Ecological applications of physically based remote sensing methods

Abstract

Global monitoring of vegetation using optical remote sensing has undergone rapid technological and methodological development during the past decade. Physically based methods generally apply reflectance models for interpreting remotely sensed data sets. These methods have become increasingly important in the assessment of terrestrial variables from satellite-borne and airborne images. Products based on satellite images currently include various ecological variables that are needed for monitoring changes in forest cover, structure and functioning, including biophysical variables such as the amount of photosynthesizing leaf area. This paper reviews variables and global products estimated from optical satellite sensors describing, for example, the amount and functioning of green biomass and forest carbon exchange. Continuous validation work as new vegetation products are released continues to be important. More emphasis is needed on the collection of field data equivalent to satellite retrievals, data harmonization and continuous measurements of seasonal forest dynamics.

Keywords:

• albedo
• chlorophyll
• fAPAR
• LAI
• LUE
• NPP
• reflectance model

Introduction

Global monitoring of vegetation is highly important in relation to the current scientific and political agenda related to global climate change. For example, forests at high and middle latitudes contain close to 30% of the all the carbon in terrestrial vegetation. Of this carbon, almost 80% is stored in the soil (WGBU, 1998) and the remaining part in the vegetation. It is not clear whether the combined effect of factors affected by climate change will push these ecosystems towards sinks or sources (Kurz et al., 2007). The land sinks of carbon absorb close to one-third of anthropogenic emissions (Canadell et al., 2007); hence, monitoring these systems using the latest technical methods is crucial. Remote sensing methodology has undergone rapid technological and methodological development during the past decade. Satellite images are used to derive products that currently include various ecological or biophysical variables needed to monitor changes in vegetation cover, structure and functioning. Biophysical variables are defined as state variables that directly control the process of radiative transfer in vegetation canopies, e.g. the amount of photosynthesizing leaf area in vegetation.

The term physically based remote sensing is commonly used for describing a set of methods which apply reflectance models to interpret the spectral–directional signatures obtained from airborne or satellite-borne images. Reflectance models describe the upwelling radiation field of an object such as a forest or an agricultural field. The reflectance spectrum of an object (e.g. forest stand) may, if interpreted correctly, be used to identify it and to quantify its structural, physical and chemical properties.

Currently, several physically based retrieval algorithms for global monitoring of vegetation are implemented operationally. For example, using observations from the Moderate-resolution Imaging Spectroradiometer (MODIS) onboard NASA's Terra and Aqua satellites, fortnightly and monthly global vegetation data sets have already been produced for a decade. These data sets are widely used by biologists, natural resources managers and climate modelers (Justice, 1998). MODIS has been followed by other missions and to date several algorithms are used operationally to retrieve land surface variables from remotely sensed optical data. Globally estimated forest variables include the leaf area index (LAI), fractional cover of vegetation (fCover), fraction of absorbed photosynthetically active radiation (fAPAR), net primary production (NPP) and forest shortwave albedo. A list of abbreviations and definitions is given in Table I.

Table I. List of abbreviations and definitions.

Acronym	Concept	Definition
APAR	Amount of absorbed photosynthetically active radiation	Total radiation in the spectral band from 0.4 to 0.7 µm absorbed by the plant canopy. Quantifies the energy available for net primary production. Calculated as the product of fAPAR and incident PAR flux
AVHRR	Advanced Very High Resolution Radiometer	A coarse spatial-resolution broadband scanner that provides data in the visible, near-infrared and thermal infrared spectral regions. Carried on the National Oceanic and Atmospheric Administration's (NOAA's) Polar Orbiting Environmental Satellites
BHR	Bihemispherical reflectance factor	The ratio of the radiant flux reflected from a unit surface area into the whole hemisphere to the incident radiant flux of hemispherical angular extent. The incoming radiation is composed of direct and diffuse irradiance. Also called blue-sky albedo
BHRiso	Isotropic bihemispherical reflectance factor	The ratio of the radiant flux reflected from a unit surface area into the whole hemisphere to the incident radiant flux of hemispherical angular extent. The incoming radiation is assumed isotropic. Also called white-sky albedo
BRDF	Bidirectional reflectance distribution function	A function characterizing the amount of electromagnetic radiation scattered from a surface as a function of view and solar angles. Defined as the ratio of the reflected radiance to the incident irradiance
CHRIS	Compact High Resolution Imaging Spectrometer	A hyperspectral–multiangular satellite sensor onboard the PROBA platform
DHR	Directional–hemispherical reflectance factor	The ratio of the radiant flux reflected by a unit surface area into the view hemisphere to the incident radiant flux, when the surface is illuminated with a parallel beam of light from a single direction. Also called black-sky albedo
ER	Ecosystem respiration	The sum of carbon released through plant respiration and heterotrophic respiration
EVI	Enhanced vegetation index	A vegetation index that has high sensitivity in green biomass regions. Based on the blue, red and near-infrared channels of a remote sensing instrument
fAPAR	Fraction of absorbed photosynthetically active radiation	The fraction of the incoming solar radiation in the PAR region that is absorbed by a photosynthetic organism
fCover	Fractional cover of vegetation	A canopy biophysical parameter that describes the fraction of a pixel covered by green biomass
fIPAR	Fraction of intercepted photosynthetically active radiation	The fraction of the incoming solar radiation in the PAR captured by canopy layers as the PAR incident at the canopy top travels down through the canopy to the ground. In a dense, green canopy fIPAR is close to fAPAR since healthy green leaves absorb most PAR
GPP	Gross primary production	The total amount of carbon fixed through photosynthesis. Can be computed as the product of the LUE coefficient, fAPAR and incident PAR
LAI	Leaf area index	One half of the total leaf surface area per unit ground surface area projected on the local horizontal datum. Is a canopy biophysical parameter
LUE	Light-use efficiency	Describes the efficiency by which APAR is converted to NPP or GPP
LUT	Look-up table	A precalculated data structure used to replace runtime computations. Can be used as an inversion method for forest reflectance models
MERIS	Medium Resolution Imaging Spectrometer Instrument	A medium spectral and spatial resolution imaging spectrometer operating in the solar reflective spectral range aboard the ENVISAT platform
MODIS	Moderate Resolution Imaging Spectroradiometer	A medium spectral and spatial resolution imaging spectrometer operating in the solar reflective spectral range instrument aboard the Terra and Aqua satellites
NBP	Net biome productivity	The total carbon that is taken up or released by an ecosystem
NDVI	Normalized difference vegetation index	A vegetation index that is sensitive to the amount of green biomass. Computed from reflectance factors measured in the red and near infrared channels of a remote sensing instrument
NEE	Net ecosystem exchange	The net amount of carbon assimilated or released by an ecosystem when dealing with measurement data based on eddy covariance techniques. Summarizes the flows of carbon dioxide to and from the ecosystem via the vegetation
NEP	Net ecosystem productivity	The net amount of carbon assimilated or released by an ecosystem (NEP = GPP–ER)
NPP	Net primary production	Describes net plant growth below and above ground
PAR	Photosynthetically active radiation	Radiation in the spectral band from 0.4 to 0.7 µm
PRI	Photochemical reflectance index	A vegetation index that is sensitive to changes in carotenoid pigments. Related to the photosynthetic efficiency. Computed from narrow-band reflectance factors measured at 531 and 570 nm
SeaWiFS	Sea-viewing Wide Field-of-view Sensor	A coarse spatial-resolution broadband scanner operating in the solar reflective spectral range instrument aboard the SeaStar platform
VALERI	Validation of Land European Remote Sensing Instruments	An international research network that provides high spatial resolution maps of biophysical variables (LAI, fAPAR, fCover) estimated from ground measurements to validate products based on satellite images
VI	Vegetation index	A mathematical transformation of reflectance factors measured in two or more spectral bands of a remote sensing instrument. Enhances the spectral contribution from green vegetation and minimizes the effects of soil, atmosphere and illumination geometry

