**1. Introduction**

#### **1.1 Leaf area index**

The green photosynthesizing leaf area of a canopy is an important characteristic of the status of the vegetation in terms of its health and production potential. At stand level, the amount of leaf area in a canopy is represented by a variable called the leaf area index (LAI), which is one of the key biophysical parameters in the global monitoring and mapping of vegetation by satellite remote sensing (Morisette et al., 2006). In this paper we adopt the, by now widely accepted, definition of LAI as the hemi-surface or half of the total surface area of all leaves or needles in the vegetation canopy divided by the horizontal ground area below the canopy. The definition is in line with the original definition of LAI, formulated for flat and (assumedly) infinitely thin leaves (Watson, 1947), as the one-sided leaf area per unit ground area. For coniferous canopies, the question arose on how to define the "one-sided" area of non-flat needles. While projected needle area formerly often has been used erroneously as a synonym to one-sided flat leaf area, it is now commonly accepted that the hemi-surface needle area represents the logical counterpart to the one-sided area of flat leaves (e.g. Chen & Black, 1992; Stenberg, 2006).

LAI controls many biological and physical processes, driving the exchange of matter and energy flow. Because LAI responds rapidly to different stress factors and changes in climatic conditions, monitoring of LAI yields a dynamic indicator of forest status and health. The link between forest productivity and LAI, in turn, lies in that LAI is the main determinant of the fraction of incoming photosynthetically active radiation absorbed by the canopy (fAPAR). The absorbed photosynthetically active radiation (APAR) quantifies the energy available for net primary production (NPP) and is thus a critical variable in NPP and carbon flux models. NPP is related to APAR by the light-use-efficiency originally introduced by Monteith (1977) for agricultural crops.

Traditionally, ground-based measurements of LAI have typically involved destructive sampling and determination of allometric relationships, e.g. between leaf area and the basal area of stem and/or branches carrying the leaves (the pipe model theory) (Shinozaki et al., 1964; Waring et al., 1982). However, such "direct methods" are quite laborious and indirect measurements of LAI using optical instruments are today the preferred choice (Welles &

Narrowband Vegetation Indices for Estimating Boreal Forest Leaf Area Index 5

of various geometric properties as the main reason for the reflectance differences between

Remote sensing of the biophysical properties, such as LAI, of a boreal coniferous forest canopy layer is further complicated by the often dominating role of the understory in the spectral signal (Rautiainen et al., 2011; Rautiainen et al., 2007; Eriksson et al., 2006; Eklundh et al., 2001; Chen & Cihlar, 1996; Spanner et al,. 1990). Coniferous forests that are regularly treated according to forest management practices tend to have relatively clumped and open canopies. Thus, the role of the understory vegetation in forming boreal forest reflectance

Canopy biophysical variables, such as LAI, can be estimated from remotely sensed data by two types of algorithms: empirical models and methods that use physically-based radiative transfer (RT) models. In empirical algorithms, the estimation is based on statistical relationships modelled between concurrent ground reference measurements and surface reflectance data. These relationships are typically expressed in the form of vegetation indices (VI). VIs include various combinations of spectral bands designed to maximize the sensitivity to vegetation characteristics while minimizing it to atmospheric conditions, background, view and solar angles (Baret & Guyot, 1991; Myneni et al., 1995). Operational LAI algorithms at global-scale typically make use of RT models, but the empirical models

The design of a VI that is optimally correlated with a particular vegetation property requires good physical understanding of the factors affecting the spectral signal reflected from vegetation. The sensitivity of a VI to a vegetation characteristic is typically maximized by including bands with high sensitivity (e.g. high absorption) to the monitored entity and bands mostly unaffected by the same entity. The simplest forms of VIs are simple differences (RB1–RB2), ratios (RB1/RB2) and normalized differences [(RB1-RB2)/(RB1+RB2)] of the reflectances of two spectral bands (RB1, RB2). (In Table 2 we give examples of common VIs used in this study.) The most apparent characteristic of the green vegetation spectrum is the pronounced difference between the red and NIR reflectances, the so called red-edge around 700 nm. For example, the normalized difference vegetation index (NDVI) utilizes this difference and has been shown to correlate with many interrelated vegetation attributes,

