Manual of Remote Sensing Andrew Rencz, editor Chapter 4 Revised,  March 15, 1997

 4.1 Introduction

Progress in developing more physical understanding has been slow due to the lack of interdisciplinary understanding between plant mineral nutrition specialists, plant ecologists, geologists, and remote sensing scientists.  Nonetheless, if remote sensing is to provide mineral characterization and mapping outside desert areas, there is a critical need to develop collaborations between these disciplines.  Ecosystem based studies, sometimes termed earth system science, offer the potential for better understanding the factors controlling plant abundance and distributions through more holistic ecophysiological models connecting plant function with edaphic and climate conditions (Ovington, 1965; Rodin and Bazilevich, 1967; Parton et al., 1987; Running and Coughlan, 1988; McGuire et al., 1992; Potter et al., 1993; Prentice et al., 1992; Perruchoud and Fischlin, 1995).

4.1.1: Chapter Objectives:  The primary aim of this chapter is to provide earth scientists with a brief review of plant properties that affect the absorption and scattering of radiation in the visible and infrared spectral region. The focus is on providing sufficient understanding of plant structure and function to understand the spectral characteristics that may be useful in earth science and geobotanical mapping of earth resources.  The second aim of the chapter is to introduce several new methods for analyzing remote sensing data that are of particular interest for extracting botanical properties.  As we move into the next century, a wide range of new satellite and airborne sensors will become available and new techniques are needed to analyze these data.  These will include hyperspectral sensors with large numbers of contiguous bands and sensors with bands in new spectral regions, including multi-band shortwave infrared, thermal infrared, radar, and others.  Case studies illustrate many approaches to geobotany and provide perspective on the range of applications.

4.1.2:  Definition of terms:   Geobotany has been defined as the study of plants as related specifically to their geologic environment (Rose et al.,1979) and by Raines and Canney (1980) as a visual survey of vegetation used to define geologic differences in the landscape.  Geobotany is composed of two fields of botanical research:  the study of  the spatial distribution of plants and plant communities as related to geology and the study of vegetation characteristics (physiology, morphology, anatomy, biochemistry) as related to geology.  The first component of geobotany has it roots in studies of plant geography which date back to Darwin (Cain, 1971).  The second component is more closely aligned with biogeochemistry. Geochemistry is the study of the chemical composition of the earth and the physical and chemical processes that have produced their distributions.  The study of geochemistry has been extended to include biological materials and their interactions in relation to earth chemicals.  Although the next two sections describe these fields as separate entities, the remainder of the chapter considers geobotany as a more general term that includes spatial distribution of vegetation and biogeochemistry and geology in the broader context of Earth Sciences.

4.1.3:  Geobotany:  Geobotany includes the study of relationships between soil and geology and the overlying vegetation type and its abundance.  It has origins in the field of plant geography and includes studies on plant distributions related to their holistic environment (Cain, 1971).  Additionally, studies by biologists, geologists and soil scientists, and paleontologiests have provided historical perspective to the geobotanical literature, where geologists and soil scientists have used plant distributions to infer soil and mineral conditions.

Some of the earliest modern ecological literature, e.g., Warming (1892) and Schimper (1898) describe the relationships between plant geography and soils from the perspective of plant function, morphology, and physiology. Warming and Schimper describe plants whose distributions are limited to a particular soil chemistry, like calcareous, saline, siliceous, serpentine, zinc and others.  Warming emphasized the importance of soil factors in plant distribution and coined the terms used today for describing the association of characteristic plant species with particular edaphic conditions:  halophyte, hydrophyte, xerophyte, etc.  While symptoms of deficiency or excess of specific elements have been described by numerous authors, caution must be used in interpreting visual symptoms of plants because different mineral deficiencies or excesses can cause the same symptoms (e.g., chlorosis), some species respond differently to similar exposures, and multple deficiencies may be present.  The books by Walter (1973) and Eyre (1968) provide a good review of the importance of soil chemistry in mineral nutrition and distribution of vegetation.

Most frequently, in North America the term geobotany is used to describe a form of mineral prospecting which relies on characteristics of the vegetation to identify the location and extent of ore bodies (Brooks, 1995a). A long list of indicator species have been identified, including those known to “hyper-accumulate” metals (i.e., tolerate and accumulate high concentrations of heavy metals) are found in Brooks (1995b).  While the field has primarily developed in this century, Brooks cites several early observations.  Roman records show that they were aware that some plant species were good indicators of subsurface water stores.  Australian aboriginals also have used plants to indicate subsurface water stores dating far back into antiquity.  Georgius Agricola in 1556 (Hoover and Hoover, 1950) described a pattern of premature plant senescence and stunted growth of plants above ore bodies.  Brooks (1995c) reports that Caesalpino (1583) associated Alyssum bertolonii with ultramafic soils and Pope Pius II (1614) suggested that Ilex aquifolium might be an indicator of aluminum.   Scandinavian miners used indicator plants like Lychinis alpina (pyrite plant) in the 17th century to locate ore deposits.  In 1979, Brooks et al. analyzed >700 herbarium specimens and identified 79 with high concentrations of copper, lead, or nickel, more than half of which came from known mining areas.

Recently, the application of remote sensing for mineral mapping has renewed interest in how plant characteristics can provide insight into soil conditions. From a broader perspective, geobotany may be considered as an integration of the physical and biological processes of interest to earth science.   Plants and microbes function at the interface between the soil and atmospheric systems, interacting with all phases of the hydrologic cycle, soil development and erosion, biogeochemical cycling, and geologic weathering.  Thus, developing tools for characterization of  vegetation has implications for a wide range of earth science research. Applications for earth science issues is possible, ranging from identification of fracture zones, weathering and erosion, sediment transport, surface and subsurface hydrology, hydrocarbon seepages, soil or water contamination, to others. The capability for synoptic mapping combined with the nearly ubiquitous cover of the terrestrial surface by plants has created a need for forging better understanding of these relationships.  Generally, they are based on understanding the mineral constituents of the soil and the plant’s nutrient requirements and tolerances.  However, plant-soil relationships are not simple;  they are often non-linear with mineral concentration, tissue age, and are strongly dependent on the water holding capacity of the soil, pH, and canopy evapotranspiration (Epstein, 1972).

4.1.4:  Biogeochemistry:  In recent years the importance of biologic agents in geochemical cycling and transport has become more widely recognized.  Plants, animals and soil microbes significantly contribute to the rates of geochemical reactions, fates of weathered minerals, and dispersion and transport processes.  Elements may be vertically transported to the surface due to reworking by soil organisms or uptake by plant roots.  Distribution of elements in the soil may be changed by preferential deposition or accumulation under plant canopies or the interstices between plants.  Organisms mediate the cycling of chemicals that have both atmospheric and geologic phases.  Biogeochemical cycles are the sequences of stages in the transfer and transformation of elements between the lithosphere, hydrosphere, atmosphere, and the biosphere.  Carbon, nitrogen, oxygen, sulfur, and water have cycles of global importance through soil, water air, and biota.  Some elements, like sodium and chloride are non-volatile but are transported by the atmosphere over great distances.  Others, like phosphorous, potassium, calcium and magnesium do not have an atmospheric phase and are only mildly soluble. These are eventually lost from the lithosphere to the oceans through the mechanism of absorption by plants and eventual leaching during decomposition (Epstein, 1972).
 
The concentrations of nutrient elements varies with the plant tissue, age and growing conditions of the plants.  Immobile nutrients like calcium, i.e., those that become fixed in plant tissue and are not subject to relocation within the plant, continue to accumulate throughout the organ’s life and are lost from the plant when the organ senesces and dies.  Mobile nutrients (e.g., nitrate and potassium) are those that may be translocated to another organ within the perennial plant.   In evergreen plants, mobile nutrients may accumulate until close to the end of the growing season (Mooney and Rundel, 1979; Chapin et al., 1980).  Adjusting root uptake kinetics provides a mechanism for regulating nutrient uptake under different soil fertilities. Chapin (1988) argues that root kinetics regulate uptake of mobile ions under all soil conditions and for immobile ions in fertile soils.  For many elements, relative concentrations are determined by stiochiometric requirements defined by plant metabolism.   Nutrients vary in concentration over eight orders of magnitude within a given tissue as illustrated in Table 4.1.   Anomalous tissue concentrations develop when soils have deficiencies or excesses in these elements or high concentrations of other elements.  Over extended periods plants respond to nutrient deficiencies, excesses or imbalances through modifications of growth.  Responses can range from reduced or enhanced growth and biomass production, changes in leaf area or root-to-crown ratios (low nutrient or water status generally leads to increased allocation of biomass to roots relative to shoots), or changes in species distributions (Schlesinger, 1991; Field, et al., 1992).

4.2  Geobotanical Remote Sensing

By 1980 Solomonson et al. reviewed several geobotanical studies to summarize the relationships between leaf reflectance and metal concentrations.  Even at this early stage in remote sensing, their paper showed that the effect of metal accumulation on leaf reflectance varied with the wavelength, metal species (copper, lead, zinc, arsenic, molybdenum, sulfates), and tree species.  A recent review by Brooks (1995d) summarizes a large number of early papers on geobotanical exploration using satellite and airborne imagery.  Milton (1978) measured field spectra of native species in a hydrothermally altered semi-arid region.  Field spectra of plants and rocks were used to construct a MSS band ratio model to map the plant communities of the East Tintic Mountains, Utah (Milton, 1983).  Chang and Collins (1980) and Horler et al. (1980) used laboratory experiments to show that chlorophyll content and reflectance changes when plants are grown in soils with heavy metals.  Horler et al. (1981) provide an early review of the use of plant reflectance to detect metal induced chlorosis. Collins et al. (1980, 1983) were the first to show narrow band spectral changes in airborne data using the 512 band MARK II spectroradiometer.  They observed a “blue shift” due to apparent reduced chlorophyll absorption associated with mineral deposits, at the “red edge,” the region between the red and near-infrared wavelengths around 700 nm, where leaf reflectance increases rapidly. Their conclusions were supported by a subsequent experimental study of mineral stressed plants (Chang and Collins, 1983).  Milton et al. (1983) used the Chebyshev polynomial waveform analysis of Collins et al. (1983) on airborne MARK II data over sites of metal anomalies in a hydrothermally altered porphyry gold system in the slate belt of North Carolina and showed strongest changes in spectral reflectance occurred over an area of high copper, molybdenum and tin.  They caution, however, that other weakly anomalous areas did not correspond to areas of high metal concentrations.  The problem in application of reflectance anomalies to image data is that reflectance depends not only on pigment content but also on plant leaf area index (total leaf area/ground area), percent cover, species composition and canopy architecture, and extrinsic but correlated factors like topography.