This paper is structured as follows. First, forest reflectance modeling from the 1980s until the current phase is briefly reviewed. Next, the chapter looks at vegetation variables and global products obtained from passive optical satellite images, describing (1) the amount and functioning of green biomass (e.g. LAI and chlorophyll content), (2) forest carbon exchange (e.g. light-use efficiency and fraction of absorbed photosynthetically active radiation), and (3) forest shortwave albedo. Finally, future perspectives of ecological applications of physically based remote sensing methods are discussed.

Forest reflectance modeling

Depending on the spatial and spectral characteristics of the remote sensing data, different forest reflectance models can be used to create the observed spectral–directional signatures using so-called “forward mode” model simulations. A range of approaches exists to describe the reflectance of vegetated surfaces mathematically (see Stenberg et al., 2008, vegetation with a probabilisticfor a more complete review of model types). The methods range from the simplest homogeneous turbid medium (Ross, 1981) to detailed three-dimensional (3D) models (Disney et al., 2000). Besides complexity, the methods differ also in their computational efficiency and the amount of detail required as input. In the turbid medium approach, the forest canopy is described as a volume filled with infinitesimally small elements, leaves or needles; the only biophysical parameters are LAI and the leaf angle distribution. In addition, the model requires leaf spectral albedo. While it is not a biophysical parameter in its own right, leaf reflectance and transmittance can be considered intermediate products carrying information about foliar biochemistry (e.g. Jacquemoud et al., 1996). The widely used Scattering by Arbitrarily Inclined Leaves (SAIL) (Verhoef, 1984) model is the best known example of a homogeneous canopy reflectance model. Although these simple models are still used, the need to include the structure present in all natural forests has led to the development of more complex and realistic models. Therefore, remote sensing research has turned to reflectance models that can explicitly include the 3D characteristics of a vegetation canopy (Schaepman et al., 2009).

In a 3D forest reflectance model, the transfer of radiation from a specified direction is uniquely defined by the leaf area density, leaf angle distribution (described by a probability density function) and leaf optical properties (described by the scattering phase function). However, contrary to the simple homogeneous model, the optical properties of the medium change continuously. In numerical computations, the continuous forest canopy is discretized into small elementary volumes, or voxels. The 3D radiative transfer computations used to predict forest canopy reflectance not only are computationally intensive but also require much more detailed information on forest structure than what is commonly available. The most detailed 3D approach is commonly known as Monte Carlo simulation or ray tracing: randomly generated photons are traced through the vegetation with a probabilistic representation of reflectance, transmittance and absorption events. A large number of simulated photons is required to predict forest reflectance in a statistically reliable manner. Despite their complexity and heavy requirements on computer resources, 3D models provide us with a reliable tool for exact calculation of the reflective properties of complex media with well-known or predefined structure (Widlowski et al., 2007).