The most commonly used VIs were designed for broadband sensors (one spectral band spans about 50 nm or more) having red and NIR bands, such as NOAA AVHRR and Landsat MSS (e.g. Tucker, 1979). However, the basic VIs in red and NIR spectral range suffer from three well-known problems in LAI estimation: (1) they are not sensitive to LAI over its natural range but tend to saturate already at moderate levels of LAI, (2) they are sensitive to canopy background variability, and (3) the VI-LAI relationships are dependent on the vegetation type. These VIs are also sensitive to atmospheric noise and correction.

The saturation of NDVI occurs typically at LAI levels of 2 to 6 depending on the vegetation type and environmental conditions (e.g. Sellers, 1985; Myneni et al., 1997). In general, NDVI saturates as the fractional cover of vegetation approaches one, although LAI still increases (e.g. Carlson & Ripley, 1998). Over conifer-dominated boreal forests, NDVI varies typically

such as chlorophyll content, LAI, fractional cover, fAPAR and productivity.

broadleaved and coniferous stands.

cannot be neglected (Pisek et al., 2011).

**1.3 Vegetation indices in LAI estimation** 

usually outperform them in more localized applications.

Cohen, 1996; Jonckheere et al., 2004). They provide inverse estimates of LAI based on the fraction of gaps through the canopy in different directions, which can be measured using devices such as the LAI-2000 Plant Canopy Analyzer (LI-COR, 1992) or hemispherical photography. A vast body of classical literature exists on the dependency between LAI and canopy gap fraction underlying these techniques (e.g. Wilson, 1965; Miller, 1967; Nilson, 1971; Lang, 1986). In short, the inversion methods rely upon the assumption that leaves are randomly distributed in the canopy, in which case Beer's law can be applied to plant canopies (Monsi & Saeki, 1953). However, as the organization of leaves (needles) in forest canopies is typically more aggregated ("clumped") than predicted by a purely random distribution, the technique causes underestimation of LAI, especially in coniferous stands (e.g. Smith et al., 1993; Stenberg et al., 1994). Instead of the true LAI, the inversion of gap fraction data without correction for clumping yields the quantity commonly referred to as the "effective leaf area index" (Black et al., 1991).

Monitoring LAI in a spatially continuous mode and on a regular basis is possible only using remote sensing. Estimation of LAI from optical satellite images is considered feasible because LAI is closely linked to the spectral reflectance of plant canopies in the shortwave solar radiation range (Myneni et al., 1997). The physical relationships between canopy spectral reflectances and LAI form the basis of retrieval algorithms used in current Earth observation programs (e.g. MODIS, CYCLOPES, GLOBCARBON products) for mapping LAI at global scales. They produce bi-weekly and monthly vegetation maps that are widely used by biologists, natural resources managers, and climate modelers, e.g. to track seasonal fluctuations in vegetation or changes in land use. The arrival of narrowband reflectance data (also known as hyperspectral or imaging spectroscopy data) opens up new possibilities for satellite-derived estimation/monitoring of variables connected to the status and structure of vegetation, including LAI.

#### **1.2 Spectral properties of boreal forests**

The boreal forest zone, which spreads through Fennoscandia, Russia, Canada and Alaska, is the largest unbroken forest zone in the world and accounts for approximately one fourth of the world's forests. The boreal zone is a major store of carbon and thus plays an important role in determining global albedo and climate.