4.2.2: Combined Geographic Information Systems Analysis:  Many mines or mineral outcrops or contaminant sources have small point-source locations within a larger geologic and ecologic context.   Often the goal of geobotany is to identify these point sources due to their contrast with the surrounding vegetation and terrain.  Numerous factors create variability in satellite images that are imposed at multiple scales.  Topographic patterns, illumination conditions, and atmospheric composition at the time of measurement are factors that vary at the larger scales of variation. Of the factors affecting local scale vegetation distribution within a regional climate zone, topography, is of critical importance. While the impact of topographic gradients is most apparent at the ecotones or margins of community ranges, it’s affect on the local microclimate--modifying temperatures and net radiation, precipitation regime, diurnal and seasonal cycles--is always present.  Vegetation type (e.g., deciduous, evergreen, or conifer) and life form (herbaceous, shrub or trees) are typically observed to vary at smaller scales in response to topographic conditions. Digital elevation maps (DEM) can be readily combined with image data sets in geographic information systems (GIS) to assess these topographic patterns.  Smith et al. (1990b) developed functions between elevation and mean vegetation cover and surface temperature, then compared the Landsat TM derived estimates of vegetation density and temperature to find vegetation distribution anomalies related to soil conditions and water availability.  Ustin et al. (1996) used the slope, aspect, and elevation data layers in a vegetation classification scheme with spectral mixture fractions.  Similarly, other ancillary data in a GIS may be used to improve relationships using geostatistical methods (e.g., co-kriging) or to train neural nets.

Composition variation in plant communities, the dispersion patterns of plants and the percentage of canopy cover also affect image data at somewhat large or intermediate scales.  Canopy structure and composition, is defined by the leaf area, number of leaf layers, and leaf angle distributions within the canopy, the distribution of different canopy components, specifically the stems, green foliage, and litter, and lastly, the soil and understory vegetation if the vegetation has multiple layers.  While the biochemistry of canopy components affects the scattering and absorption of light, it has a small effect on reflectance relative to other landscape scale variation.  Both physically based and empirical models attempt to account for these properties and their interactions in producing at-satellite and top-of-canopy radiances.

There are a number of methods to examine images for texture or spatial patterns in vegetation distribution.  One example of a spatial analysis that has been used for geological hydrocarbon exploration is lineament analysis (Cetin et al., 1993).  Anomalous linear features in an image are assumed to represent structural geologic features, particularly joints and faults.  These geologic features have been shown to be associated with zones of higher soil permeability, and often have different plant species or changes in plant density growing over these fracture zones.  However, by themselves lineament interpretations are typically noisy and ambiguous because many lineaments in satellite images are cultural features or random associations.  By combining multi-date imagery, or using GIS topographic databses, the reliability of lineament analysis can be improved.

Later in this chapter we introduce wavelet analyses as an image method that utilizes spectral and spatial information and temporal information imbedded in remotely sensed images.

4.3:  Spectral Characteristics of Foliage and Plant Canopies

The foliar absorptions are primarily caused by photosynthetic pigments in the visible spectrum, water, cellulose and other carbon-based compounds in the infrared.  Cell wall, water-air interfaces account for much of the scattering in the infrared (Gates et al.1965; Gates,  1970; Gausman, 1977;  Gausman et al., 1978).  Major pigments include chlorophyll a and b,  b carotene, lutein and xanthophyll cycle pigments (Lichtenthaler, 1987).  It has long been noted that extracted chlorophyll absorption peaks are shifted about 20 nm to shorter wavelengths than observed from leaf reflectances. Seasonal shifts of the long-wavelength edge of the chlorophyll absorption also have been observed in foliage (Gates et al., 1965; Horler at al., 1983) as have stress related shifts, e.g., due to heavy metals (Chang and Collins, 1983, Milton et al., 1983).  Despite 20 years of work, defining the relationship between the red edge and chlorophyll concentration remains an area of active research (Curran et al., 1995, Pinar and Curran 1996).  Partly this is due to difficulties in measuring the suite of pigments and intermediates and conformational changes that occur during handling.  In the intact chloroplast, pigments are arrayed in protein complexes associated with thylakoid membranes in the light harvesting apparatus. Gamon et al. (1990), observed short-term changes in reflectance near the green peak at 530 nm that were due to changes in the distribution of xanthophyll cycle pigments in response to the light environment.  This physiologically mediated rapidly reversible conformation change occurs in response to leaf exposure to high light levels and is part of the plants mechanism for light intensity regulation.

Water dominates leaf absorption in the shortwave infrared (1.0-2.5 mm) region and major water absorption bands occur at 1.45, 1.94, and 2.7 mm and secondary features at 0.96, 1.12, 1.54, 1.67, and 2.20 mm (Knipling, 1970, Wolley, 1971, 1973; Wessman, 1990).  Protein, cellulose, lignin and starch affect leaf reflectance (Marten et al., 1985; Weyer, 1985; Card et al. 1988) but have been more difficult to characterize, even in dry leaves (Peterson and Hubbard, 1992; Curran, 1994; ACCP, 1994).  Spectra of these organic compounds are primarily composed of C-H, N-H, and O-H bonds that have fundamental molecular absorptions in the 5-8 mm region and exhibit overtones and combination band overtones in the shortwave infrared region (Curran 1989; Wessman, 1990).  While absorption spectra of these molecules have been developed (e.g., Card et al., 1988, Barton et al., 1992, Wessman, 1990), they are not identical to the specific absorption coefficients predicted by inversion of the PROSPECT model (Jacquemoud et al., 1996), presumably due to bending and stretching bond vibrations in the intact leaf.

4.3.1.2:  Fluorescence and Thermal infrared spectra

4.3.2:  Variability at the Canopy Scale:  The first attempt to mathematically represent the bulk transfer of radiation in a plant canopy probably goes back to the pioneering work of Monsi and Saeki (1953), who described the transmission of light through a cloud of leaves as an extinction process.  The solution of this problem is a simple negative exponential of the amount of leaf material present in the canopy, a variable related to the leaf area index.  The Kubelka-Munk theory of radiation transfer in layered media was later applied by Allen and Richardson (1968) to the problem of estimating the reflectance and transmittance of plant canopies, and this simple approach thus provided an initial paradigm which ultimately lead to the development of a suite of instrumental approaches to derive the quantity of leaf material in plant canopies from observations of the downwelling radiation field, including fish-eye photography and other sensors with a wide field of view.

Ground breaking results were obtained in the 1960's and 1970's by various scholars.  It will not be possible to do justice to this field and to make an exhaustive review of the many contributions, but we have elected to highlight two outstanding groups. The first  concerns Professor Gates and his collaborators, who largely contributed to the introduction of spectroscopy in plant physiology and fathered the field of biophysical ecology (Gates, 1980).  Another notable case is provided by the Estonian school, lead by Professor Ross, who applied the classical methods of radiation transfer developed in astrophysics to the description of the directional reflectance of vegetation canopies (Ross, 1981). Both of these seminal books contain extensive bibliographies and can be used as introductions to these and related fields.

The development of remote sensing techniques during the 1970's and 1980's, and in particular the launch of space instruments dedicated to the observation of the Earth's geophysical environments, such as the Landsat MSS and, later, the AVHRR series of instruments, stimulated research and promoted new applications. Progress in the monitoring and characterization of plant canopies and ecosystems on the basis of space observations was achieved in various directions, some of which will be briefly outlined below.  It is important to recognize at the outset, however, that the interpretation of remote sensing data can only be achieved by exploiting the variations of the observed radiometric signals with respect to the independent variables that describe the conditions of observation (Verstraete et al., 1996). The five domains of variation that have proven useful so far in remote sensing correspond to the spatial, temporal, spectral, directional and polarimetric independent variables of the radiation transfer equation Gerstl, 1990).  Furthermore, the quality and quantity of information derived from these data depends strongly on the degree of sophistication of the tools used in their interpretation. More advanced models, that better take into account the details of the relevant processes, will lead to more appropriate and more accurate information retrieval.

The following general approaches can be distinguished in the utilization of remote sensing for the identification and characterization of plant canopies:

1. The first concerns the spatial analysis of the data. Indeed, right from the start, imaging space sensors provided data on large areas much faster than could be gathered by a field team. Space instruments also provided a spatial resolution sufficient to imagine, and indeed to implement, applications that could take advantage of this new technology.  Besides purely cartographic applications, even the primitive (by tomorrow's standards) spectral resolution of early instruments stimulated exercises in pattern recognition and land surface type identification.  This capability, coupled with the possibility of revisiting the same sites at least periodically, generated a lot of interest in monitoring seasonal and interannual evolution of land surfaces.  Today, land cover classification and changes studies remain a priority, as witnessed by the recently adopted IGBP Land-Use and Land-Cover Change Science/Research Plan (Turner et al.,  1995), and remote sensing will continue to provide critical data to address these issues (e.g., Townshend, 1992).  The analysis of temporal changes in land surface properties, in particular, has permitted distinguishing different land cover types or the impact of human activities on the landscape (e.g., Lambin and Ehrlich, 1995; Ehrlich and Lambin, 1996).