Often a compromise among the level of detail, robustness and accuracy is sought in the form of geometric–optical (GO) models (Li & Strahler, 1985; Chen & Leblanc, 1997). These models represent the canopy as an aggregation of geometric tree crown envelopes whose locations are described by a statistical distribution. Thus, the stand parameters used by these models correspond more closely to what can actually be measured in a forest: crown dimensions, canopy closure, etc. In its simplest form, a GO model predicts the fractions of four areas of different brightness in a remotely sensed image: directly illuminated tree crowns and ground, and shadowed tree crowns and ground (Chen & Leblanc, 1997). Improving such models to allow for a quantitative prediction of the radiance of the four areas leads to a hybrid model (e.g. Li et al., 1995; Kuusk & Nilson, 2000). Hybrid models combine features from the GO and turbid medium approaches: the turbid medium approach is used for within crown radiative transfer, while the GO model is typically used for modeling between crown interactions. Clumping at scales smaller than the crown can be accounted for in reflectance calculations by choosing an aggregated unit (e.g. a coniferous shoot) as the basic structural unit, for which the scattering and absorption properties are defined (Smolander & Stenberg, 2005).

Since the very beginning of the research field, the general trend has moved away from homogeneous models to ones including more structural variables. An exciting example of new developments in the field of vegetation reflectance modeling is the p-theory. This theory states that simple algebraic combinations of leaf and canopy spectral transmittance and reflectance become wavelength independent and determine a small set of canopy structure-specific variables (Knyazikhin et al., 1998a; Panferov et al., 2001). These spectrally invariant variables include the recollision probability (p) and the escape probability (1 – p). The recollision probability is defined as the probability that a photon, after being scattered from a leaf (needle) in the canopy, will interact within the canopy again (Smolander & Stenberg, 2005). The recollision and escape probabilities can be expressed as a function of canopy gap fractions and LAI (Stenberg, 2007), and depend on the structure of the canopy at a variety of scales (Rautiainen & Stenberg, 2005; Mõttus & Stenberg, 2008). Knowing p, the fractions of radiation scattered and absorbed by the canopy at any specific wavelength λ are obtained as a simple function of the leaf (needle) scattering coefficient ω _L(λ), at the same wavelength. Only recently, Schull et al. (2007) presented a method to retrieve the recollision probability and the directional escape factor (escape probability in the view direction normalized by the canopy interceptance) for a forest canopy from hyperspectral remote sensing data. A limitation of the method, however, is the assumption of a non-reflecting background (“black soil”). Thus, it may not be successfully applied, e.g. to sparse forests with significant understorey reflectance.

Inversion of a forest reflectance model can be used to obtain biophysical variables from remote sensing data. The task of interpreting the spectral–directional signature of a forest can be reduced to matching the measured reflectance signal to a known simulated signal. However, inversion is an ill-posed problem as many parameterizations of a reflectance model may correspond to the same measured reflectance. Even the simplest vegetation canopy reflectance models, which can realistically take into account the complex structure of forest canopies, are not easily invertible. For computationally less demanding models, iterative optimization may be used: the values of the unknown parameters are slightly modified between consecutive model runs until the model-predicted reflectance spectrum fits the observed one closely enough. In operational environments, a look-up table is usually computationally more efficient if it is used repeatedly (e.g. Knyazikhin et al., 1998a). Theoretically, neural networks should be the fastest inversion method (Liang, 2007). Inversion of forest reflectance models to retrieve biophysical variables is further discussed later in this article, in the section on Operational retrieval of ecological variables at global scale.

Ecological variables of current interest in remote sensing

Variables describing the amount and functioning of green biomass

Leaf area index

The LAI characterizes the amount (area) of leaves in a plant canopy. LAI was originally defined for flat leaves as the one-sided leaf area per unit ground area (Watson, 1947). It has been convincingly argued that for coniferous needle leaf canopies, the logical counterpart to the one-sided area of flat leaves is half of the total surface area of needles (e.g. Lang, 1991; Chen & Black, 1992; Stenberg, 2006). Subsequently, the currently most widely accepted definition of LAI, applicable to both deciduous and coniferous species, is the “hemisurface LAI”, i.e. one half of the total leaf surface area per unit ground surface area projected on the local horizontal datum (Morisette et al., 2006; Gonsamo, 2009).

LAI is a key characteristic of interest in forest ecosystems because the green leaves control the processes driving the exchange of matter and energy. LAI is also the main determinant of the fraction of intercepted radiation by the canopy, i.e. the portion of the incoming radiation that “collides” with canopy elements (predominantly leaves). In particular, as leaves absorb strongly in the photosynthetically active radiation (PAR, wavelengths from 0.4 to 0.7 µm) portion of the solar spectrum, most of the intercepted PAR is absorbed by the canopy. fAPAR is defined as the fraction of the incoming solar radiation in the PAR region that is absorbed by a photosynthetic organism. Thus, fAPAR is closely related to the fraction of intercepted photosynthetically active radiation (fIPAR). For radiation arriving from a single direction, the general form of the relationship between LAI and fIPAR is given by fIPAR = 1–exp(kLAI), where k is a canopy specific extinction coefficient that varies with the direction of incoming radiation (e.g. solar elevation). The relationship, known as Beer's law, was originally applied to plant canopies with randomly distributed leaves (Monsi & Saeki, 1953), in which case k is simply a function of the leaf angle distribution. For natural (forest) canopies, however, k also includes the effect of grouping or non-random spatial distribution of leaves. As an approximation, Beer's law is also applied to global radiation arriving from the whole upper hemisphere. Together with the close relationship between fIPAR and fAPAR, it underlies the physical connection between LAI and PAR absorption. The exponential attenuation of PAR provides the basis for indirect optical methods to estimate LAI from canopy gap fraction data (e.g. hemispherical photographs, LAI-2000 Plant Canopy Analyzer) (Welles, 1990; Jonckheere et al., 2004). However, the indirect estimates are commonly referred to as “effective LAI”, because the effect of grouping at small scales cannot be accounted for by these methods.

