. ... 87

Figure 28. Flight 0805-2009 geometric correction ... 88

Figure 29. Translation of lidar identified treetops to reflectance imagery... 92

Figure 30. Correlogram comparison of spectral extraction methods... 93

Figure 31. Average chlorophyll-a and reflectance at 682nm... 94

Figure 32. CRDN treetop spectra and chlorophyll-a correlation ... 95

(8)

Acknowledgments

Many people contributed to this work and I am grateful for their help and guidance. I want to acknowledge the Greater Victoria Watershed District staff and management who supported field reconnaissance and ground reference data acquisition. I would like to thank Dr. Barbara Hawkins and Samantha Robbins at the University of Victoria’s Centre for Forest Biology, who provided expert consultation and facilities for pigment extractions. Kevin Pellows and Dr. Rona Sturrock from the Pacific Forestry Centre, Natural Resources Canada, were invaluable in identifying appropriate sites through demonstrating techniques for field diagnosis of root disease; we are thankful for the financial support of The British Columbia Transmission Corporation. I would like to personally thank Fabio Visintini, Rafael Loos, Roger Stephens and Laura Duncanson for assistance in field campaigns. Thanks are given to Diana Parton for lidar data processing. We are appreciative for Terra Remote Sensing for operating the AISA sensor. I would like to thank Professor Olaf Niemann and Dr. David Goodenough for their guidance and the opportunity to work with them, which has been a truly enriching experience. I would like to thank my family, my Dad and Lorrie have always been a big inspiration for me and I am very grateful for all their support over the years. Most importantly, I need to thank my loving wife Ashley, who should be recognized for her continual encouragement.

(9)

Chapter 1

1.1 Research motivations

With the onset of global climate change, it is important that northern forests are monitored in parallel with climatic parameters. The Canadian Boreal Forest extends the full latitudinal ranges that exhibit the forest species tolerable temperature ranges. The Boreal Forest species are thus likely to respond to changes in global temperature, firstly by health status and ultimately, in their spatial extents. Changes in Canadian forests are likely to occur before the effects in more equatorial forests become conspicuous (Root et al. 2003). Additionally, diseased trees have decreased CO2 assimilation rates and reduced

photosynthetic rates (Manter 2002), which have implications for national and global carbon accounting. Forest health research has the potential to provide a better understanding of ecosystem functioning and trajectories, while also documenting significant departures from the norm. A considerable challenge for monitoring and modeling forest health is the characterization of normality when much variability is expected in natural systems. In the coastal forests of the Pacific Northwest, substantial naturally induced variation makes determining normal forest attributes difficult. Much of this variation is the manifestation of natural agents of forest disturbance whose effects must be investigated and quantitatively measured to characterize normal forest form and function.

Laminated root disease is a naturally occurring forest pathogen. This disturbance factor is manifested in forest stand canopies as an increased frequency and magnitude of canopy gaps and reduced growth rates. While the disease may have a negative identity from a timber production perspective, the disease promotes increased biodiversity and therefore the resilience of the ecosystem. The traditional field-based methods for mapping disease infected stands are costly, inaccurate, and limited to low sampling densities. Spectroscopy applied in the field, lab and from airborne platforms has emerged as a potential tool for mapping the spatial distribution of vegetation health by means of foliar

(10)

chemistry estimation. This approach has a proven capacity for the detection of moisture and nutrient deficient environments, air pollution, contaminated soils, photo-inhibition and a wide range of biological stressors (Suarez et al. 2008; Claudio et al. 2006; Stimson et al. 2005; Solberg et al. 1998; Whitehead et al. 2005; Campbell et al. 2004; Noomen et al. 2006; Schuerger et al. 2003; Gamon et al. 1992; Falkenberg et al. 2007; Muhammed 2005; Apan et al. 2004).

The methods applied in early remote sensing stress detection have exploited multispectral reflectance data for the estimation of foliar pigmentation. Multispectral sensors do not adequately sample the electromagnetic spectrum to sufficiently characterize the degree of chlorosis or the variation in foliar pigment concentration (Leckie et al. 2004). As a result previous efforts applying broadband spectral indices have been successful in distinguishing coarse health classes but typically fail with more precise chlorosis class intervals. Additionally, remote sensing approaches have relied on coarse spatial resolution datasets that typically do not resolve fine scale patterns and processes. Forest research applied to physical processes occurring at the individual tree level requires a scale that is smaller than, or equivalent to, that of the target tree canopies. Therefore, imaging spectroscopy with a 2 meter raster applied to mature tree canopies, with diameters greater than 6 metres, of the maritime Costal Western Hemlock and Douglas-fir Biogeoclimatic zone should be sufficient for this task. At spatial resolutions which are coarse relative to tree canopies, the contributions of non-target reflective components are expected to be significant and contribute a considerable source of error (Niemann et al. in progress).

Data that are accessible to imaging spectroscopy include the spatial attributes of a targeted subject including its size, shape and texture and of course its spectral characteristics which are largely determined by the material’s chemical composition. Root disease sites are expected to exhibit specific traits for these attributes. The pathogenic fungus, Phellinus sulphurascens, causes a reduction in the efficiency of root functioning which is expected to impact foliar chemistry. While the size of disease centers is largely dependent on the age of the infection, the shape is expected to be

(11)

roughly circular, as the infection and subsequent windthrow radiates out from an inoculum source.