2. The second is marked by the design and exploitation of simple and largely empirical tools, most notably spectral mixture models and vegetation indices. The former approach assumes that the multispectral observations actually gathered by a remote sensing instrument can be understood as combinations of the spectral signatures of the individual objects present in the scene, called `endmembers.'  Initial studies (e.g., Smith et al., 1990a) assumed that such combinations occurred linearly, but more recent investigations have shown that for canopies, non-linear terms may be important.  Nevertheless, the main outcome of this approach is to provide estimates of the relative importance (often interpreted as coverage) of the elementary surface types in the observed pixels.

For their part, vegetation indices, which are effectively two-endmember linear models, attempt to directly estimate a given plant or canopy property from a suitable manipulation of the spectral measurements. Many such indices have been proposed in the literature over the last twenty years. Rather than listing them or reviewing the many applications in which they contributed, we propose to underscore three critical issues related to the proper utilization of these indices. The first concerns the design of appropriate mathematical formulae to address specific questions.  Most of the vegetation indices proposed so far were arrived at empirically, but a rational approach to the design of optimal spectral indices has recently been proposed by Verstraete and Pinty, 1996). The second issue relates to the interpretation of these indices. Although most proposed formulae attempted to take advantage of the high spectral contrast between the red and near-infrared reflectances of vegetation, only recently has there been an effort to quantify the possible physical reasons for the observed correlations (Baret and Guyot, 1991; Pinty et al., 1993;  Myneni et al., 1995). The third essential aspect of the proper use of spectral indices relates to their evaluation and therefore to the selection of the most relevant formula to address a particular problem.  A general solution to this issue has been proposed by Leprieur et al. (1994),  who propose to rank vegetation indices according to a suitably estimated signal to noise ratio. Here again, the development and exploitation of advanced spectral indices will allow the delivery of more reliable information.

3. Work has also been pursued in a third direction, namely the design and exploitation of models describing explicitly the anisotropy of the reflected radiation field. These so-called bidirectional reflectance distribution function (BRDF) models are conceived to describe how the observed reflectance depends on the particular geometry of illumination and observation, given the structural and optical properties of the simulated surface.  BRDF models fall into two broad categories: either they attempt to simulate the observed anisotropy of the target in terms of measurable physical quantitites (physically-based or causal models), in which case the reliability of the results critically depends on the quality of the model, or they aim at representing the observed effects (empirical or phenomenological models). In either case, the inversion of a BRDF model against a set of directional measurements permits, under specific mathematical conditions (e.g.,  Verstraete et al., 1996), to reconstitute the entire reflectance field and to estimate the directional hemispherical reflectance (albedo) of the surface. The opportunity to critically evaluate simple physically-based BRDF models through inversion has been discussed by Pinty and Verstraete (1992).

4.4:  Vegetation Reflection Models

4.4.1.1:   Ray tracing models:  Among various approaches, only ray tracing techniques can account for the complexity of internal leaf structure as it appears in a photomicrograph. They require a detailed description of individual cells and their unique arrangement inside tissues. The optical constants of leaf materials (cell walls, cytoplasm, pigments, air cavities, etc.) also have to be defined. Using the laws of reflection, refraction, and absorption, it is then possible to simulate the propagation of individual photons incident on the leaf surface. Once a sufficient number of rays have been simulated, statistically valid estimates of the radiation transfer in a leaf may be deduced. The technique has been applied with a number of variants. The first studies were performed at the cell level (Haberlandt, 1914 ; Gabrys-Mizera, 1976), in particular with epidermal cells the shape of which might influence the path of the incident beams: convex cells of some plants act as lenses that focus light within the upper region of the palisade parenchyma which contains many chloroplasts adapted to high light. This phenomenon has been mainly presented as an adaptation to the low light environment on the tropical forest floor (Bone et al., 1985), but Martin et al. (1989) showed that the epidermal lenses of Medicago sativa (alfalfa), could increase absorption of light at low sun angles. Research efforts were also directed toward understanding the transmission path of light through entire leaves: Allen et al. (1973), and afterwards Brakke and Smith (1987) modeled an albino maple leaf by 100 circular arcs and of two media: intercellular space air and cell walls characterized by their index of refraction. The model was used to test the specular and the diffuse nature of reflection at the cell walls. Simulations led to an under-estimation of the reflectance and an overestimation of the transmittance in the near-infrared plateau.  This was demonstrated shortly afterwards by Kumar and Silva (1973) who found that the actual reflectance and transmittance could be better reproduced by adding two more media into the model, cytoplasm and chloroplasts, thereby increasing the internal diffusion. Whatever the approach, the absorption phenomena that characterize leaf optical properties outside the near-infrared plateau has been ignored.  Moreover in all these models, leaves were always described as two-dimensional objects although the three-dimensional structure of these organs is very important to their physiological function (e.g., for CO2, H2O, O2 diffusion) and to light scattering (Parkhurst, 1986 ; Vogelman and Martin, 1993).  To address these problems, Govaerts et al. (1996) used Raytran and successfully simulated the optical properties of a virtual 3-D dicotyledon leaf.  This ray tracing code, primarily designed to calculate photon transport in a plant canopy, requires a 3-D representation of leaf internal cellular structure that conforms to the constraints of plant anatomy and physiology.

4.4.1.2:  N-flux models:  These models, derived from the Kubelka-Munk theory, consider the leaf as a slab of diffusing (scattering coefficient, s) and absorbing (absorption coefficient, k) material. The N-flux equations are a simplification of radiative transfer theory:  the solution of these equations yields simple analytical formulae for the diffuse reflectance and transmittance. A two-flux model (Allen and Richardson, 1968) and a four-flux model (Fukshansky et al., 1991; Martinez v. Remisowsky et al., 1992 ; Richter and Fukshansky, 1996) have been successfully used in the foreword mode to calculate the s and k optical parameters of plant leaves. Yamada and Fujimura (1991) later proposed a more sophisticated version in which the leaf was divided into four parallel layers: the upper cuticle, the palisade parenchyma, the spongy mesophyll, and the lower cuticle. The Kubelka-Munk theory is applied with different parameters in each layer, and solutions are coupled with suitable boundary conditions to provide the leaf reflectance and transmittance as a function of the scattering and absorption coefficients. But these authors went further, interpreting the absorption coefficient determined in the visible region in terms of chlorophyll content. By inversion, their model became a nondestructive method for the measurement of photosynthetic pigments. The leaf biochemistry has been introduced by Conel et al. (1993) who used a two-flux model to study the influence of water, protein, cellulose, lignin, and starch on leaf middle infrared reflectance. However this model wasn't validated. Finally, a very simple model, directly issued from the expression of the reflectance, has been used to estimate the chlorophyll content of wheat leaves (Andrieu et al., 1988).

4.4.1.3:  Plate models:  The first plate model was introduced by Allen et al. (1969) who represented a leaf as an absorbing plate with rough surfaces giving rise to Lambertian diffusion. Parameters are an index of refraction and an absorption coefficient. This model was successful in reproducing the reflectance spectrum of a compact corn leaf characterized by few air-cell wall interfaces. The same authors rapidly extended the model to non compact leaves by regarding them as piles of N plates separated by N-1 air spaces (Allen et al., 1970). The solution of such a system, provided in the last century by Stokes (1862), has been extended to N being a real number: this is the so-called generalized plate model. This additional parameter N actually describes the leaf internal structure and plays a role similar to that of the scattering coefficient s in the Kubelka-Munk model. The PROSPECT model (Jacquemoud and Baret, 1990) has been designed this way: it requires only three parameters, the structure parameter N, the chlorophyll and water contents, to calculate the reflectance and transmittance of any fresh leaf over the whole solar domain. Furthermore, it can be inverted to estimate these leaf biophysical properties. In order to upgrade PROSPECT, a laboratory experiment associating visible/infrared spectra of plant leaves both with physical measurements and biochemical analyses was conducted at the Joint Research Centre in Italy, leading to the LOPEX database (Hosgood et al., 1995).  Two new parameters, the protein and cellulose + lignin contents, permitted the simulation of dry leaf spectra (Jacquemoud et al., 1996). An exhaustive decomposition of leaf biochemistry was attempted by Fourty et al. (1996) but, due to difficulties each of these authors had in estimating protein content, Baret and Fourty (1997) reduced the biochemistry to the specific leaf area.

4.4.1.4:  Other models:  Tucker and Garatt (1977) proposed an original stochastic model where the radiation transfer is simulated by a Markov chain. A black maple leaf is partitioned into two independent tissues, a palisade parenchyma and a spongy mesophyll. Four radiation states (solar, reflected, absorbed, and transmitted) are defined, as well as the transition probabilities from one radiation state to another, between the different compartments. These probabilities are set on the basis of the optical properties of the leaf material. Starting with an initial state vector representing the incident radiation, the steady state is computed by iteratively applying the one-step transition matrix, and yields both the reflectance and transmittance.

Compared with canopy level, only few models directly use the radiative transfer equation at leaf level. The poor information we have on leaf internal structure and biochemical distribution leads to strong simplifications which make such an approach less efficient as compared to more robust equations. In Ma et al. (1990), the leaf is described as a slab of water with an irregular surface containing randomly distributed spherical particles. In Ganapol et al. (1997), it is compared to a homogeneous mixture of biochemicals which scatter and absorb light. Each model was able to reproduce a faithful simulation of leaf optical properties.