The total amount of absorbed photosynthetically active radiation (APAR) quantifies the energy available for net primary production (NPP) and is a critical variable in forest growth and carbon flux models (see section on Variables describing the forest carbon exchange, below). LAI also has a major impact on the surface albedo (see section on Variables describing forest shortwave albedo, below). Finally, because forest leaf area responds to different stress factors and changes in climatic conditions, LAI serves as an important indicator variable in global change research (Myneni et al., 1997).

An increased interest in LAI and its applications emerged with the feasibility of satellite remote sensing for large-scale vegetation assessment (Sellers et al., 1995). The physical background of remote sensing of LAI lies in the close relationship between LAI and canopy gap fractions in the sun and view directions (bidirectional gap probability), together with the characteristic spectra of green leaves. Thus, LAI is closely linked to the spectral reflectance of plant canopies in the shortwave solar radiation range.

Leaf chlorophyll content

Chlorophyll, a green foliar pigment, has a key role in photosynthesis, where it absorbs and transports the energy of incoming sunlight to photosystem reaction centres. Chlorophyll molecules serve as an interface between incident solar light and inner cell photosystem reaction centres in photosynthesis. Five different chlorophyll structures have been identified: C_a, C_b, C_c1, C_c2 and C_d. C_a and C_b have the highest concentrations in plant tissues, Ca concentration being three times higher than C_b. Chlorophyll exhibits a selective absorption of light in a range of visible wavelengths, with low absorption in green and high absorption in blue and red wavelengths (absorption peaks of 430 and 662 nm for C_a; 465 and 642 nm for C_b), resulting in the green color of leaves. Leaf chlorophyll concentration is the determining factor of leaf spectral reflectance in many visible wavelengths, although other foliar pigments (particularly carotenoids and xanthophylls) also play a role (Sims & Gamon, 2002).

C_{a + b} concentrations differ for species during their phenological cycle. Therefore, recognition of different species based on C_{a + b} content is possible (Asner, 1998). Chlorophyll concentration responds to leaf physiological stress (Rock et al., 1988), and overall concentration of C_{a + b}, or ratio between C_a and C_b (Fang et al., 1998) can be used as a proxy of vegetation health status and preliminary indicator of stress caused by insects (Lawrence & Labus, 2003). Moreover, leaf chlorophyll content is strongly related to nitrogen content via additional protein absorption features (Schlerf et al., 2010).

Estimation of chlorophyll content by remote sensing became feasible with hyperspectral remote sensing missions. Hyperspectral data enable relatively narrow chlorophyll absorption features to be observed. Quantitative prediction of C_{a + b} content can be performed by defining an empirical relationship between a vegetation index and ground C_{a + b} measurements (Zagolski et al., 1996), or by radiative transfer (RT) modelling (Demarez & Gastellu-Etchegorry, 2000).

Many chlorophyll vegetation indices (VIs) have been designed for leaf-level scale (see review by le Maire et al., 2004), but only a few of them are applicable at crown level. A properly designed canopy-scale VI should be sensitive to changes in chlorophyll concentrations but insensitive to influence of understorey and canopy 3D structure. VIs for chlorophyll content retrieval are usually based on factors such as (1) neighboring wavelength of maximum chlorophyll absorption (i.e. saturation of absorption feature at low C_{a + b} values) normalized by a C_{a + b} non-sensitive band (TCARI; Daughtry et al., 2000), (2) soil line of vegetation (OSAVI; Rondeaux et al., 1996), or (3) transformation of the spectral curve in a chlorophyll-sensitive range of wavelengths (Kokaly & Clark, 1999). Schlerf et al. (2010) reviewed spectral transformations for C_{a + b} and nitrogen content retrieval for Norway spruce (Picea abies (L.) Karst.). VIs have been successfully applied at both local scale and global scale (Dash & Curran, 2007).

Variables describing the forest carbon exchange

Forest carbon exchange concepts

Ecosystem carbon budgets are characterized using different variables (Grace et al., 1999): the total amount of carbon fixed through photosynthesis is known as the gross primary productivity (GPP). This variable describes the photosynthetic conversion of carbon dioxide (CO₂) into carbohydrates used for growth and maintenance of the trees. About half the carbon assimilated in the photosynthesis is lost in the respiratory gas exchange known as plant respiration. The remaining part is known as the net primary productivity (NPP), which describes the net plant growth below and above ground. Heterotrophic organisms that feed on organic matter, such as dead plant and animal matter, will release CO₂ to the air. The sum of carbon released through plant respiration and heterotrophic respiration is known as ecosystem respiration (ER). Subtraction of ER from GPP provides the net amount of carbon assimilated or released by the ecosystem, which is referred to as net ecosystem productivity (NEP). When dealing with measurement data based on eddy covariance techniques (Baldocchi, 2003) this parameter is generally referred to as net ecosystem exchange (NEE). It summarizes the flows of CO₂ to and from the ecosystem via the vegetation and can, over longer periods, be dominated by assimilation (the vegetation is a carbon sink) or by release (the ecosystem is a carbon source). However, carbon can be lost from the ecosystem in other ways such as harvest, fire, erosion, changes in land use, or through drainage to groundwater and streams. The total carbon that is thus taken up or released by the ecosystem is termed the net biome productivity (NBP).