Previous efforts to infer health status through spectroscopy have focussed mainly on the visible and near infrared region (400 through 1000 nanometres) of the spectrum. Likely the most popular approach applies the red edge feature (Kochubey and Kazantsev 2007; Baranoski and Rokne 2005; Smith et al. 2004; Demetriades-Shah et al. 1990; Rock et al. 1988); however, this method has limited application to airborne sensors with decreased spectral resolutions relative to laboratory spectrometers. The variation in the position of the red edge is approximately 30 nanometres which must be sufficiently sampled to represent the deviation in the feature. Airborne studies that apply the red edge approach have a necessity to apply curve fitting methods that may, or may not, adequately represent the true material response (Cho and Skidmore 2006). Other more broad features applied to airborne vegetation stress detection are likely to produce better results as these features are more aligned with the spectral characteristics of modern airborne sensors. Because these spectral features are wider, the likelihood of overlapping spectral absorption features is increased. These methods are subjected to complicated transformations to reduce the effects of such background features (Blackburn 2007; Kokaly et al. 2007; Noomen et al. 2006; Stimson et al. 2005; Pu et al. 2003; Curran et al. 2001; Kokaly and Clark 1999; Clark and Roush 1984). The red edge position and the continuum removal transformation are two examples of spectroscopic analysis methods that are not feasible for multispectral datasets.

A study to address some of the issues identified above related to the early detection of root rot was devised. This study was divided into two parts. The first involved a laboratory-based experiment to firstly investigate the impact of the disease on foliar biochemistry and secondly to determine, under optimal conditions, the potential for inferring biochemical concentrations from reflectance spectra. The second portion of this study considered an airborne-collected reflectance dataset and expanded on the relationships found at the leaf or needle level. This second part of the study provided a

(12)

proof of concept for the pre-visual identification of unhealthy tree canopies through the application of hyperspectral remote sensing.

This thesis, detailing the procedures, motivations and results of the remote detection of root disease, begins with a needle level assessment. Through this laboratory study, potentially confounding factors expected at the canopy level could be avoided. Therefore, the needle level analysis would determine the feasibility of an operational airborne detection methodology. For example in the laboratory, samples could be prepared such that non-target materials did not contribute to the observed spectral responses. At the canopy level, non-foliar tree materials (shadow, bark, cones and lichen) would contaminate pure needle spectra, not to mention the effects of the underlying ground materials (soil and understory vegetation). In addition, the illumination environment varies significantly at the canopy level, especially in the spatial domain, as a consequence of bidirectional reflectance distributions and the surface structure of the target. The illuminating conditions may also vary in the temporal domain as light cirrus clouds pass between the target and source causing decreased incident irradiance levels. In the lab, illumination conditions are constant and irradiance, in this case a full range lamp, is static and consistently monitored. Following the needle level analysis an airborne dataset was investigated. Careful consideration was given to data quality and the potential for spectral mixing. Leveraging on previous studies demonstrating methods for pre-processing and an integrated lidar dataset, these concerns were addressed. Finally, the results found through this research are summarized and future research directions are suggested.

(13)

Chapter 2

2.1 Introduction

Laminated root disease, like many other forest pathogens, is considered a natural agent of forest disturbance that promotes increased biodiversity, forests structure and the health of the forest ecosystems (Frazer et al. 2005; Gilbert 2002; Hansen and Goheen 2000; Hennon 1995; Holah et al. 1997; McCarthy 2001). Once the disease is established within the root network, the anchor systems of contiguous colonized trees are compromised and ultimately, canopy gaps are created as high wind velocities topple the infected trees. In free-to-grow forest stands, early seral species are established within these gaps, which are resistant to infection. The resistant tree species outlive the fungus, which is not known to live much beyond fifty years without a live host (Hansen 1979). It is expected that forest management practices including mono-species stand replacement has increased the frequency of root disease populations to as high as 80% in some cases (Bloomberg and Reynolds 1985). Increased disease occurrence in managed forests is attributed to the absence of the typical species succession. Alternatively, it has been demonstrated that harvesting practices effectively reduce soil microbial diversity (Outerbridge et al. 2009). Harvesting practices have the potential to reduce populations of antagonistic fungal species and consequently increase disease virulence (Glodfarb et al. 1989; Hutchins 1980).

Laminated Root Rot is a disease caused by the fungal pathogen known as Phellinus

sulphurascens. The fungus is native to the Pacific Northwest and may have coevolved in

a sort of evolutionary arms race with its primary host tree species, Douglas-fir (Pseudotsuga menziesii) (Gilbert 2002; Hansen and Goheen 2000). The fungus is classed in the Phylum Basidiomycota, a taxonomic group without a singular defining feature, but that is widely characterized by the occurrence of spore dispersing fruiting bodies.

P.sulphurascens sporophores are rarely documented in the field, therefore the

(14)

of mycelium (Lim et al. 2007). As the root systems of healthy Douglas-fir trees expand and make contact with the roots of diseased trees, the mycelium of P.sulphurascens grows proximally and distally as ectotrophic mycelium to colonize the surface of the once healthy roots. Only recently was the mechanism by which the fungus infects healthy host roots determined. At hyphal tips, specialized ultrastructures are formed called appressoria. The appressoria, which are not specific to P.sulphurascens, are swollen and flattened structures that adhere to the host root surface. They produce sufficient turgor pressure to mechanically penetrate the host tissues, these structures are also known to produce host cell-wall degrading enzymes (Islam et al. 2009). The fungus penetrates the root through intact bark, killing phloem and cambium as it enters and initiates decay in the xylem (Wallis and Reynolds 1965). Both lignin and cellulose are metabolized (Hansen and Goheen 2000). Unlike Armillaria ostoyae root disease, which produces specialized structures called rhizomorphs, P. sulphurascens mycelium is incapable of growing through the soil substrate and exclusively lives on the surface of host roots. Therefore, the disease has the tendency to occur as contiguous pockets of infection.

The disease has become a major topic within forestry discourse due to the increased occurrence of the fungus in second-growth timber stands and due to the threat it poses to the forest crop investments. Laminated root disease has been estimated to cost the forestry sector of North America 4.4 million m3 of timber annually through reduced growth rates (Nelson et al. 1981; Sturrock et al. 2006; Thies and Sturrock 1995). These reduced timber volumes equate to approximately $528M (CAD) of lost revenue for the North American economy per year. To date, little progress has been made towards the efficient control of the disease spread, although a long term project assessed the efficacy of a number of fumigation techniques and fungicides, but achieved relatively little success (Thies and Westlind 2006). The message communicated through long term scientific studies is that the only seemingly effective method of containment is through the reduction of residual inoculum and the establishment of resistant species (Thies and Westlind 2005). One need is, therefore, the effective and accurate mapping of infected trees, a significant challenge for traditional field-based methods which are costly, inaccurate and limited in spatial extent.