None of these models are adapted to needle-shaped leaves. The size of individual conifer needles makes the measurement of their optical properties tricky. In practice, only the infinite reflectance of stacked samples can be performed. Dawson et al. (1997) recently designed a model, LIBERTY, which has the capacity of accurately predicting the spectral response of both dried and fresh stacked pine needles.

4.4.1.5:  Summary:  The simulation of leaf optical properties has resulted in three categories of models:  models based on Monte Carlo ray tracing techniques and requiring a detailed description of the leaf internal structure ; stochastic models using a Markov chain approach ; and finally models derived from the radiative transfer equation where the leaf is considered a slab of diffusing and absorbing material. Each type is able to accurately and coherently simulate the reflectance and transmittance of plant leaves. However, if the goal is to retrieve information on the leaf anatomy or constituents, only radiative transfer models can be inverted. The latter also have the advantage of being easily coupled with canopy reflectance models, connecting the canopy reflectance directly to the foliar biochemicals. Despite this advance in understanding the contribution of leaf biophysical characteristics to canopy reflectance, the search for improved leaf optical properties models is not concluded, both for remote sensing or physiology purposes.

4.4.2:    Models of Canopy Reflectance:  Interest in radiation transfer in plant canopies is therefore relatively widespread, but the particular focus depends on the intended application. The biological, ecological, and agronomic communities are concerned about the interaction between solar light and vegetation through the processes of absorption, photosynthesis and productivity. For their part, the meteorological, climatological and remote sensing communities are more involved with the modeling of scattering processes, the estimation of canopy reflectance, or the albedo of the surface. Since absorption and scattering are two aspects of the same problem, it would be desirable to develop and use common standard models and approaches in both cases. This would facilitate both the exploitation of remote sensing techniques in the first case, and the integration of plant processes in climate models in the second. The rest of this section discusses in somewhat more detail these canopy radiation transfer models, and especially the BRDF models, since they constitute the most developed tools to represent the reflectance of the entire canopy.

BRDF models can be classified, depending on the way they represent the transfer of radiation. A first group of models includes all those that ultimately derive from the classical theory of radiation transfer initially developed to describe the propagation of light in stellar atmospheres (Chandrasekhar, 1944, 1960). Models derived from these theories were first extended to take into account the shadowing effect noticeable in relatively dense media, and applied to the representation of the reflectance of planetary surfaces (Hapke, 1981; Lumme and Bowell, 1981), and, later, of bare soils on Earth (Pinty et al., 1989). The application of these `turbid medium' concepts to vegetation canopies (e.g., Camillo, 1987; Shultis and  Myneni, 1988; Nilson and Kuusk, 1989; Pinty et al., 1990; Knyazikhin and Marshak, 1991) required, however, to take into account the finite size of the scatterers, which are much larger than the wavelength of solar light, and their orientation (Knyazikhin et al., 1992). This lead to the development of hot spot models (Marshak, 1989; Verstraete et al., 1990;  Jupp et al., 1991; Kuusk, 1991).  Another recent development in this direction is the progressive accounting of the effect of the lower boundary condition, namely the role of the soil reflectance (eg., Privette et al., 1995; Gobron et al., 1997). These models, however, remain one-dimensional. Although new techniques discussed below provide alternative solutions to this problem, few authors have tried to extend the classical approach of radiation transfer to describe the anisotropy of the reflectance field in terms of the three-dimensional characteristics of the environment, so the pioneering work of  Myneni and Asrar (1993) should be underscored. The typical physical parameters entering these models include the single scattering albedo and phase function (or, equivalently the reflectance and transmittance of the scatterers), the optical thickness of the medium, and the orientation distribution of the scatterers. Excellent reviews of these approaches have been published, for instance by Goel, 1988; Myneni et al., 1989;  Myneni and Ross, 1991; Hapke, 1993; and Strahler, 1994).

A second category includes BRDF models designed to take into account the `macroscopic' properties of vegetation, in particular the size and shape of the trees or assemblages of plants themselves. This line of research dates back at least to the work of  Brown and Pandolfo (1969).  An important series of contributions was stimulated by Suits (1972), and further work by Suits  (1983) and others (e.g., Verhoef, 1984) lead to the development of the well-known SAIL model.  Early models aimed at estimating the interception of solar light by complex plant arrangements for biological or agronomic purposes (Richardson et al., 1975; Jackson et al., 1979), or to compute the radiation balance of arid regions (Otterman, 1983). Geometric-optic models were further developed by  Li and Strahler (1985, 1986, 1992) and  Li et al., (1995) to estimate the directional reflectance of these structured surfaces. Geometric models typically describe the reflectance of the surface in terms of the dimension, shape, orientation and average optical properties of the scattering volumes and surfaces considered (boxes, cylinders, cones, etc.).

A third class comprises a whole range of models exploiting recent advances in computer graphics and visualization techniques. Early approaches involved ray tracing through a series of boxes affected by average optical properties (e.g., Kimes and Kirchner, 1982), while later developments relied on the explicit representation of the position, shape, orientation and optical properties of all the relevant scatterers in the scene of interest, with such methods as Constructive Solid Geometry or L-systems (Ross and Marshak, 1988; Goel and Rozehnal, 1991; Govaerts and Verstraete, 1994). The radiation transfer problem itself is then solved through the computation of the relative contributions of each object to the reflectance of all others in the case of radiosity methods (Goel et al., 1991; Borel et al., 1991; Gerstl and Borel, 1992), or through the propagation of a large number of individual rays in the case of Monte Carlo Ray Tracing techniques (e.g., Govaerts, 1996; North, 1996). Both approaches have their own merits, each offers the potential to describe the reflectance of complex and quite realistic scenes, and both can be demanding in terms of input data to set up the scene and in computing resources to generate the solution. The main applications of these models include the establishment of high quality standards against which other simpler models (Govaerts and Verstraete, 1995) or spectral indices (Goel and Qin, 1994) are evaluated, as well as sensitivity studies to determine the exact role of particular plant canopy properties in the overall reflectance field (Ross and Marshak, 1989). For completeness, it should be added that any of the methods mentioned above can be further exploited with artificial intelligence techniques, as proposed by Kimes et al., 1994. Research in this direction is still in its infancy, but significant advantages will most probably be derived from further effort in this area, for instance for the purpose of land cover classification.

The last important class of BRDF models includes empirical functions capable of representing the overall shape of the observed reflectance fields in terms of parameters not associated with any particular physical meaning.  Work in this direction also derives from research initially performed in a planetary context, since Minnaert, 1941 first proposed a simple trigonometric formula to describe the variations of reflectance of the Moon. Other models and developments in this category include Walthall et al., 1985;  Roujean et al., 1992 and Rahman et al., 1993a.  The latter was further developed by Engelsen et al., 1996 to support the analysis of MISR data, while the former, together with other models, has been formally expressed as a linear kernel and packaged in a way suitable for the analysis of MODIS data (Wanner et al., 1995. The main advantages of these empirical models are that they may, in principle, be applied to a wide variety of surface types, and that they may provide a simple and efficient solution to the estimation of surface albedo. However, they cannot be used to derive a fundamental understanding of the physical processes responsible for the observed anisotropy, nor can they be inverted to retrieve information on the physical state variables which condition the measured signal.

Many of the models surveyed above can and will be further improved in the near future. The main research issues at the time of this writing, however, concern the proper representation of the boundary conditions, and in particular the accurate accounting of the multiple scattering processes within the canopy as well as between plants, the soil and the atmosphere.  Coupled canopy-atmosphere models have already been proposed (Kriebel, 1978; Tanre et al., 1983; Myneni and Asrar, 1993; Rahman et al., 1993b; and Liang and Strahler, 1993, 1995).  An alternative approach is to incorporate some of the physically-based 1D or empirical models cited earlier as a selectable lower boundary condition in atmospheric radiation transfer codes, as was done by Vermote et al., 1995 with the popular 6S atmospheric radiation transfer code, but much more needs to be done to truly integrate all relevant physical processes into a coherent, comprehensive model.

Further research will be motivated by parallel developments in field instrumentation to characterize the anisotropy of natural surfaces (Deering and Leonoe, 1986; Deering, 1989) and by the upcoming generation of Earth Observation satellites to be launched over the next few years. The data generated by these advanced space sensors will certainly contribute to an increase and a diversification of user requirements, as well as stimulate new expectations regarding the exploitation of remote sensing data in a wide range of applications. This unique situation will create both challenges (because a number of instruments may have similar characteristics or comparable performances) and opportunities, such as the synergistic use of multiple sensors.

4.4.3:  Summary:  Environmental resources, contamination and pollution, and impacts of anthropogenic and climate induced changes are among the concerns of earth sciences today.  Remote sensing, at all scales from the laboratory to space, has significant potential to aid monitoring and detection efforts. This review has attempted to provide prospective on the history of the use of vegetation properties for earth sciences mapping and illustrate some current approaches to data analysis that may become more widely used as we move from qualitative to quantitative spectral analysis approaches.  We have provided a brief review of physical edaphic factors that affect plant species distributions and growth potential.  The actual relationships between vegetation characteristics and soil properties are often difficult to resolve with respect to understanding which of many potential interacting factors is significant in a particular locality.  Thus, both the physical relationships and the image processing steps are subject to errors and misinterpretation.  We have provided examples over scales ranging from leaf modeling to regional analysis of satellite images.  Resolving the significant spectral and spatial properties over the full range of environmental scales is critical to successful interpretation of remote sensing data for addressing problems in earth sciences.  Greater attention to combining spatial and spectral information, topographic data and other ancillary datasets may reduce ambiguity in image analysis and provide new insight into ecosystem processes and interactions.  We illustrate several recent techniques for image processing, from spectral mixture analysis and wavelet modeling, to radiative transfer models of leaves and canopies using a variety of approaches.  While the latter models may be to complex for for most earth science applications, these models provide greater confidence for linking empirical models to physical processes.  The large number of new satellite sensors, with improved spatial, spectral, and temporal resolution, should open many opportunities for integrated earth science assessments.