Absorbed photosynthetically active radiation and light-use efficiency

Remote sensing is not capable of monitoring all aspects of the forest carbon cycle, but can provide a useful addition to other monitoring techniques in providing data on some of the relevant variables at high spatial and temporal resolution. A variable that has commonly been estimated using optical satellite sensor data is NPP. The monitoring of this variable rests on the assumption that NPP is determined by the APAR (Monteith, 1972, 1977). The efficiency of the biochemical processes converting CO₂ into carbohydrates with the aid of radiative energy is expressed with a light-use efficiency (LUE) coefficient (ε _n). This coefficient expresses the ratio of NPP to APAR. APAR is the total radiation energy in the PAR spectral band (0.400–0.700 µm) absorbed by the plant canopy. In practice, it is often calculated as the product of fAPAR and incident PAR flux. In general, GPP is expressed using a similar linear expression, but with a different LUE coefficient (ε _g rather than ε _n ) which would imply that the relationship between GPP and NPP is nearly constant. However, several researchers have argued that there may be significant variation in the NPP to respiration ratio, and that the relationship should be built around GPP instead (Prince, 1991; Goetz & Prince, 1999). Hence, GPP can be computed as the product of the LUE coefficient, fAPAR and incident PAR. To obtain NPP it is necessary to subtract plant respiration.

The general validity in estimating GPP or NPP via remote sensing depends to a large extent on the possibility of modeling the LUE coefficient. The degree to which LUE converges towards a constant value or range is a matter of scientific debate. It has been suggested that there may be convergence in LUE, owing to adjustments of light capture due to resource shortage (e.g. drought, nutrient shortage), and that this convergence may have an evolutionary basis (Field, 1991). However, owing to differences in carbon-use efficiency among functional types with different carbon allocation strategies this does not appear to hold for ε _n. There is, however, evidence suggesting convergence in the amount of carbon assimilation per unit APAR, i.e. ε _g (Goetz & Prince, 1999). Hence, application of the LUE concept in remotely sensed data for terrestrial primary production modelling should be based on GPP, and when estimating NPP autotrophic respiration needs to be estimated as a separate term.

Estimation of the LUE coefficient is critical to the LUE concept, however, with no unanimously adopted solution owing to the obvious difficulty in summarizing complex biochemical processes into a single parameter. A common approach is to regard the photosynthetic efficiency in terms of environmental stresses (Goetz & Prince, 1999; Prince & Goward, 1995). The LUE coefficient is then separated into a vegetation-specific maximum LUE factor, and scalars representing stress suppression related to limitations in water availability, temperature and nutrients (Prince & Goward, 1995; Goetz et al., 1999; Seaquist et al., 2003). This concept has been applied in the global NASA MODIS GPP product (MOD17) (Heinsch et al., 2006). Different approaches have been applied to the estimation of the stress scalars, either by modeling them based on climate data, or by estimating them from remotely sensed data. The possibility of evaluating these estimates has improved dramatically with the increased availability of eddy covariance flux data. In Scandinavian needle-leaf forests LUE is strongly temperature dependent, with a marked seasonal profile (Lagergren et al., 2005).

Another approach to estimating LUE is to observe directly its effect on chlorophyll fluorescence and reflectance changes. Chlorophyll fluorescence can be efficiently measured using laser pulses at the leaf level, but also at a short distance from the canopy (Kolber et al., 2005). There is also potential for measuring fluorescence using passive techniques (Corp et al., 2006); however, because of low energy levels the technique remains challenging. Nevertheless, plans have reached an advanced stage regarding including a fluorescence sensor onboard a European Space Agency (ESA) satellite. Spectral reflectance changes in response to the epoxidation state of the xanthophyll cycle have also been observed (Gamon et al., 1990). These changes have been quantified with a spectral vegetation index, the photochemical reflectance index (PRI), which is related to the photosynthetic efficiency (Gamon et al., 1997; Gamon & Surfus, 1999; Hall et al., 2008). Although the relationship seems to hold over a range of species, and PRI has been shown to correlate with satellite measurements (Rahman et al., 2004; Drolet et al., 2005), upscaling remains difficult owing to the temporal dynamics in plant photosynthesis, small changes in reflectance at 531 nm, and effects related to view angle, soil background and leaf area (Hall et al., 2008). However, novel studies (Hall et al., 2008; Hilker et al., 2009) have demonstrated that simultaneous multiangular measurements of canopy shadow fraction and PRI would enable efficient space-based estimation of LUE.

The simplicity of the LUE concept and the possibility of using it with remotely sensed data have led to the development of a large number of LUE models over the past 20 years. They differ in details and complexity, but are built on similar principles. The incident radiation (I _p) that drives the production can be estimated from gridded climate data or directly based on remote sensing. The former approach depends on the availability of gridded PAR irradiance data, but if these are not directly available the PAR irradiance may be derived as a constant fraction of the total shortwave flux, with acceptable accuracy in Scandinavia when integrating over monthly or yearly periods (Olofsson et al., 2007). The satellite approach is based on radiative-transfer modeling and has been implemented with data from MODIS (Van Laake & Sanchez-Azofeifa, 2005; Liang et al., 2006; Olofsson et al., 2007; Liu et al., 2008). There is room for further improvements, particularly for daily PAR flux estimation.

Absorbed fraction of photosynthetic radiation

The fAPAR is a critical ecosystem variable that has been shown to be near-linearly related to vegetation indices, based on theoretical studies, radiative transfer modeling and analysis of empirical data (Tucker & Sellers, 1986; Goward & Huemmrich, 1992; Myneni & Williams, 1994; Fensholt et al., 2004; Olofsson & Eklundh, 2007). Most commonly, the normalized difference vegetation index (NDVI, Rouse et al., 1973) has been used for estimating fAPAR, but the enhanced vegetation index (EVI) has also been applied (Xiao et al., 2004a, 2004b) as it has higher sensitivity in dense vegetation (Huete et al., 2002) and to leaf chlorophyll light absorption (Xiao et al., 2004b). The remote sensing literature is dominated by empirical approaches to estimating fAPAR; however, it should be noted that fAPAR retrieval is sensitive to soil reflectance, non-photosynthetic plant components, viewing angle variations and atmospheric conditions (Goward & Huemmrich, 1992; Myneni & Williams, 1994). Hence, empirical relationships may not be transferable to other areas or measurement conditions.