(15)

Monitoring the spread of the disease has proven to be an illusive feat, as observable canopy symptoms consistently underestimate the extent of infection (Wallis and Reynolds 1965). In addition, non-destructive sampling of below ground indicators of infection have a tendency to fail in identifying positives (Sturrock et al. 2006). Consequently, there is a need for a more accurate detection method of infected trees. Recent advancements in the field of molecular pathology have demonstrated a chemical signal response to P.sulphurascens infection (Ekramoddoullah et al. 2000; Robinson et al. 2000; Zamani et al. 2003). Ekramoddoullah et al. (2000) were able to detect a chitinase-like pathogenesis-related protein in the foliage of infected Douglas-fir trees. The authors suggest that this chemical is increased in response to fungal infection and is capable of metabolizing the chitin-containing cell walls of invading hypha. These methods are limited to low sampling densities and are therefore not practical for large scale studies. However, chemical signalling might prove valuable in the validation procedure for an operational root disease detection protocol. Furthermore, while the abundance of chitinase may not be conducive to detection in electromagnetic reflectance, other more plentiful foliar substances may be sufficiently impacted by the disease to enable sensible changes in reflectance.

The discipline of spectroscopy has evolved greatly from its early beginnings when Joseph von Fraunhofer observed more than 500 narrow spectral gaps in the refracted solar spectrum. Through elemental spectroscopy these dark spectral features were linked to the frequency-specific absorption of electromagnetic energy by certain molecular species of the solar and terrestrial atmospheres (Liu et al. 2005; Schaepman et al. 2009). Analytical spectroscopic methods have developed sufficiently, and are now applied widely. Motivated by NASA’s Accelerated Canopy Chemistry Project (Goetz and Davis 1991) a substantial amount of research has been concentrated on the decomposition of the vegetation spectral signal. Recently spectroscopic methods have demonstrated a capacity for the estimation of moisture (Ceccato et al. 2001; Ceccato et al. 2002; Niemann et al. 2002), lignin and cellulose (Curran et al. 2001; Kokaly et al. 2009; Serrano et al. 2002), nitrogen (Goodenough et al. 2007; Wang et al. 2007), chlorophylls (Blackburn 2007; Richardson et al. 2002), carotenoids (Blackburn 1998; Ustin et al. 2009) and even

(16)

xanthophylls (Evain et al. 2004; Gamon et al. 1992) many of which also include canopy level assessments applying imaging spectroscopy data. Although the connections between foliar chemistry and reflectance in deciduous species have been thoroughly characterized, there has been comparatively little research conducted on biochemistry and coniferous species (Moorthy et al. 2008). Not unlike early remote sensing stress detection projects, the prior efforts to map the foliar effects of root disease have applied multispectral reflectance data.

Beginning in June of 1996 and funded by Forest Renewal British Columbia, an industry and government cooperative project titled “Development of Certified Forestry Applications Using Compact Spectrographic Imager Data” was conducted in partnership with Weyerhaeuser Ltd., Macmillan Bloedel Ltd. at the time, Itres Research Ltd. and the Canadian Forestry Service (Pacific Forestry Centre). As part of the larger project, a study was performed to assess the capacity of airborne imaging technology for mapping root disease at a site near Nanaimo, BC. A high spatial resolution (60cm) Compact Airborne Spectrographic Imager (CASI) dataset was acquired in multispectral mode. The sensor collected reflectance from eight discrete bands sampled at roughly 50nm intervals within the VNIR portion of the spectrum (438-861nm). The authors delineated tree canopies using a valley-following algorithm and extracted mean-lit reflectance values for each identified canopy. Having qualitatively assessed ground reference trees for chlorosis, a broad classification strategy attempted to classify disease severity. The Jeffries-Matusita distance between health classes was used to perform feature selection, which suggested that blue, red, NIR red/NIR and NDVI produced the best separation of health classes. The classification of isolated canopies did not successfully identify diseased canopies, suggested by a maximum likelihood classification with an overall accuracy of 57%. Although some infected trees were correctly identified, a reconnaissance survey determined that many trees scattered throughout the imagery were erroneously classed as affected (Leckie et al. 2004). It is suggested that, opposed to performing subjective qualitative assessments of canopy condition, foliar samples be collected and quantitatively assessed for pigment contents. In addition, multispectral sensors do not adequately sample the spectrum to characterize the degree of chlorosis or variation in

(17)

foliar pigmentation. Therefore, in such studies imaging spectroscopy should be applied in place of multispectral data with lower spectral sampling densities. The study provided some support for a large body of literature that suggests foliar pigmentation and reflectance at the corresponding spectral absorption regions are capable of detecting vegetation stress.

Pigment content determination has been widely applied to the detection of stress imposed by moisture (Suarez et al. 2008) and nutrient deficits (Solberg et al. 1998), air pollution (Rock et al. 1988; Smith et al. 2004) photo-inhibition (Demmig-Adams and Adams III 2006) as well as other biological stressors. In the case of P.sulphurascens root disease, once the infection has sufficiently developed to alter root function, the infected specimen will begin to exhibit the effects of reduced nutrient and water availability. Stressed trees experience decreased photosynthetic rates resulting from a lack of water and the consequent disruption of electron transfer and ultimately photo-inhibition where reactive oxygen species inhibit the repair of damaged photosystems (Manter 2002). Alternatively, a lack of nitrogen uptake may translate into increased nitrogen translocation to developing foliage and the subsequent senescence of older foliage (Hawkins and Henry 1999).