4.5:  Data Analysis:  Remote Sensing Analysis and Applications of Geobotanical Techniques

The development and evolution of geobotanical parameters from remote sensing data has been based solely on obtaining positive correlations to local “ground truth” measurements rather then from first principles.  The ensemble of established remote sensing techniques have a number of inherent embedded assumptions that have negative implications for extendibng solutions (i.e., solutions outside the region of model calibration). Consequently, the remote sensing literature is crowded with examples of non-repeatable results.  One example is that vegetation endmembers and /or feature classes can be uniquely identified by multispectral measurements (i.e., that each object class has a unique spectrum), another example is that validation using local “ground truth” measurements (e.g., from a field experiment or observation of a single image), ensures that measurements are extendible over time or to regional scales. This focus has limited the extendibility of results because the solutions are non-unique (i.e., the analytical framework was not immune to the many factors affecting radiance measurements unrelated to the property of interest) when applied to conditions that are not identical to those from which they were derived.  Despite considerable interest in scaling problems, there has been little advancement in understanding the scale relationships of laboratory spectral measurements and image pixels.

Increasingly, earth science research involves large-scale regional-to-global estimates of biophyscial processes like biogeochemical cycling, desertification, erosion, or air pollution.  For addressing these problems, we need to apply newer spectral detection approaches. Among these, wavelets and neural networks provide a means of validating remote sensing techniques that extend beyond accuracy assessments using local “ground truth” measurements. While early sensor technology, e.g., MSS, had only a few broad bands and limited radiometric resolution, these empirical techniques were reasonable approaches to data analysis.  Now, however, detector technology has matured, spectral and spatial data can be obtained (at least with airborne sensors) at almost any resolution, data quality has improved, as has our ability to calibrate radiance to surface reflectance using physically based atmospheric radiative transfer algorithms (e.g., MODTRAN or 6S).  We must now develop analyses that can transform the data to geobotanical parameters that filter out the non-relevant factors that cause spectral change. For a transformation to be effective it must satisfy two objectives: 1) be positively correlated to the parameter of interest and 2) be immune from extrinsic factors causing a response that mimics a given parameter. Verstraete et al. (1996) formalize these arguments and describe 10 approaches for developing the models capable of relating measured radiance to the physical parameter of interest.

4.5.1:  Spectral Methods:  Many spectral methods employed in vegetation analyses are those described elsewhere in this volume, especially in Mustard and Sunshine.  Others, like band ratio vegetation indices are widely used in remote sensing and a number of reviews are available (Huete et al., 1994, Baret and Guyot, 1991; Verstraete and Pinty, 1996; Verstraete et al., 1996). We will highlight several newer spectral and image analysis techniques that are under development and hold promise for addressing the primary needs of geobotanical image analysis.  Many computationally simple methods work adequately where analysis goals are focused on visual interpretations of geologic structures e.g., lineaments, presence or absence of vegetation, or major seasonal conditions (like wet vs. dry season or large interannual changes).  Where these analyses often fail is in detection of minor landscape changes, where gradients are important, or spectral features associated with the information content are weakly differentiated.

4.5.1.1:  Vegetation Indices:  One of the earliest procedures developed for identifying and enhancing the vegetation contribution to remote sensing pixels has been some form of a vegetation index (VI), a ratio between the red and the near-infrared spectral regions. Tucker et al. (1986) showed that global composite AVHRR vegetation index images exhibited monthly latitudinal changes in greenness that fluctuated in synchrony with measurements of atmospheric CO2 concentration, demonstrating a correlation with planetary biospheric activity. However, even the Normalized Difference Vegetation Index (NDVI), probably the most widely used remote sensing index, has been carefully validated only at a few sites (Myneni et al., 1995; Verstratete et al., 1996), e.g., the Kansas FIFE field campaign (Sellers et al., 1995; Friedl et al., 1995), MAC VI (Qi et al., 1995a), and Jasper Ridge (Gamon et al., 1993). Many questions remain uncertain, such as whether vegetation indices are equivalent when extracted at different spatial, temporal and spectral scales (Qi et al., 1995a).  Jackson and Huete (1991) review methods of calculation, including digital counts, satellite radiances, apparent or surface reflectances.  The ratio can be calculated from any sensor having red and NIR spectral bands but each yields different results for the same surface conditions (Huete et al., 1994).  In fact, despite the many studies, precisely what is measured by VI’s is uncertain.  Although clearly correlated with chlorophyll absorption by foliage, it has been related to many plant properties including LAI (Qi et al., 1995a), percent green cover and biomass (Gamon et al., 1993, 1995), productivity (Weigand and Richardson, 1987), and APAR (Asrar et al., 1984) and biophysical properties like photosynthetic capacity (Sellers, 1985).  In fact, VI’s are sensitive to the life form (e.g., grasses, shrubs, or trees), community composition, and factors like canopy architecture and plant distribution, which, with the autocorrelation among plant properties, accounts for some of the many plant attributes VI’s are correlated with.  In recent years a number of modified vegetation indices have been developed to minimize some extraneous factors affecting the measurements.  The first of these was the NDVI, designed to reduce the influence of variation in illumination conditions.  Subsequently vegetation indices were developed to reduce soil brightness and atmospheric variations (Huete and Jackson, 1987; Kaufman and Tanre, 1992; Pinty and Verstraete, 1992; Huete et al., 1994).  Sun-view geometry also affects VI’s (Epiphanio and Huete, 1995; Qi et al., 1995b).

4.5.1.2:  Red Edge Detection:  The “red-edge” is the reflectance inflection point observed in the spectrum of green plants at the transition between the visible and near-infrared wavelengths.  It is the long-wavelength edge of the chlorophyll absorption feature and has been used to describe the variation in leaf and canopy chlorophyll concentration.  This boundary point in reflectance typically occurs between 690-740 nm in fresh leaves and is determined by the interaction between chlorophyll absorption of red light and the internal scattering processes in the leaf (Gates et al., 1965; Horler et al., 1983; Curran et al., 1990).  Typically the wavelength of this inflection is determined by calculating a first or second derivative but more recently, Chappelle et al. (1992) developed a ratio based analysis.  Increased chlorophyll concentration causes a deepening and broadening of the chlorophyll absorption feature while stresses causing chlorosis usually increases reflectance and narrows the absorption feature, making the inflection point appear to move to longer or shorter wavelengths, respectively (Horler et al., 1983; Curtiss and Ustin, 1989).  The relationship has been utilized to detect symptoms of mineral deficiency (Horler et al., 1980, Milton et al., 1991), air pollution damage (Ustin et al., 1988; Rock et al., 1988; Hoque and Hutzler, 1992).  Using the red edge to quantify environmental stress can be difficult and several authors have noted red-edge changes related to leaf age,  developmental stage, and canopy architecture that were partially independent of pigment composition (Gates et al., 1965; Collins, 1978; Shutt et al., 1984; Vanderbilt et al., 1985, Ustin et al., 1989; Rondeaux and Vanderbilt, 1993).  These effects could be due to changes in the light harvesting mechanisms, e.g., altered chloroplast structure or changes in light scattering and refraction, such as cell wall thickness or chemistry.  The relationship can also be affected by the presence or absence of other pigments, such as the non-photosynthetic leaf pigment anthocyanin (Curran et al., 1991).  Railyan and Korobov (1993) examined changes in the red edge of Triticale (hybrid of wheat and rye) over the growing season and found that the red-edge wavelength varied with developmental stage.  At earlier growth stages, the feature moved from shorter to longer wavelengths (a shift of about 5 nm was observed) while later they found a shift back to shorter wavelengths.  This type of developmental pattern was also observed in pine needles of different ages (Ustin et al., 1989), red spruce and eastern hemlock (Rock et al., 1994) and in Loblolly pine and Norway spruce (Ustin, unpublished data).  Quantitative prediction of pigment composition from the red edge is complicated by interactions among the pigment complexes and the bound proteins in in situ chloroplasts (Curtiss and Ustin, 1989; Railyan and Korobov, 1993).  Curtiss and Ustin (1989) found broadening of the chlorophyll absorption band (a red-edge shift to longer wavelengths) in Ponderosa pine needles in sites having greater atmospheric ozone exposure. This pattern mimicked increased chlorophyll concentration, except that red reflectance increased under ozone exposures and decreased under higher chlorophyll concentrations.  They attributed ozone induced reflectance changes to an apparent increase in the disorder of the absorbing medium, a pattern consistent with observations that an early sign of ozone injury is granulation of the thylakoid stroma in chloroplasts.  These factors, and the correlation between chlorophyll concentration and leaf mass or area at the canopy scale, make remotely sensed detection and prediction of pigment concentration or state of health of a canopy difficult to interpret.  Nonetheless, within a vegetation type, spatial patterns in the red edge coupled with greenness patterns, may be useful to identify areas of potential vegetation stress that could be indicators of soil conditions or contamination.

4.5.1.3:  Multi-Temporal Vegetation Indices:  Sometimes simple statistical or mathematical approaches are sufficient to identify geobotanical sites because of their large areal extent or the impact of edaphic properties on plant growth.  The analyses may be computationally simple but rely on expert knowledge of the interpreter for understanding the geology and ecology of the area.  Analytical complexity arises when multi-temporal data sets are used to resolve mineralized locations.