The reflectance spectra of coniferous forests (even if they have the same leaf area) are very distinct from similar broadleaved forests. The reasons for the special spectral behaviour of coniferous forests are versatile, yet primarily related to their structural, not optical, properties. Firstly, a high level of within-shoot scattering of conifers was originally noted nearly four decades ago (Norman & Jarvis, 1975). More recently, Landsat ETM+ data and a forest reflectance model were used to show that the low near infrared (NIR) reflectances observed in coniferous areas can largely be explained simply by within-shoot scattering (Rautiainen & Stenberg, 2005). Secondly, absorption by coniferous needles is higher than that by broadleaved species (Roberts et al., 2004; Williams, 1991), a phenomenon which can partly contribute to the lower reflectances of conifer-dominated areas. Other explanations include, for example, that the tree crown surface of coniferous stands is more heterogeneous than in broadleaved stands (Häme, 1991; Schull et al., 2011). In other words, when surface roughness (i.e. crown-level clumping) increases, the shaded area within the canopy increases, thus leading to lower reflectances. Overall, these results highlight the importance of various geometric properties as the main reason for the reflectance differences between broadleaved and coniferous stands.

Remote sensing of the biophysical properties, such as LAI, of a boreal coniferous forest canopy layer is further complicated by the often dominating role of the understory in the spectral signal (Rautiainen et al., 2011; Rautiainen et al., 2007; Eriksson et al., 2006; Eklundh et al., 2001; Chen & Cihlar, 1996; Spanner et al,. 1990). Coniferous forests that are regularly treated according to forest management practices tend to have relatively clumped and open canopies. Thus, the role of the understory vegetation in forming boreal forest reflectance cannot be neglected (Pisek et al., 2011).

#### **1.3 Vegetation indices in LAI estimation**

4 Remote Sensing – Applications

Cohen, 1996; Jonckheere et al., 2004). They provide inverse estimates of LAI based on the fraction of gaps through the canopy in different directions, which can be measured using devices such as the LAI-2000 Plant Canopy Analyzer (LI-COR, 1992) or hemispherical photography. A vast body of classical literature exists on the dependency between LAI and canopy gap fraction underlying these techniques (e.g. Wilson, 1965; Miller, 1967; Nilson, 1971; Lang, 1986). In short, the inversion methods rely upon the assumption that leaves are randomly distributed in the canopy, in which case Beer's law can be applied to plant canopies (Monsi & Saeki, 1953). However, as the organization of leaves (needles) in forest canopies is typically more aggregated ("clumped") than predicted by a purely random distribution, the technique causes underestimation of LAI, especially in coniferous stands (e.g. Smith et al., 1993; Stenberg et al., 1994). Instead of the true LAI, the inversion of gap fraction data without correction for clumping yields the quantity commonly referred to as

Monitoring LAI in a spatially continuous mode and on a regular basis is possible only using remote sensing. Estimation of LAI from optical satellite images is considered feasible because LAI is closely linked to the spectral reflectance of plant canopies in the shortwave solar radiation range (Myneni et al., 1997). The physical relationships between canopy spectral reflectances and LAI form the basis of retrieval algorithms used in current Earth observation programs (e.g. MODIS, CYCLOPES, GLOBCARBON products) for mapping LAI at global scales. They produce bi-weekly and monthly vegetation maps that are widely used by biologists, natural resources managers, and climate modelers, e.g. to track seasonal fluctuations in vegetation or changes in land use. The arrival of narrowband reflectance data (also known as hyperspectral or imaging spectroscopy data) opens up new possibilities for satellite-derived estimation/monitoring of variables connected to the status and structure of

The boreal forest zone, which spreads through Fennoscandia, Russia, Canada and Alaska, is the largest unbroken forest zone in the world and accounts for approximately one fourth of the world's forests. The boreal zone is a major store of carbon and thus plays an important