Thomson et al. (1996) reported on a study to asses the effects of Phellinus weirii (later recognized as Phellinus sulphurascens) infection on the foliar chemistry of Douglas-fir trees. The study investigated nutrient content and pigmentation of current and non-current foliage from 25 trees at each of eight plots and at three dates. This study determined the infection status of sample trees by, not only canopy symptoms but also through an examination of root collars to confirm the presence of ectotrophic mycelium. The authors observed reduced foliar moisture in affected specimens. Affected samples were also found to have decreased foliar chlorophyll-a and nitrogen relative to healthy samples. Chlorophyll-b was found to be highly variable and of little use in distinguishing between the health classes. It was determined that finding significant differences between healthy and affected samples would be dependant on the nutrient assessed, age of foliage, phenology and the desired accuracy. Of the nutrients assessed, nitrogen exhibited the

(18)

lowest coefficient of variation while older foliage was found to be less variable than new foliage and among others, chlorophyll-a and nitrogen increased throughout the growing season. These findings support a sampling strategy that samples non-current growth and applies chlorophyll-a as a surrogate for nitrogen at a single sampling campaign to distinguish between healthy and affected samples.

To build on and extend from earlier studies on the remote detection of vegetation stress, this work is concerned with the detection of subtle pigment concentration changes within Douglas-fir foliar samples from laboratory reflectance measurements. The objectives of this leaf level research are two fold. The primary objectives are to spectrophotometrically characterize the response of foliar chemistry to P.sulphurascens infection and secondly, to establish an accurate pigment estimation protocol by applying new vegetation spectroscopy analysis procedures in a needle level reflectance study. The knowledge and experience obtained through these experiments will be up-scaled to a canopy level study utilizing data acquired from airborne platforms. Ultimately, the goal is the development of a large scale operational root disease mapping method.

2.2 Methods

2.2.1 Field sample collection

A standardized specimen selection protocol was devised loosely based on the spread pattern of the disease. Given the location of viable residual inoculum, and assuming that it was the original source of the infection, infected and otherwise healthy trees could be selected for sample acquisition. Provided that infection centers did not intersect, the likelihood of selecting an infected specimen would decrease with increasing distance from the site center, or the identified inoculum. Six transects were established radiating at 60° intervals from the identified source of infection. Along each transect two specimen were selected. The closest tree to the site center was most likely to be infected and, given disease spread estimates are approximately 30cm/year (McCauley and Cook 1980), and the stand ages, which are roughly 60 years, healthy trees were expected to occur ~18 meters from the site center or beyond. The infection status of sample trees was

(19)

Figure 1. a) Tree 31 at site G2 with melanized pseudosclerotial plate partially excised

revealing white mycelium. b) Scanning electron micrograph cross-section of

pseudosclerotial plate illustrating swollen, fused hyphae (adapted from McDougall 1996)

ultimately determined through lateral root excavations. Excavations were approximately 0.5m deep and extended to ~1m from the tree stem. Exposed roots were assessed for the presence of ectotrophic mycelia. The development of the mycelia mat begins as thin filaments, which converge and produce a white homogenous layer of swollen hyphae. After approximately a week, the surface mycelia turn a brownish color, a structure that mycologists refer to as the pseudosclerotial plate (McDougall and Blanchette 1996). Both, the early white stage and the later melanized form of the mycelia were used as evidence for the presence of the fungus on a selected tree specimen (Fig. 1).

Due to the within-tree variability that is expected for pigment and reflectance measurements, a sampling strategy was devised that would capture the range of within-tree pigmentation. In considering the indirect influence that aspect exerts on photosynthesis, it was decided that from each sample specimen, foliage from all cardinal directions would be sampled. As a result of the diurnal fluctuations in photosynthetic photon flux density, samples collected from north and southerly aspects were expected to occur at the extremes of pigment and reflectance ranges, while East and West aspects were anticipated to occur somewhere within that range. In the northern hemisphere, south facing aspects are exposed to higher photosynthetic photon flux densities than north

5cm 5µm

(20)

Figure 2. Diurnal photosynthetic photon flux densities at north and

south aspect, subscript indicates upper (U) and lower (L) canopy regions (Adapted from Abadia 1996).

facing aspects yielding marked differences in light harvesting apparatuses. Chlorophyll is found in greater concentrations in shade grown Douglas-fir trees than open conditions, although this finding is not specific to Douglas-fir (Abadia et al. 1996; Khan et al. 2000; Major et al. 2009). The four cardinal direction canopy sampling strategy was applied to control for the variability in photosynthetic pigment concentrations of sun exposed and shade foliage that are attributed to the differing irradiance regimes of the cardinal directions (Fig. 2).

A compromise was struck between selecting areas of the canopy that are visible to distant platforms and those that are comprised of sufficiently developed foliage. Newly developed foliage tends to have different chemical signatures and higher variability in foliar chemistry than older foliage in Douglas-fir canopies (Goodenough et al. 2009; Thomson et al. 1996). In terms of needle age, samples needed to only satisfy the non-current requirement to ensure samples collected were at roughly the same developmental stage. Lepedus (2005) demonstrated, in Picea abies, a lack of statistical significance in total chlorophyll between samples of successive growth years beyond second year samples. The fourth whorl from the leader was selected as the target branch from which samples would be collected. As further justification for selecting the upper portion of the

(21)

canopy for sample acquisition, shade foliage and sun-exposed foliage are not comparable because nitrogen composition, pigment concentrations and specific leaf area are found to differ significantly (Han et al. 2004a; Khan et al. 2000; Nippert and Marshall 2003).