The European Commission's Joint Research Centre (JRC, Ispra, Italy) has developed a fAPAR algorithm that relies on rectification of the reflectance data for atmospheric and illumination and viewing geometry in combination with the derivation of optimized vegetation indices based on radiative transfer modeling (Gobron et al., 2000). It has been applied to various data types such as SeaWiFS and Envisat MERIS. NASA has developed a MODIS based global LAI and fAPAR product at 1×1 km resolution (MOD15) that is based on radiative transfer modeling and uses a look-up table with biome-specific parameters (Myneni et al., 2002).

Variables describing forest shortwave albedo

Forest broadband shortwave albedo is a critical variable affecting the Earth's climate (e.g. Dickinson, 1995). However, it is still among the main uncertainties of the radiation budget in current climate modeling (Liang, 2007). Seasonal and long-term changes in forest cover and vegetation dynamics induced by human activities have changed the surface albedo, or the extent to which incoming solar radiation is reflected back to the atmosphere, around the globe. Variations in snow cover, flooding, seasonal vegetation dynamics of agricultural crops and natural vegetation communities, as well as burning and clearing of land surface all result in a change in surface albedo (e.g. Schaaf et al., 2008; Román et al., 2009). From the perspective of remote sensing of forests, reliable quantitative predictions of the effect of forest structure and its global cover on albedo are needed for climate modeling. For example, forest management practices may influence forest albedo (Betts, 2000) through changes in canopy cover, stand density, LAI and the amount of snow visible through the canopy (Manninen & Stenberg, 2009).

Before satellite-based estimation of land surface albedo became possible, albedo values were usually obtained from field measurements at specific meteorological study sites and then extrapolated to cover the whole study region. However, variation in both the temporal and spatial patterns of surface albedo caused problems in interpreting the extrapolated results. Currently, using satellite remote sensing, albedo can, in theory, be determined pixel by pixel for the entire globe.

A problem related to the albedo discussions has been the confusing use of terminology. Albedo is commonly defined as the fraction of the incident radiation that is reflected by a surface. However, in scientific considerations, there are three distinct albedo quantities: black-sky albedo (directional–hemispherical reflectance, DHR), white-sky albedo (isotropic bihemispherical reflectance, BHR_iso) and blue-sky albedo (bihemispherical reflectance, BHR) (Schaepman-Strub et al., 2006). Black- and white-sky albedos are the two extreme cases of completely direct and completely diffuse illumination. Blue-sky albedo corresponds to ambient illumination conditions, and is thus often also called “actual albedo”.

The black-sky albedo is the ratio of the radiant flux reflected by a unit surface area into the view hemisphere to the incident radiant flux, when the surface is illuminated with a parallel beam of light from a single direction (uncollided irradiance, i.e. assuming no scattering). It corresponds to pure direct illumination (a “clear day”) and depends on the actual illumination angle of the direct component. The white-sky albedo (BHR_iso) corresponds to a “cloudy day”, and is the ratio of the radiant flux reflected from a unit surface area into the whole hemisphere to the incident radiant flux of hemispherical angular extent. In other words, BHR_iso is an approximation of the albedo of the surface under diffuse (atmospherically scattered) irradiation; the incoming radiation is assumed isotropic, and for pure diffuse illumination conditions, it does not depend on solar angle. Finally, the blue-sky albedo (BHR) is a function of aerosol optical depth, i.e. it is influenced by the combined direct and diffuse irradiance, and has also been referred to as the actual albedo. BHR can be approximated through the weighted combination of the black- and white-sky albedos (Lucht et al., 2000). Finally, all the above-mentioned albedo quantities can be further classified as spectral (or narrow-band) albedos or broadband albedos. Spectral albedo refers to an albedo value calculated (or measured) for one single wavelength or spectral band, whereas broadband albedo corresponds to an integrated value of spectral albedos.

Theoretical studies on the relationship between forest structure and broadband albedo can be carried out using vegetation reflectance models (see section on Forest reflectance modeling above). Forest reflectance models which have been coupled with an atmospheric radiative transfer model (or empirical data on the downwelling flux) can be used to simulate the spectral albedo (DHR, BHR and BHR_iso) for a given forest. GO- and hybrid-type forest reflectance models are especially useful for this purpose: they use typical forest inventory parameters as their input and can thus generate realistic scenarios. However, the composition and phenology of the understorey layer (e.g. shrubs, grasses, mosses, snow) also influence forest albedo, yet field data on it are largely lacking in forest inventory databases.

Nevertheless, we must remember that remote sensing instruments do not directly measure (forest) surface albedo. Typically, satellite instruments record the radiance at the top of the atmosphere in one viewing angle (or sometimes in several viewing angles, e.g. CHRIS, PROBA, MISR, POLDER). The instruments also measure only in narrow wavelength regions or bands, and the albedo values retrieved under one specific atmospheric condition may not be applicable to other atmospheric conditions (Liang et al., 1999). Thus, the challenges in estimating forest surface albedo from optical satellite images are related to modeling radiative transfer through the atmosphere (i.e. accounting for the atmospheric effects), modeling the bidirectional reflectance distribution function (BRDF) of the forest (or other land surface) with a forest reflectance model, and developing methodologies for converting spectral albedos to broadband albedos for extensive areas. Finally, satellite-derived albedo estimates need to be validated with ground reference measurements (e.g. Liang et al., 2002, Román et al., 2009).