4.5.1.4:  Continuum Removal Feature Mapping:  One method for identifying and mapping an absorption feature, such as are associated with many minerals in imaging spectrometer data is to remove the convex hull continuum from the observed spectrum (Clark and Roush, 1984).  The continuum is the generalized shape of the spectrum in the absence of the particular narrow-band absorption feature.  This background or continuous spectrum is due to the combination of all the scattering and absorbing characteristics of the material.  This approach works well for comparisons to spectral libraries when the continuum is removed by dividing both library and pixel (unknown spectrum) by the continuum reflectance (see Clark or Mustard and Sunshine this volume).  The technique has been applied to mineral identification (Clark et al., 1990a, b) using the USGS spectral library in the Tricorder algorithm (Clark et  al., 1993).  These analysis tools are available from their internet site (http://speclab.cr.usgs.gov/).  Clark et al. (1995) apply the methodology to the detection of plant stress and species mapping in the San Luis Valley, Colorado, where they were able to detect small spectral shifts of the “red edge” for crops in various stages of senescence.

The continuum removal method was calculated and regressed against measured leaf water content to estimate of water content of leaf, canopy and AVIRIS datasets for chaparral communities in the Santa Monica Mountains, California (Ustin et al., 1997).  The sensitivity to wildfire in the chaparral varies with canopy water content, among other factors like species/community composition and biomass.  Measurement of spatial and temporal variation in water content is a potentially significant step for improving fire hazard assessments of wildlands like this, which are border the Los Angeles urban region.

The continuum removal technique has also been applied to a study of water content variation at canopy and AVIRIS image scales in salt marshes along the northern shore of the San Francisco Bay estuary (Zhang et al., 1997; Sanderson et al., 1997).  The canopy water content of the dominant species varies spatially with interstitial soil salinity. Open water does not express the narrow-band absorption feature of the plant canopy (just the continuum) so the technique is insensitive to water bodies in the marsh.  A simple interpolation of field sampling points revealed a spatial pattern that approximated the AVIRIS pixel distribution (Zhang et al., 1997).  The accuracy and validity of the water content estimate was investigated in a detailed scaling exercise that compared field spectra and water content samples measured in the salt marsh and AVIRIS derived estimates of water content (Sanderson et al., 1997).  This study used a variogram to describe the spatial correlation structure of the canopy water content.  Then ordinary kriging estimates were calculated over blocks that corresponded with the co-registered AVIRIS pixels.  In this case, co-kriging with more sparsely sampled canopy water content measurements did not improve the model.  The relationship was applied to the larger Petaluma River salt marsh, an area approximately 3 km x 6 km.  This technique is appropriate to consider when there is reason to expect a change in a known abosorption feature caused by a geobotanical anomaly or to detect changes in pigment composition or water content of the plant canopy.

4.5.1.5:  Spectral Mixture Analysis and Residuals:  Linear spectral mixture modeling (SMA) is useful for vegetation mapping because of the continuous variation in pixel composition at the scale of many satellites.  Multiple scattering phenomena in plant canopies creates a nonlinear mixing problem as described above.  However, a number of studies have examined linear subpixel separation of vegetation and soil spectral components and obtained results in reasonable agreement with supporting field data (Adams and Adams, 1984;  Huete, 1986, Smith et al., 1990a, b).  A description and review of the methodology is found in Mustard and Sunshine (this volume). The procedure assumes that a pixel is a linear combination of the proportional spectral contribution of each of the components.  The assumptions of this approach contrast with classification methods which assume no within-class variance.  While endmembers (the pure spectral types or classes) in SMA may be similar in concept to traditional classification, they are not constructed the same way.  That is, endmembers can be considered basic building blocks from which pixels vary in the proportions of common environmental components (soil, vegetation, etc.).  Classes defined in clustering algorithms are generally determined by other criteria, e.g., taxonomic category (serpentine, volcanic tuff, corn, soybeans, etc.).  Variation in proportions of the scene elements, specifically, classes of green foliage, soil, and litter can reveal important patterns related to soil processes.  For example, Smith et al. (1990b) found that the percentage of green foliage could be used to calculate the spatial distribution of evapotranspiration rates for the Owens Valley, California.  The valley, about 90 km by 40 km, is predominantly semiarid shrubland on the alluvial fans, with riparian corridors along perennial stream channels and a large halophytic community on the valley floor.  Vegetation abundances were generally less then 40% cover, and were to first order, related to soil moisture variation, salinity, soil types, and past disturbance.  One feature of SMA is that it minimizes the impact of albedo variation due to topography and maximizes the spectral shape information in the spectral signature.  One feature that has been noted is that plant dispersion and biomass contribution to the shadow fraction, which in the absence of topography could provide an additional source of information about the structure and growth dynamics of the vegetation (Smith et al., 1990a,b; Ustin et al., 1993).

Generally acceptance of the SMA model is dependent on the endmember fractions producing a close fit to the measured pixel reflectance (or radiance).  This is estimated by the residuals, often the RMS error, which can be displayed as a gray scale image and areas of high error can be visualized in their spatial context.  This information has implications for locating and mapping anomalous vegetation growth or small mineral outcrops.  If a hyperspectral sensor is used, the specific spectral bands having high residuals can be examined for absorption features associated with particular conditions.  Sabol et al. (1992) examined the separability of different materials to better understand conditions when different spectral components can be detected. The primary limitation of simple mixture models is the effect that variability in the spectral endmember has on predicted endmember fractions.

Roberts  et al.(1996) found that for many pixels, reflectance is adequately modeled by combinations of only two or three endmembers.  In fact, solving for more endmembers in such cases causes instability in the predicted fractions.  Roberts et al. (1997) developed a modified SMA procedure that focused on variability within potential endmember classes.  In this form of SMA, instead of finding the four or five endmembers that provide the best overall fit for all pixels in a scene, the approach is to search a spectral library for the two endmember combinations that best fit each pixel.  All pixels that are unsatisfactorily modeled (i.e., have high residuals), are re-analyzed for three endmember combinations.  Few pixels need more than four endmembers to account for the subpixel mixing.  Once the SMA steps are complete, images can be produced to show which endmember spectra were chosen for each pixel.  These new images produce maps that that approximate the vegetation community distributions for the area.
 
4.5.1.6:  Foreground/Background Ananysis:  Foreground background analysis (FBA) a recent advance in remote sensing techniques (Pinzon et al., 1997; Smith et al. 1994; Ustin et al., 1997; Harsanyi, 1993; Harsanyi and Chang, 1994)  is equivalent to a single neuron in a neural net.  Unique to this technique is the training of the neuron to satisfy the two objectives stated above, where the first objective (correlation to the parameter of interest) is defined by the foreground and the second objective (immunity from extrinsic factors) is determined by the background.  The initial development of FBA examples are limited to the spectral dimension and do not consider the spatial context. FBA can be applied to give the exact same results as residual analysis from SMA, thus SMA is a special case of the FBA.

Pinzon et al. (1997) show how spectral features associated with foliar chemistry, like water nitrogen, and carbon compound content can be identified in the fresh leaf spectrum. In this analysis, the background is defined as the set of extrinsic factors and the foreground as the objects of interest.  For example, in the case of NDVI, the background would be those effects caused by changes in canopy architecture, spectral differences between plants, and the effects of soils of different brightness and elevational differences arising in steep topography.  Pinzon et al. (1997) adopted a hierarchical procedure to reduce the nonlinear dependencies of the spectral response.  Their procedure relied on the spectral variance to provide an initial separation of the spectra into classes that corresponded with different ranges of the chemical variable.  Following this initial classification (which linearized the problem), they trained new vectors for each of the spectral ranges and these were used to predict chemical concentration of the foliage samples.

4.5.1.7:  Wavelets and Neural Networks: An integrated approach:  Despite questions of consistency and extendibility, parameters derived from remote sensing measurements are used at an increasing rate as inputs to and validation for large scale biosphere models such as CASA and SiB-II (Potter et al., 1993; Field et al., 1995; Ruimy et al., 1996; Sellers et al., 1996).  Wavelet decomposition techniques provide an analytical framework to separate and quantify the effects of scale that have only been superficially speculated about (see Mustard and Sunshine, this volume, for a discussion of these techniques). Parameters computed using wavelet decomposition techniques provide increased immunity to extrinsic factors (e.g. atmosphere, viewing geometry, spectral variability, surface composition, etc) by focusing on specific temporal, spectral, or spatial scales distinct for the parameter of interest.  Similarly, neural networks, when properly trained also have increased filter capacity over the past techniques of SMA, classifiers, and indices. The integration of wavelets and neural nets allow models which have a much better ability to separate processes that cause similar effects in spectral measurements.

The continuum removal algorithms of Clarke and Roush (1984) and above, are similar to fine spectral scales resulting from wavelet decomposition of images in the spectral domain.  The advantage of using wavelets is that we can develop solutions specific to the spectral scales. Unlike minerals which have narrow band absorptions, the spectral scales from vegetation are quite different.  The spectra of plants typically have variability at larger spectral scales than most minerals, with much of the variability controlled by the plant phenology and condition.  Additionally, at the smaller scales plant species are more consistent than minerals since plants contain similar metabolic constituents while mineral chemistry is varied.

An overall evolution of simpler image processing techniques to wavelets and neural networks is necessary to move remote sensing from local observational studies toward an operational science that takes advantage of the non-uniqueness of solutions in remote sensing images. The approach to training neural networks is the key to satisfying application objectives. Constrained energy minimization (CEM) can be viewed as a neural net problem in the simplest form. FBA and CEM techniques are both elementary neural nets comprised of single neurons that differ in the training of a finite impulse response (FIR) filter.  In both cases the filter is applied in the spectral domain.  Sub-band coding is an electrical engineering term that refers to filters that utilize information in sub-bands. For remote sensing applications, these may be spatial, spectral, or temporal scales that have potential for the extraction of geobotanical parameters.