Once the samples were received, they were stored to minimize protein degradation and ensure that pigment and moisture contents would reflect the growing conditions rather than the sampling strategy. Shortly after samples are cut from a tree, the photosynthetic efficiency promptly begins to decline followed by pigment degradation and a concurrent decrease in needle water content. Nippert and Marshall (2003) illustrated that CO2

assimilation rates decline 50 minutes following the excision of Douglas-fir needles. The sensitive nature of sample acquisition requires that samples be maintained at low temperature and light levels to minimize protein degradation. A sub sample of approximately 15 needles were taken from each sample and placed in a cryogenic vial, which was then stored on dry ice. Once received at The University of Victoria, samples to be processed for pigment analysis were transported to a freezer capable of sustaining temperatures of -80°C, where they would remain until wet laboratory methods could be performed. Samples intended for spectral analysis were stored on ice rather than dry ice, as freezing and thawing is known to affect reflectance characteristics (Carter and Knapp 2001; Nicotra et al. 2003). In total 192 samples were intended to be collected. However, the needle loss of two selected sample trees was so severe that sufficient sample quantities could not be collected. Ultimately 184 samples were collected, (4 samples per tree x 12 trees per site x 1 site per day x 4 days). Forty-eight samples for spectral analysis, pigment analysis and moisture analysis were taken from a study site per day. The sample collection was severely time constrained. Due to the volatile nature of harvested samples and the restriction that they must not be cryogenically stored, reflectance samples needed to be processed immediately and therefore limited the number of samples that could be processed in the remaining hours of a sample acquisition day. The number of samples chosen to be collected was selected based on the estimated number that could be processed in the evening, after sample acquisition.

(22)

2.2.2 Laboratory reflectance spectroscopy

Immediately following the storage of pigment samples, reflectance samples were processed. The FieldSpec Pro spectrometer (Analytical Spectral Devices, Inc.) was used in a darkroom laboratory setting to acquire reflectance data from Douglas-fir needle samples. The spectrometer is composed of three subsystems covering the range of VNIR (350-1050 nm), SWIR1 (900-1850 nm) and SWIR2 (1700-2500) (Analytical Spectral Devices Inc. 2002). A fibre optic cable with an FOV (field of view) of approximately 25° receives light from a target, which is then reprojected onto a diffraction grating that separates the captured light energy into its component wavelengths. The refracted light is directed onto a linear array of 512 silicon photodiode detectors; here the resultant voltage from absorbed photons is converted into a 16-bit digital signal that is stored onto a portable lap top computer. The VNIR sensor has a sampling interval of ~1.4nm and a spectral resolution of ~3nm (full width half maximum at 700nm). In contrast to the VNIR which has 512 detectors, each SWIR sensor has only one. The SWIR sensors cope with the lack of detectors by utilizing a concave grating that oscillates to complete a scan of the SWIR range every 100 milliseconds. Thus, different energy frequencies are stored consecutively rather than simultaneously like the VNIR sensor. The sampling interval of the SWIR sensors is ~2nm and the spectral resolution varies between 10nm and 12 nm (Analytical Spectral Devices Inc. 2002). Applying the RS3 software, reflectance spectra were viewed and recorded. The software receives the sensor signal with variable bandwidths and sampling intervals. With the sensor specific calibration and polynomial responses for each detector, a response weighted function is applied to interpolate the signal and output the spectra with reflectance values at each nanometre within the sensor spectral range (Fager 2009).

Prior to measuring reflectance, an automated signal optimization was performed and the dark current was recorded. The internal software of the ASD optimizes the integration time for capturing reflectance to maximize the signal without saturation; the function is analogous to the shutter speed and aperture of a common film camera or the pupil of an eye that dilates to regulate the amount of light received (Analytical Spectral Devices Inc. 2002). A Spectralon (Labsphere, Inc.) white reference standard was used to acquire

(23)

Figure 3. Spectrometer and illumination source geometry for

darkroom laboratory reflectance measurements

comparative reflectance spectra defined as the ratio of sample radiance (W/m2/sr/nm) to irradiance (W/m2/nm). The white reference standard is a target characterized as a near perfect diffuse reflector and is therefore an appropriate representation of the illumination quantity and quality (Labsphere 2006). Upon completing the measurement of a white reference standard the ASD automatically initiates the dark current measurement. In order to compensate for the additive effect of the instrument and its associated electronic currents on the measured signal, the dark current is subtracted from each spectral curve recorded thereafter. A Pro-light (Lowel Light, Inc.) with a 50 watt, quartz-halogen cycle tungsten filament lamp was applied because the emission spectrum of the lamp closely approximates the solar spectrum with a 3200°K color temperature.

In preparation for spectroscopy the laboratory apparatus was assembled. On a dark background material and in a dark room, one light source, a stage and the spectrometer were arranged (Fig. 3). An illumination and viewing geometry was established as a standard, which would remain constant throughout the duration of lab spectral acquisitions. The light source was adjusted to an elevation of roughly 40° and

(24)

approximately half a meter from the target sample. The bare fibre optic was placed in the pistol grip and oriented nadir at a distance of 2.3cm from the sample stage and stabilized by a tripod that would remain constant throughout the course of spectroscopy. Half the bare fibre optic FOV and a desired view spot radius of ~0.5cm was used to calculate an appropriate distance to the sample stage (d= r/ tan∝). Samples were transported to the lab on ice and in a cooler. The ASD was turned on and allowed to stabilize for a period of approximately one hour.

Individual samples were randomly selected from the cooler and processed for reflectance until all samples were used. From each sample in turn, 20 needles were removed and carefully aligned abaxial side down on black tape and mounted to a sample stage. The black polystyrene sample stage with a flat finish was used to produce a low, flat, featureless and diffuse background spectral response. Sample identification numbers and spectral file names were documented. A white reference was recorded ten consecutive times, each of which was measured as the arithmetic mean of ten measurements. With the specified geometries and acquisition parameters, the ASD RS3 software optimized an integration time of 136 ms/scan. The response of the white standard was monitored to ensure a stable sensing condition, verified by a consistently flat curve at a value of one. The white reference panel was removed from the stage and replaced with a prepared needle sample. The sample was positioned with needles at an approximate angle of 45° to the illumination azimuth. In following the protocol for the white target, the mean reflectance of ten measurements was recorded ten times. The sample was reoriented with needles orthogonal to the direction of illumination and measurements were taken in the same manner. Finally, the sample was repositioned for the last time with needles parallel to the azimuth of incidence and reflectance was recorded. The orientation of the samples were altered throughout the process of data collection in order to capture the variability that might be attributed to the bidirectional reflectance distribution function, which could then be accounted for in averaged spectra (Vaiphasa et al. 2005). After spectra were recorded, the sample was discarded. Another white reference was recorded while the next sample was prepared. This procedure was repeated until all samples had been measured.