Operational retrieval of ecological variables at a global scale

Reliable information on vegetation status at a global scale has become a prerequisite for decision making. The Global Climate Observing System (GCOS) second adequacy report (GCOS, 2003) highlighted a number of essential climate variables (ECVs), including LAI, fAPAR and land surface albedo. Vegetation maps are required for carbon budgeting. For many regions, reasonable maps are already available from different national sources. However, various mapping and land-use information projects and harmonization efforts of the national inventories, as well as United Nations Food and Agriculture Organization (UN FAO) global forest resources assessment, have revealed the difficulties in obtaining comparable and reliable information at a global scale.

Operational satellite observations refine estimates of the distribution and variability of Earth's vegetation. Continuous monitoring of vegetation structure and functioning at regional and global scales has improved considerably since the launch of new-generation medium-resolution satellite sensors (e.g. SPOT VEGETATION, Terra and Aqua MODIS, Terra MISR, SeaWiFS and ENVISAT MERIS). Medium-resolution sensors provide observations of the Earth's surface at various spatial and temporal resolutions, starting from daily observations at 250×250 m resolution. Using these data, sensor science teams have developed many algorithms to supply various operational products for LAI (Myneni et al., 2002; Bacour et al., 2006; Deng et al., 2006; Baret et al., 2007), fAPAR (Myneni et al., 2002; Gobron et al., 2006, 2007; Plummer et al., 2008), fractional cover of vegetation (Hansen et al., 2003), broadband albedo (Schaaf et al., 2002) and GPP/NPP (Running et al., 2004). Global vegetation products are routinely assessed by the Committee on Earth Observation/Land Product Validation Subgroup (CEOS/LPV). The mission of CEOS/LPV is “to foster quantitative validation of higher level global land products derived from remote sensing data and to relay results so they are relevant to users”.

Canopy biophysical variables (e.g. LAI and fAPAR) have been estimated from remotely sensed data by two types of algorithms: by empirical models that are calibrated using in situ measurements and by methods that use physically based reflectance models. In empirical algorithms, the retrieval is based on statistical relationships modeled between the concurrently acquired in situ and surface reflectance data, which are typically expressed in the form of a vegetation index (VI). VIs include various combinations of signal values in multispectral bands designed to maximize the sensitivity to vegetation characteristics while minimizing the sensitivity to atmosphere, background, view and solar angles (Baret & Guyot, 1991; Myneni et al., 1995). VIs are the simplest method for matching a measured spectral signature with biophysical parameters. Nevertheless, although VIs are computationally efficient and well suited to the broad spectral bands of many satellite instruments, it is evident that a few spectral bands cannot completely describe the reflective properties of a vegetation canopy: VIs represent composite properties of the different biophysical variables and they are moderately useful in predicting individual canopy properties (Glenn et al., 2008).

Although pure empirical approaches have been employed at the regional scale to retrieve Canada-wise LAI products (Chen et al., 2002; Fernandes et al., 2003; Rochdi & Fernandes, 2010) and to retrieve fractional cover at a global scale (Hansen et al., 2003), current operational algorithms typically make use of radiative transfer-based reflectance models. Empirical models are site, sensor and time specific, which makes it difficult to adjust them to variable vegetation and observation situations. Considering this, the possible advantages of canopy reflectance models (see section on Forest reflectance modeling) become clear: reflectance models can predict vegetation reflectance properties in continuous spectral intervals and they are sensor independent. Furthermore, a properly formulated canopy reflectance model is more robust and can identify noise-contaminated reflectance values, indicating errors in the measuring system or a non-vegetated surface as a source of the signal.

Reflectance models can be used in several ways to retrieve canopy variables. The simplest approach is to use model simulations to calibrate the relationships between model parameters and VIs (Myneni et al., 1997; Gobron et al., 2000; Buermann et al., 2002; Deng et al., 2006). However, the VIs do not fully utilize the information of spectral–directional signatures. The more advanced methods to invert reflectance models include iterative optimization algorithms, look-up tables (LUTs) and advanced statistical methods, such as neural networks (Liang, 2004). Iterative optimization algorithms require numerous model runs for each reflectance value (or pixel in common remotely sensed imagery) and are computationally too demanding for retrieving land surface variables at a global scale (Kimes et al., 2000). LUT and neural networks can be used for faster operational inversion. A LUT is a precomputed database of reflectance values where combinations of model parameter values are varied. The LUT is used for searching a set of reflectance values that are most similar to the measured signature, the goodness of the fit being measured with a merit function. Accurate inversion may require large LUT dimensions, which may slow down the search process. The LUT method is used, for example, for deriving the MODIS and MISR LAI and fAPAR products (Knyazikhin et al., 1998a, b). In neural networks (Liang, 2004), a database of model simulations is used only once when calibrating the networks. After the calibration has been done the inversion is very efficient. Neural networks are applied, for example, to retrieve a set of biophysical parameters (LAI, fAPAR, fCover and LAI×C_{a + b}) from MERIS (Bacour et al., 2006) and VEGETATION (CYCLOPES products; Baret et al., 2007).