4.5.1.8: Summary:  The presence of neural networks and wavelets generalize the analytical framework of many individual techniques developed and used in the past.  These new frameworks provides the capability to generate multiple solutions that trade contrast in parameter detection to obtain insensitivity to variable background factors.  This trade-off is necessary to make parameters extendible beyond the local study.  The training of networks based on remote sensing data is crucial to obtaining accurate regional estimates of parameters.  The goal is not to use the most complicated network to develop solutions, but to find the simplest network necessary to achieve a specified level of extendibility.

4.6:  Case Studies

Case Study 1: Vegetation responses to heavy metals in Quebec (Bélanger /Rencz)

The natural vegetation of the region has been characterized by Marie-Victorin (1964) as deciduous boreal forest.  The dominant trees types include Picea mariana (black spruce), Picea glauca (white spruce), Abies balsamea (balsam fir), Betula papyrifera (white birch) and Populous tremuloides (trembling aspen).  The geobotanical studies in this area were designed to investigate the physiology, morphology, and pathology of the dominant plant species.  Several  species, including coniferous and deciduous trees, were selected for biogeochemical analyses.  For each species, foliar samples were analyzed for trace elements and chlorophyll content.  Correlation coefficients were calculated to illustrate relationships between levels of elements in the soil and in plant tissue.  There was a relatively high correlation for several elements in several of the tree species, notably for sugar maple and white birch, whereas several plants displayed no obvious correlation notably, trembling aspen (Table 4.2).

Generally, (Bélanger, 1988) concluded that correlations between plant samples and soils were relatively weak for Cr and Co, but slightly higher for Ni, Mg and Ca.  However, he notes that there was variation between species.

The influence of heavy metals on leaf development was monitored and showed some relationship between chlorophyll and element concentrations in leaf tissue (Rencz et al., 1980) as shown in Table 4.3. Subsequent studies generally found that chlorophyll content was lower in the anomalous areas than the background area and consistently related to low calcium and excessive magnesium. Ca/Mg ratios are in the 1-20 range for healthy vegetation (Walker et al., 1955), while anomalously low values around 0.5 were found in birch, aspen and red maple (Bélanger,1988).

The utility of remotely sensing this biogeochemical anomaly was investigated in several studies that included ground measurements, airborne measurements and LANDSAT MSS.  The MSS analyses monitored variations in biomass during the growing season using a vegetation index.  Additionally, delay in leaf flush (Bell et al., 1985) and premature leaf senescence (Schwaller and Tkach, 1980) were observed.  Biomass indices were calculated for four images: June, August,  September, and October.  The biomass index used a simple infrared/red ratio.  Although a straight comparison of the index identifies tonal differences between the dispersal train and background conditions, this approach is insufficient to monitor vegetation types or chlorophyll content because local conditions can vitiate spectral signatures of vegetation. To monitor variations in biomass index and possible anomalous behaviour of certain species, scatter diagrams and corresponding maps were plotted to show the relationship between biomass for June versus August, August versus September, and August versus October.

A scatter diagram of biomass indices from the June/August dates shows the relationship between early, normal, and late vegetation flush. Early flush is characterised by relatively high production of chlorophyll in early June (index greater than 17) and average production of chlorophyll in August (index 12 to 18). Late flush corresponds to vegetation in which the  biomass is low in June but high in August. High biomass index for both June and August correspond to hardwood deciduous species, such as sugar maple and birch, the spectral signatures of which are not affected by stress conditions. A map of late versus early and normal flush (Fig. 4.3) shows a close relationship between the dispersal train and late flush whereas early flush and high biomass index vegetation are located mainly outside the dispersal train.

Bélanger (1988) concluded that while Landsat MSS was effective in describing the geochemical anomaly in the Thetford Mines area, their result was dependent on multi-date cloud-free imagery and the fact that the target zone was large and enhanced by glaciation dispersed mineralized debris extending over a distance of at least 70 km.
 

Case study 2:  Combining digital elevation data with remotely sensed  imagery for geobotanical mapping

Geobotanical remote sensing remains challenging because the substrate is generally not the primary factor in determining the ecology of a particular location.  The regional climate gradient, with local factors such as micro-climate and site drainage, and stochastic elements like disturbance and regeneration, have confounding influences on vegetation patterns.  Consequently, ancillary data can be extremely useful as an aid in reducing uncertainty associated with very weak geobotanical associations.  DEM data is one of the most often used ancillary data sources, because elevation is so important in determining local climate, as are aspect and slope on local micro-climate and site drainage, soil development and weathering, and vegetation distribution.

In a study of the East African Rift, Warner et al. (1989) found that NOAA ETOPO-5 DEM data improved modeling remotely-sensed vegetation distribution, and the identification of anomalous regions related to tectonic provinces.  The vegetation distribution was determined by integrating six NDVI AVHRR multi-temporal composite scenes (February, June and September) from 1984 and 1987 (Figure 4.4, left panel).  In contrast to the severe drought suffered by most of East Africa in 1984, 1987 was characterized by near normal precipitation.  A second order general linear model was developed, relating elevation, latitude, longitude to the NDVI data.  All the independent variables had a greater than 99 percent significance in the model (Figure 4.4 middle).

The model residuals were useful to identify sites where local factors are more important than elevation and the regional climate gradient in determining vegetation growth (Figure  4.4 right panel).  In some places, such as along the coast of the Indian Ocean, this factor can be related to local sources of moisture.  However, in many cases the residuals are strongly correlated to geology.  Large negative residuals were found in the geologically young Lake Turkana rift basin, as well as on the Pleistocene-Holocene shield complexes and flows, where soil development was limited due to the geologic youth of the area.  In these regions the vegetation is anomalously sparse.  In contrast, the more clay-rich sedimentary units and the older Miocene-Pliocene basaltic shield complexes support a more dense vegetation than predicted by the model, and are therefore associated with large positive residuals.  When slope and aspect were included in the model, they did not improve accuracy and were not statistically significant.  Yet, the observed pattern is one in which steep slopes, particularly those seaward facing, have greater orographic precipitation. Because the slope and aspect are calculated over a matrix of adjacent pixels this decreased  the spatial resolution below the smaller scale topographic relationships.  In addition, aspect cannot be considered in isolation from slope, since it is only on the steeper slopes that aspect is important.  These two factors suggest that geobotanical methods that incorporate elevation, slope and aspect as “logical channels” (Strahler et al., 1978), will not exploit all the information present in the topographic data.

One method to improve topographic data is the TOPOVEG algorithm, developed for analysis of DEM data, Landsat TM imagery, and aeromagnetic geophysical data (Warner et al., 1994).  The study site for the algorithm was Quetico Provincial Park, Ontario, Canada (Figure 4.5, upper panel).  Most of this region is underlain by metamorphosed granites and granodiorites, which surround elongate units of metamorphosed mafic lithologies, principally hornblende amphibolite, with occasional seprentinized dunite, gabbro, gabbornorite and lherzolite.  Economic deposits, including gold, silver, copper, nickel and platinum group metals are associated with the mafic rocks.  Weathering of the aluminum-rich granitic rocks results in soils that tend to be more acid than those derived from the mafic units.  Quetico Park is on the southern ecotone of the boreal forest, and is dominated by jack pine (Pinus banksiana), black spruce (Picea mariana), aspen (Populus tremuloides) and paper birch (Betula papyrifera).  Southern species such as black ash (Fraxinus nigra) and American elm (Ulmus americana) are generally limited to more fertile sites.

The TM data were converted to nPDF Deciduous Vegetation Index values (Cetin et al., 1993), a measure of the pixel’s position on the spectral mixing line between coniferous and deciduous vegetation (Figure 4.5, middle panel). The nPDF is a suite of programs centered around a data reduction technique in which the original image bands are combined, based on their Euclidean distance from selected corners of the multispectral data space.  A nPDF index developed, based on the distance of each multispectral pixel from the corner comprising the minimum in bands 4, 5, and 7, and the maximum in the remaining bands.  This has the effect of maximizing the difference between the pixels with the geobotanical anomaly, compared to the surrounding areas.  Thirty meter elevation data was interpolated from digitized 1:50,000 topographic maps. These areas are displayed in purple and blue on the color map.

A landscape classification was developed from the digital data, in which slopes greater than 3 degrees were classified into either north-and east-facing, or south- and west-facing groups and slopes less than 3 degrees were classified by drainage characteristics, yielding ecologically relevant site classes.   Similarly, the geophysical data were converted from raw measurement units to significant geobotanical classes. The aeromagnetic geophysical data were then classified into anomalously high and background value classes, which were broadly correlated with the mafic rocks and granitic rocks respectively.

The TOPOVEG algorithm is a rule-based classification (based on a higher nPDF Deciduous Forest Index) predicting increased southern hardwoods on the less-acidic, south-facing slopes underlain by mafic rocks (Figure 4.5, bottom panel).  Other mafic sites and sites underlain by acidic granites favor a more boreal coniferous community (Warner et al., 1991).  However, these are generalized patterns and coniferous and deciduous trees are found at all sites.  Thus a direct association between vegetation and substrate results in a noisy association, as shown in the TM maximum likelihood classification (see Figure 4.5).  In the TOPOVEG classification, the average nPDF Deciduous Forest Index is calculated for each group of contiguous pixels that forms a slope facet.   If the south-facing facet is enriched in deciduous species and the site is near a geophysical anomaly, then the rock type for the slope facet is classified as mafic.  Otherwise, a granitic composition is assumed.  The adjacent topographic facets (ridge tops and then north facing slopes, and finally wet sites) are assumed to be similar in geology to the adjacent topographic facets. The classification takes place at the landscape scale rather than the pixel scale yielding greater classification accuracy for the original study site (86%), and for adjacent areas (86 - 79%), when compared to the maximum likelihood classification (71% - 64%).  More significant, however, is the smoothing of high frequency noise in the TOPOVEG output, making the classification easier to interpret.
 