(25)

2.2.3 Moisture content determination

Before samples were collected, 96 cryogenic vials were labelled and weighed to a precision of 0.01mg with an OHAUS AP250D analytical balance. The manufactures’ stated repeatability of the balance was +/-0.02mg. This particular model of balance has a measurement capacity of 52g and an average stabilization time of 12 seconds (Ohaus Corporation 1999). The moisture samples were composed of 30 whole needles and a vial with an approximate mass of 1.5g. The proposed sample and carrier fell well within the measurement capacity of the balance. Upon establishing the balance on a level surface, in an environment with minimal air currents, vibrations, temperature and humidity extremes, the balance was allowed to stabilize for 20 minutes. Three measurements were taken and the average recorded for each empty cryogenic vial. The standard error, calculated from thirty measurements from each of several vials with constant weight was found to be ±0.01mg. Therefore, a single measurement could be expected to occur within ±0.01mg of the samples true weight. An average of three measurements produced a marginal increase in accuracy with an error of ±0.009mg when compared to the true weight, approximated as the average of thirty measurements.

Once in the field, samples were collected for water content analysis. From every sample collected for pigment and spectral analysis, 30 needles that were adjacent to those used in pigment analysis were removed and placed in the pre-labelled and weighed cryogenic vials. Justification for the use of sealed cryogenic vials was that after needle abscission transpiration rates were expected to change and a concurrent reduction in moisture might also occur. Therefore, the vials needed to be air tight to ensure the mass of escaping water vapour was maintained within the measured sample. Samples intended for moisture analysis were stored and transported on ice in a cooler. After the spectroscopic analysis was completed the moisture samples were weighed using the analytical balance. For each sample, needle fresh weight was calculated as the difference in the mean sample mass and the corresponding mean empty vial mass. Because two measurements were required, the final mass uncertainty for each sample was ±0.02mg. After sample masses were recorded the projected leaf area of each sample was measured in order to determine the equivalent water thickness.

(26)

Equivalent water thickness is a variable that is more closely related to reflectance parameters than percent moisture content (Ceccato et al. 2002). Once all time-critical measurements were complete, moisture samples were processed for leaf area. A prior experiment demonstrated that samples stored in a fridge would experience a decrease in leaf area of approximately 2% after four days. A LI-3100C area meter was warmed up for 30 minutes before samples were measured. Using a transparent calibration mask of known area the leaf area meter was calibrated to an area of 8.8cm2. For each sample the 30 needles were removed from their vials and aligned adaxial side down between two transparent plates. The needles were then passed through the area meter five times, each measured area was recorded and the average area was calculated. The method of area measurement produced an accuracy of approximately 0.05cm2 assessed through calibration transparencies. After area measurements were completed, the samples were replaced into the labelled vials and the next sample was measured for projected leaf area.

Once the area of each sample had been recorded the samples were dried to remove all moisture from the needles. The samples were placed in clean ceramic crucibles, which were placed in an oven and heated at 80°C for 24 hours. The dry mass of each sample was recorded immediately after the samples were removed from the oven. Moisture content was calculated as the difference between fresh and dry weight expressed as a percentage of fresh weight and as equivalent water thickness, grams of moisture per leaf area.

2.2.4 Pigment extraction

The determination of needle pigment concentration was performed implementing the methods of Warren et al. (2004). The instrumentation required for the procedure included: an analytical balance, mechanical homogenizer, precision pasture pipette, vortex mixer, a centrifuge and a spectrophotometer. Prior to pigment extraction samples were stored in sealed cryogenic vials and maintained at -80°C temperatures. The workspace and fume hood were prepared as illustrated in figure 4. The stored samples were transported on dry ice to the work station. From each sample in sequence, 50±1mg

(27)

Figure 4. Lab apparatus for pigment extraction in fume hood

(0.049-0.051g) of needle material was weighed and relocated into a new labelled 2ml cryogenic vial. Five millimetre diameter glass beads were then placed into the vials containing the needle material. A ball mill was used in place of the traditional mortar and pestle for material maceration. The glass bead functionally replaced the use of a pestle. The vial labels with corresponding sample identification codes and sample masses were documented. After the mass of each pigment sample was documented and then prepared for extraction in their respective vials, the vials were again placed on dry ice to limit protein degradation.

The spectrophotometer was given sufficient time to stabilize prior to processing the extracted pigment solutions (~30 minutes). The samples were placed in the grinding mill (Wig-L-Bug model 30) for approximately five seconds, until the sample was thoroughly macerated. Photosynthetic pigments are severely sensitive to light and are thus predisposed to photooxidation (Lichtenthaler 1987). As a result the remaining duration of the extraction process was performed under low diffuse light conditions. Four randomly selected samples were processed at a time to minimize the exposure to increased ambient light levels and temperature; the remaining samples were kept in the cooler on dry ice. The four samples were briefly placed in a centrifuge (short pulse) to remove any needle

(28)

material that may have settled on the caps of the vials. Using a precision pipette one millilitre of pure N,N-dimethylformamide (DMF) was added to each cryogenic vial. All four vials were held to the mixing surface of the vortex for approximately four minutes before running them through a centrifuge cycle at 10,000g (relative centrifugal force) for four minutes. The resultant supernatant was removed from each vial and placed in individual graduated cylinders using pasture pipettes. The procedure was repeated and the supernatants from consecutive extractions were pooled into sample specific graduated cylinders. The extractions were complete once the fibrous pellet no longer exhibited the green hue of chlorophyll pigments. Extractions were typically complete after three to six extractions, depending on the original state of chlorosis of the samples. The final volume of the sample supernatants were documented, which were later used to calculate sample pigment concentrations. Quartz cuvette cells were rinsed with clean DMF in preparation for spectrophotometric sampling.