In addition to the inversion method used, current physically based algorithms differ in terms of the reflectance models, preprocessing level of the satellite data and use of prior information (Baret & Buis, 2008). Most of the LAI retrieval algorithms couple separate models for leaf, soil and canopy. Furthermore, if an atmospheric radiative transfer model is also coupled, canopy variables can be retrieved directly from the top-of-atmosphere reflectance values. The compositing techniques, and the filtering and temporal smoothing of satellite data or retrieved products also vary. Although some algorithms are designed to operate for all types of vegetation (Bacour et al., 2006; Baret et al., 2007), land cover maps are commonly used for the adjustment of reflectance model simulations to a specific land cover type and to limit the retrievals to the expected range of variables. For example, the MODIS LAI and fAPAR algorithms use an eight-class biome classification system. A disadvantage is the sensitivity to classification errors (Myneni et al., 2002; Yang et al., 2006).

The inversion of canopy variables is characteristically an ill-posed problem as different combinations of model parameters can produce similar reflectance values (e.g. Baret & Buis, 2008). Thus, for example, the retrieval of LAI from surface reflectance values is unstable in that small variation in reflectance can result in a large change in LAI, particularly when the reflectance signal saturates (e.g. Knyazikhin et al., 1998a). This also makes retrievals very sensitive to preprocessing and measurement uncertainties of the satellite data. Incorporating more a priori information and data from different instruments with complementary spectral and directional sampling may provide improvements in estimation accuracy.

The intercomparison of LAI products has shown considerable differences between the data sets, particularly over forest areas (Garrigues et al., 2008). One major source of uncertainty is the different representation of the grouping of foliage at different scales, which is particularly important in needleleaf forests. Depending on the reflectance model, foliage grouping and landscape heterogeneity are treated differently. In addition, understorey LAI and woody material are considered with diverse methods (Garrigues et al., 2008). Substantial land cover dependent differences have also been noted among global fAPAR data sets over northern Eurasia (McCallum et al., 2010). Potential reasons for these discrepancies include differences in, for example, the retrieval methods, the use of LAI and land cover data, or snow effects.

Quantitative validation of satellite-derived products (i.e. direct comparison to in situ measurements) has become increasingly important as global-scale retrievals are applied more and more and a multitude of products has become available (Morisette et al., 2006). Validation is also an essential part of algorithm improvement (Yang et al., 2006). The ground-based networks devoted to the validation of satellite products (e.g. VALERI) have been crucial in establishing validation protocols for in situ measurements and their upscaling to the much coarser spatial resolution of satellite sensors. The validation networks typically consist of measurement sites distributed over a range of land cover types supplied with fine spatial resolution maps of biophysical variables based on fine-resolution satellite images (Morisette et al., 2006).

Future perspectives

Retrieval of a growing number of forest variables from remote sensing data requires an increase in the information content of remote measurements. Ideally, in the future, the continuous reflectance spectrum should be used for signature matching (Schaepman et al., 2009). The spectral resolution is determined by the instruments: common field spectroradiometers used for ground measurements have a resolution of a few nanometres, whereas the resolution of remote sensing instruments is somewhat lower. The number of spectral bands in some upcoming instruments exceeds 200 (Buckingham & Staenz, 2008), covering the region of 400–2500 nm. However, approximately 30–40 spectral bands characterize virtually all the information (variability) in the data, with the precise number of useful channels depending on data interpretation (Price, 1997). In the future, an additional source of information would be the use of multiangular reflectance data (e.g. Heiskanen, 2006; Mõttus & Rautiainen, 2009) as the information content of the spectral and angular domains has been found to be highly complementary (Verhoef & Bach, 2007).

Despite advances in estimating primary production based on remote sensing, considerable progress is needed before the full carbon balance can be monitored from space. Ecosystem respiration is generally of similar magnitude to GPP; however, it is more difficult to monitor since it depends on temperature as well as a range of environmental factors. Several studies have modeled NEP based on a combination of remotely sensed and climatic data using models of different complexity (Veroustraete et al., 1996; Oechel et al., 2000; Hunt et al., 2004; Chiesi et al., 2005; Olofsson et al., 2008). Recognizing the difficulty in obtaining accurate climate and additional data for large-scale estimation of NEP, some authors have investigated possibilities to estimate NEP solely from remotely sensed measurements. Rahman et al. (2005) and Sims et al. (2006) found that GPP could be modeled directly from EVI without explicitly modeling LUE, and suggested that NEP can be obtained solely from satellite data based on observed relationships between MODIS surface temperature and ER. Xiao et al. (2010) modeled GPP based only on remotely sensed data products. Olofsson et al. (2008) found that large-scale modeling of NEP is feasible in Scandinavian forests; however, it requires spatially explicit data on the base respiration rate (R ₁₀).

Finally, the improvements both in forest reflectance modeling and remote sensing data processing will contribute to global retrieval of biophysical variables in the future. Any improvement in the quality of the surface reflectance data, including better instrumentation, radiometric calibration, cloud and snow masking and atmospheric correction, is likely to advance retrieval of land surface variables as those are limiting the accuracy of the current retrievals (Yang et al., 2006). The importance of long time series of key variables in ecological forest monitoring and modeling places emphasis on both the reprocessing of the historical satellite data record (AVHRR data since 1981) and incorporation of data from future sensors into spatially and temporally continuous data sets with novel algorithms (Ganguly et al., 2008). Continuous validation work as the new products are released continues to be important. Nevertheless, more emphasis is still needed in the collection of field data equivalent to satellite retrievals, data harmonization and continuous measurements of seasonal forest dynamics.

Acknowledgements

Our work has been supported by the Emil Aaltonen Foundation, University of Helsinki Research and Postdoctoral Funds, Academy of Finland, the Swedish National Space Board and Czech Ministry of Environment (ForChange SP/2d1/70/08, MZP).