Case Study 3:  Boreal taiga/tundra and fire scars.

Case study 4:  Geobotanical exploration for hydrocarbon microseeps in the eastern deciduous forest.

The preliminary analyses of microseeps in Wood and Ritchie Counties, in the dissected plateau of West Virginia, show that by combining methods of geobotanical interpretation and lineament analysis, many random associations can be discounted, and a more comprehensive analysis generated.  The study site is the historic Volcano Oil field, on the crest of the North-South trending Burning Springs Anticline, one of the few major structural features (500 m relief) in the region. The crest is almost flat for nearly a mile, and the flanks are characterized by very steep dips, resulting in box-fold structure at the surface.  The oil field was discovered about 1864, and has produced over two million barrels of oil.  Production was from four sandstone horizons within the Connoquenessing Sandstone, the Pottsville, the Greenbrier, and the Pocono Big Injun Sands.  Despite the extensive production, hydrocarbons are still present in parts of the structure.  In the 1950’s, cores from the Sandhill Well, the first deep well drilled in the center of the Appalachian Basin (West Virginia Geological Survey, 1959), led to the recognition of thrust faulting in the lower and middle Devonian in the subsurface of the anticline.

A geobotanical anomaly associated with the Volcano Oil Field was identified in a late summer/early fall (September) Landsat TM image.  At the time the image was acquired, trees had not yet begun to senesce.  The anomaly is characterized by higher radiance in bands 4, 5 and 7, and a less prominent reduction in the radiance of bands 1, 2, 3 and 6. The geobotanical anomaly is either weakly developed or absent in scenes acquired in April, July, and October.  Maximum contrast between the oaks (Quercus), tulip poplars (Liriodendron tulipifera), and maples (Acer), the dominant species of the area, was obtained in the October image, concurrent with peak fall colors.  Unlike the September image, there is little difference between the area underlain by the Volcano Field and the surrounding region.  This suggests that the geobotanical anomaly observed in the September image is a stress-induced feature, rather than a difference in community composition.  In addition, the stress is apparently only sufficient to cause optical changes immediately prior to senescence.

The October image was used to produce a geobotanical anomaly map based on an nPDF transformation (Figure 4.7).  The nPDF is a suite of programs centered around a data reduction technique in which the original image bands are combined, based on their Euclidean distance from selected corners of the multispectral data space.  A nPDF index developed, based on the distance of each multispectral pixel from the corner comprising the minimum in bands 4, 5, and 7, and the maximum in the remaining bands.  This has the effect of maximizing the difference between the pixels with the geobotanical anomaly, compared to the surrounding areas (see Figure 4.5).  A field analysis provided an independent validation.  A four-foot probe extracted soil gas which was analyzed for light hydrocarbons.  As expected, there was general agreement between the light hydrocarbon halo over the Volcano Field, and the geobotanical anomalies shown in the color map, but there is less agreement in detail.

Lineaments were mapped on the spring image, by two independent observers. The lineaments were then categorized based on whether their strike was parallel or near-perpendicular to the strike of the Burning Springs Anticline, the major structural feature controlling the formation of the Volcano Oil Field.  The results are displayed as an overlay on the October geobotanical anomaly image (Figure 4.7).  A comparison of the lineament analysis with the soil gas data suggests that not all lineaments are of structural significance.  Figure 4.7 shows that the soil ethane concentration appears to be highly correlated with distance to cross-strike lineaments, but not with distance to lineaments that are sub-parallel to the strike. Cross-strike lineaments appear to be of greater significance for zones of hydrocarbon microseeps than lineaments that are sub-parallel to strike.  The combination of the lineament analysis and the geobotanical investigation provides a more comprehensive and less ambiguous interpretation then either do independently.
 

Case study 4:  Vegetation phenology in the Amazon basin derived from wavelet reconstruction techniques.

A second filter was applied, a one dimensional median wavelet filter (Donohoe 1993), to each channel in the Amazon FIR filtered data in the temporal direction for each pixel to remove additional residual atmospheric effects.  For many months a single pixel will not have a cloud free day and this second step of filtering uses temporal patterns to further reduce the effects of variable atmosphere.  Even though each pixel was filtered independently over time, the net effect was to further smooth the images in the spatial domain.  The median filter was chosen because it is better at removing non-linear noise caused by clouds which are analogous to Cauchy noise.

Figure 4.8 illustrates a sequence of monthly cloud free images centered on the Amazon Basin for Channels 1, 2 and 3 for the 1983 El Niño year.  For comparison, Figure 4.9 shows the maximum NDVI computed over the same period.  The NDVI images indicate a significant drop in NDVI that corresponds to the El Niño peak in 1983, especially prominent in the highland terra firme centered around Manous, Brazil.  In contrast, the color composites of Figure 4.10 reveal an increase in red reflectance (Channel 1) that is not accompanied by an increase in NIR (Channel 2).  Smith et al. (1997) interpreted the reduction of  NDVI corresponding to the 1983 El Niño as due to increased smoke in the atmosphere from burning during the dryer El Niño periods rather then decreased vegetation abundance as previously interpreted.

In contrast to the 1983 year, the 1987 El Niño (Figure 4.11) pattern is marked by a seasonal phase shift in the vegetative phenology of the Amazon savannas.   The spectral response in both space and time is significantly different between the 1983 and 1987 El Niño periods.  These spatial patterns are uniquely different than those resulting from global biosphere models using the standard NDVI inputs (Field et al 1995).  We find that the dominant El Niño spectral effects observed by AVHRR are represented by temporal phase shifts in the vegetative phenology of different communities.  These phase shifts themselves can be used to further define and map vegetative communities where greenness alone is insufficient to distinguish them.  Furthermore, temporal phase shifts are an important geobotanical cue for detecting vegetation stress from other causes, e.g., presence of heavy metals or contaminants.

We have examined the mapping of wetlands in the Amazon using SMRR data as a means for validating wetland classifications from AVHRR data.  The Pantanal wetland (Figure 4.12) is easily discernible and initially one might think a simple index similar to NDVI could be used to map wetlands.  Water beneath the vegetation would reduce the radiance reflected from the canopy yielding a negative relation between the wetlands presence and albedo in channel 1 and 2.  Fig. 4.13 illustrates the interaction of spatial scales from the respective coherence between the scales of digital terrain models derived from 37 GHz SMMR data (Sippel and Hamiliton, 1994) and AVHRR Channel 2.  The energy between scales was computed using a Symmlet wavelet decomposition of AVHRR channel 2 and digital terrain images. We found that a strong negative coherence existed only at the 100 km spatial scale.  Changes in Channel 2 at larger and smaller scales have little correspondence with the temporal appearance of wetlands.  Spectral variability at the pixel scale in AVHRR Channel 2 has almost no coherence with wetlands.  These results indicate that mapping wetlands using AVHRR GAC data in the Amazon basin is a sub-band filter problem that is clearly scale dependent.  Wetlands of 4 km (e.g. the pixel resolution) will not be accurately mapped using this data.  Because of the small extent of many mineral exposures and point sources of contamination, similar scaling problems exist in many remote sensing datasets.  Methods such as wavelet decompositions can provide insight into the spatial and spectral scaling of the features of interest, thus avoiding misinterpretation of remote sensing data that is inappropriate for the scale of the process and the image data.
 

Case study 5:  Detection of land use change.

Our attempts to retrieve cutting history using conventional remote sensing techniques from single images covering periods of time extending before the Landsat satellite program have failed.  The cutting history for the Gifford Pinchot National Forest, Washington was supplied by the U.S. Forest Service and we applied SMA, FBA and an integrated wavelet/neural net to a 1992 Landsat image.  We found the simple mixture model applied to single multispectral  images had high uncertainties in quantifying seral stage (successional sequence) in this forest.  All results indicate a steadily increasing areal coverage with stand age as shown in Figure 4.15 (i.e., area covered by older stands is more than by younger stands). These results conflict with Forest Service records (Figure 4.14) which clearly depict that most of the area has been cut in the recent past (i.e., a negative slope between stand age and area covered for each age class).  The uncertainty of stand age determined from simple mixture models was highest for the youngest ages (+20 years, Figure 4.15) and lowest and constant for stands above 60 years of age.  Endmember fractions did not vary significantly in forest stands  older than 60 years of age.  Even though younger stands contrast spectrally with older stands, the variability of endmember fractions in old and young stands alike make areal stand age estimates using these methods highly uncertain. This may be due to several factors, e.g., differences in growth rates between conditions at better and poorer sites, failures of reseeding, or incomplete harvests.
 

Case Study 6:  Quantifying Land Use Change and other Change Detection.

SMA and FBA predictions are consistent with the trend in declining uncertainty in stand-age uncertainty as a function of age (Figure 4.16).  The spectral contribution to variance in apparent stand age arises from two main sources: 1) the physical components within the scene, and 2) lighting geometry (e.g., terrain effects).  The precision with which stand age can be estimated by either SMA or FBA is highest for young stands, and converging to a low standard deviation of <5 yrs for stands older than ~60 years (Figure 4.16).  Although FBA produces more precise estimates for young stands, FBA and SMA yield similar results for mature stands.

The spectral variability unique to stand-age is at a finer spatial scale than that arising from  terrain and size of clear cuts.  We used wavelet decomposition to extract the two spatial scales (corresponding to 60 m and 120 m) to isolate the spectral variance attributable solely to seral stage. Younger stands have much greater spectral variability than older stands at these spatial scales due to the compositional variability in the initial stages of regrowth and to less shadowing in gaps between young trees.  Utilizing spectral variation from the finest spectral scales to apply SMA and FBA also uncouples stand age estimates from large scale (regional) moisture gradients and reduces the impact of annuals on the apparent “greenness” during the first few years after cutting (Figure 4.17).

4.7:  References