The Philips PU8620 series, single beam UV/VIS/NIR spectrophotometer was used, which employs a tungsten-halogen source and a deuterium arc for sensing wavelength shorter than 325nm. Upon powering the instrument an automated wavelength calibration, stated to be accurate to within one nanometre, was performed. The manufacturer claims an absorption value accuracy of five percent (+/- 0.002A) for the solid-state photodiode detector which, for example, translates to an accuracy of 5.12µg/g for chlorophyll-a. The spectral sampling of the PU8620 is one nanometre with an eight-nanometre bandwidth throughout the wavelength range of 195 to 1100nm (Philips PU8620 UV/VIS/IR Spectrophotometer manual).

From each pooled sample pigment extract solution, one cuvette was filled. A fifth cuvette was filled with clean DMF, which would be used to calibrate the spectrophotometer. The sensing wavelength of the instrument was set to 664nm. The cell with clean DMF (blank) was inserted into the chamber of the spectrophotometer and the instrument was calibrated to zero absorption. Each prepared cuvette was then measured for absorption at 664nm; the absorption coefficient was recorded once the instrument had stabilized. This procedure was repeated for wavelengths 647 and 480nm. Sample solutions, which had

(29)

measured absorption values greater than 0.8 in either wavelength, necessitated the dilution of the sample with DMF. The sensitivity of absorption readings to solute concentrations begin to decrease at higher concentrations. High solute concentrations have the tendency to produces absorption coefficients that saturate beyond 0.8 and large increases in concentration beyond this value produce very little response in absorption readings. At such high concentrations, the relationship between absorption and concentration is no longer well represented by the same model used to predict lower concentration levels. For such a case the solution was replaced to the corresponding graduated cylinder and approximately five millilitres of clean DMF was added to the pooled supernatant. The new volume replaced the previously recorded volume for that sample and the absorption of electromagnetic energy was again recorded. Alternatively, if any sample solution yielded absorption values less than 0.2, the approximate limit of sensitivity, the entire pooled supernatant for that sample was discarded. In such cases a new sub sample of 50mg of the remaining original sample (~100mg) was processed. Once the sample extractions were complete the pasture pipettes were discarded and the graduated cylinders were rinsed with clean DMF in preparation for the next four samples. This method was repeated until all the needle samples had been processed. Using the equations developed by Lichtenthaler (1987) and corrected by Wellburn (1994) the chlorophyll a, chlorophyll b, total chlorophylls and the total carotenoid contents were calculated from fresh weight, solution volume and the three absorption coefficients.

2.2.5 Data analysis

2.2.5.1 Health class establishment

Irrespective of reflectance, the biochemistry of needle samples were considered. To facilitate the comparison between stressed and healthy samples it was required that the data set be segregated into the appropriate groupings. Needle samples were partitioned by means of the binary occurrence of mycelium on the field specimen from which they originated. As is the case for non-destructive sampling, root excavations were not exhaustive and the extents of endotrophic (as apposed to ectotrophic) mycelia were not known. Consequently, two health classes based on the absence or occurrence of the

(30)

fungus would undoubtedly yield groupings of uncertain integrity. In order to compensate for trees that have only minimal exposure to the fungus and are less likely to be responding to infection (endotrophic mycelium), those samples originating from trees exhibiting relatively little fungal growth on excavated roots were omitted from the analysis. The resulting health classes were defined as healthy; those with no identifiable mycelium on exposed roots, and infected, a class which was composed of samples collected from trees that had fungal growth above a chosen threshold on exposed roots. In order to evaluate the influence of colonization, the lower threshold of colonized roots for the infected class was varied at thirty percent intervals and the groupings were graphically and statistically analyzed.

2.2.5.2 Chemistry analysis

Statistical analyses were performed to resolve the separability of the defined health classes and to indicate which of the biochemicals were most responsive to the fungal infection. Varying the threshold of root colonization for defining the infected class permitted the investigation of at which level of root colonization, given the extents of the excavations, tree canopies could be expected to show signs of decline. Moreover, by altering the health class membership, an effort was made to reduce the impact of errors of commission, where healthy/asymptomatic samples could be assigned to the colonized health class. The healthy class was objectively represented by samples collected from trees without above or below ground indicators of colonization. The affected class was defined by a threshold of root colonization. The first affected class was composed of samples acquired from trees that had any amount of mycelium on excavated roots. In intervals of 30%, the threshold of root colonization was increased, where ultimately the last class exhibited 90% colonization of excavated roots.

Site conditions were anticipated to have a significant influence on the health of tree canopies. Soil parameters including organics, substrate parent material (mineral composition, hydrophilic properties), mycorrhizal and microbial diversity are known to have implications on vegetation growth and productivity. In addition, site micro-climate and topography is expected to have an effect on tree health through slope, aspect and

(31)

elevation. Slope has a direct influence on soil moisture retention, while aspect, at sufficient slopes, will impact productivity due to differences in radiation levels. Elevation is known to impart a stress on vegetation composition and health also by controlling moisture and temperature levels. Therefore, the biochemistry data were first stratified by site and analyzed, and then grouped by those sites deemed similar through statistical testing. Conceptually, this stratification protocol would reveal differences in biochemistry attributed to health classes if site differences were such that they obscured health class differences when the data were pooled together.

The effects of site, aspect and health class on moisture and pigment variables were investigated. By generating basic statistics and box plots, significant differences could be quantitatively determined. In order to test the significance of these observed differences, one-way analysis of variance tests were employed with Tukey’s honestly significant difference and Bonferroni post-hoc tests. Where violations of variance homogeneity were identified by the Levene’s test, Welch’s robust test for mean equivalence was applied with the Games-Howell and Tamhane post-hoc test (Hsiung and Olejnik 1991; Welch 1938). In cases where the assumption of normality was violated, determined by the Shapiro-Wilk test, a nonparametric equivalent of the one-way ANOVA called the Kruskal-Wallis test was applied followed by the Mann-Whitney U for pair-wise comparison in place of traditional post-hoc statistics. All statistical analyses were performed using the SPSS software suite.

To further evaluate the legitimacy of distinguishing health classes based on colonized roots, a K-means cluster classifier was run. The classification was used to resolve stressed and healthy samples based on biochemistry, the results of which were compared against root colonization. Additionally, pigment ratios were assessed for their apparent ability to identify stressed samples (Bracher and Murtha 1993; Campbell et al. 2004; Heimler et al. 1989). Subsequent to the analysis of water and pigment contents, lab spectra were processed and analyzed.

(32)

2.2.5.3 Spectroscopy

In the preliminary reflectance analysis, needle spectra from the four canopy aspects were averaged to represent tree reflectance responses (Wang et al. 2007). Specimen reflectance spectra were compared with below ground indicators of

P.sulphurascens colonization. As performed in biochemistry analyses, thresholds of root

colonization were used to generate discrete health classes. The average class reflectance responses were used to illustrate broad spectral regions that provide the potential for distinguishing between the health classes. Sample trees that exhibited evidence of below ground indicators of colonization were compiled into the infected class, while those without sufficient evidence of infection were classed as healthy. Average infected and healthy foliage spectra were generated and plotted along with a difference spectrum.

A continuum removal process was applied to biochemical absorption features to remove the background spectra and wavelength dependent scattering to reduce the effects of spectrally adjacent and overlapping features, and to normalize the spectra (Clark and Roush 1984; Kokaly 2001). The method of continuum removal yields comparable sample spectra that are more closely related to the concentration of the target substance within samples than reflectance spectra (Curran et al. 2001). The continuum removal process begins by establishing the wavelength range that contains the absorption feature. This wavelength interval is used to construct all the continuum removed sample spectra. A linear equation defines the continuum line that joins the extremes of the absorption feature. The continuum line is an approximation of the spectra without the effects of absorption attributed to the target material. The continuum removed spectrum (CR) is then calculated as the ratio of reflectance to the continuum line at each wavelength within the absorption feature. The ratios are subtracted from the value of one, in order to invert the curves. Continuum removed spectra are then normalized to the band depth, this is done by dividing each continuum removed value by the maximum within the curve, the product of which yields curves that range between zero and one. In order to interpret the final curves as normalized reflectance data rather than absorption, the curves were again inverted by subtracting from one (equation. 1). Attributes were then extracted from the

(33)

band depth normalized continuum removed spectra (CRDN), which were used in regression analysis to derive relationships for the prediction of foliar biochemistry. Attributes included the band center, absorption feature symmetry, depth, total area, left area and right area (Lawrence and Labus 2003; Pu et al. 2003).

CRDN = (1-(Ri/Ci)/CRmax)

Equation 1. Where Ri = reflectance at band i, Ci = value on continuum line at band i and

CRmax = maximum depth after continuum normalization.

For comparative purposes the red edge position (REP), was regressed with chlorophyll-a. The red edge position is found as the maximum first derivative (Campbell et al. 2004; Rock et al. 1988). Conceptually, the red edge position is the wavelength location of the maximum positive slope within the transitional wavelengths of chlorophyll absorption and near-infrared scattering. The red edge position has proven to be highly correlated with chlorophyll concentration in laboratory reflectance studies, but because the shift in REP is typically less than 25nm, instruments with 1-5nm resolutions are required (Ustin et al. 2004), otherwise, models (linear, linear extrapolation, Gaussian, polynomials), must be fit to generate the continuous REP data (Cho and Skidmore 2006; Ustin et al. 2009). In addition, a series of vegetation indices that pertain to vegetation health and productivity were applied to the laboratory reflectance dataset. The indices assessed were selected, as they were developed or were related to biochemistry, health and or productivity.

The relationship between biochemistry and reflectance was investigated. In previous research, correlograms have been used to illustrate regions within the electromagnetic spectrum that are significantly correlated with a substance of interest (Blackburn 1999; Curran et al. 2001; Datt 1998). A correlogram is a graph of Pearson’s R coefficient of correlation values on the y-axis and wavelengths on the x-axis. The plot is constructed by computing and then plotting the correlation between reflectance (at a given wavelength) and a biochemical variable. Since the ASD spectra are output to 2150 wavelengths, a correlogram for chlorophyll for example, would have 2150 correlation coefficients

(34)

relating reflectance at each wavelength to the pigment content. The utility of these plots are realized by identifying electromagnetic regions where a biochemical of interest may have strong absorption features. Statistically significant correlations are identified as those values that exceed a threshold coefficient of correlation, which is dependent on sample size (N-2 degrees of freedom) and a significance level. This, illustrates which wavelengths are strongly correlated with the content of a biochemical of interest. The information obtained from correlograms, augmented with health class spectral difference results, suggested which of the sampled chemicals is responsible for the observed differences in reflectance. The difference spectra and correlograms were used for feature extraction, where a wavelength range was identified as a feature of interest. In addition to correlation analysis throughout the spectrum, a correlation matrix was constructed using an extensive vegetation index library of 154 indices.

2.2.5.4 Regression analysis

The estimation of biochemistry from reflectance was a significant objective throughout this study. Although the in-vivo absorption characteristics of the target pigments are well documented for deciduous tree species, the same can not be said for their gymnosperm counterparts (Moorthy et al. 2008). Correlograms were used to suggest electromagnetic regions that were significantly associated with biochemistry and were, therefore, useful in selecting target wavelength ranges for spectral feature analysis. In addition to highly correlated wavelengths and indices, attributes extracted from spectral feature analysis (continuum removal) were used to develop empirical models, which were applied to the prediction of the biochemical composition of samples. Additional, widely available eigenvector-based geometric projection algorithms were applied to derive latent components that were also used in regression analysis. Ultimately, the models that produced the most promising results in regression analysis whether produced from feature attributes, indices, or latent variables were identified.

Towards an operational root disease mapping methodology through lidar integrated imaging spectroscopy

Supervisory Committee

Abstract

Table of Contents

List of Tables

List of Figures

Acknowledgments

Chapter 1

Chapter 2