The reflectance spectra of coniferous forests (even if they have the same leaf area) are very distinct from similar broadleaved forests. The reasons for the special spectral behaviour of coniferous forests are versatile, yet primarily related to their structural, not optical, properties. Firstly, a high level of within-shoot scattering of conifers was originally noted nearly four decades ago (Norman & Jarvis, 1975). More recently, Landsat ETM+ data and a forest reflectance model were used to show that the low near infrared (NIR) reflectances observed in coniferous areas can largely be explained simply by within-shoot scattering (Rautiainen & Stenberg, 2005). Secondly, absorption by coniferous needles is higher than that by broadleaved species (Roberts et al., 2004; Williams, 1991), a phenomenon which can partly contribute to the lower reflectances of conifer-dominated areas. Other explanations include, for example, that the tree crown surface of coniferous stands is more heterogeneous than in broadleaved stands (Häme, 1991; Schull et al., 2011). In other words, when surface roughness (i.e. crown-level clumping) increases, the shaded area within the canopy increases, thus leading to lower reflectances. Overall, these results highlight the importance

the "effective leaf area index" (Black et al., 1991).

vegetation, including LAI.

**1.2 Spectral properties of boreal forests** 

role in determining global albedo and climate.

Canopy biophysical variables, such as LAI, can be estimated from remotely sensed data by two types of algorithms: empirical models and methods that use physically-based radiative transfer (RT) models. In empirical algorithms, the estimation is based on statistical relationships modelled between concurrent ground reference measurements and surface reflectance data. These relationships are typically expressed in the form of vegetation indices (VI). VIs include various combinations of spectral bands designed to maximize the sensitivity to vegetation characteristics while minimizing it to atmospheric conditions, background, view and solar angles (Baret & Guyot, 1991; Myneni et al., 1995). Operational LAI algorithms at global-scale typically make use of RT models, but the empirical models usually outperform them in more localized applications.

The design of a VI that is optimally correlated with a particular vegetation property requires good physical understanding of the factors affecting the spectral signal reflected from vegetation. The sensitivity of a VI to a vegetation characteristic is typically maximized by including bands with high sensitivity (e.g. high absorption) to the monitored entity and bands mostly unaffected by the same entity. The simplest forms of VIs are simple differences (RB1–RB2), ratios (RB1/RB2) and normalized differences [(RB1-RB2)/(RB1+RB2)] of the reflectances of two spectral bands (RB1, RB2). (In Table 2 we give examples of common VIs used in this study.) The most apparent characteristic of the green vegetation spectrum is the pronounced difference between the red and NIR reflectances, the so called red-edge around 700 nm. For example, the normalized difference vegetation index (NDVI) utilizes this difference and has been shown to correlate with many interrelated vegetation attributes, such as chlorophyll content, LAI, fractional cover, fAPAR and productivity.

The most commonly used VIs were designed for broadband sensors (one spectral band spans about 50 nm or more) having red and NIR bands, such as NOAA AVHRR and Landsat MSS (e.g. Tucker, 1979). However, the basic VIs in red and NIR spectral range suffer from three well-known problems in LAI estimation: (1) they are not sensitive to LAI over its natural range but tend to saturate already at moderate levels of LAI, (2) they are sensitive to canopy background variability, and (3) the VI-LAI relationships are dependent on the vegetation type. These VIs are also sensitive to atmospheric noise and correction.

The saturation of NDVI occurs typically at LAI levels of 2 to 6 depending on the vegetation type and environmental conditions (e.g. Sellers, 1985; Myneni et al., 1997). In general, NDVI saturates as the fractional cover of vegetation approaches one, although LAI still increases (e.g. Carlson & Ripley, 1998). Over conifer-dominated boreal forests, NDVI varies typically

Narrowband Vegetation Indices for Estimating Boreal Forest Leaf Area Index 7

biochemical composition of vegetation (water, nitrogen, cellulose and lignin), on the other hand, are reviewed by Kokaly et al. (2009). Many of the developed indices have been designed to work at leaf level and do not necessarily upscale to canopy level, because of the high sensitivity to canopy structure, background, solar and view geometry. Another approach is to find iteratively the simple combinations of bands that give the best

Most chlorophyll indices exploit the information in the red edge around 700 nm (Ustin et al., 2009). Imaging spectroscopy data also enables the estimation of the red edge position (REP), which is particularly sensitive to changes in chlorophyll content (e.g. Dawson & Curran, 1998). Water indices, on the other hand, utilize the water absorbing regions in the SWIR region of the spectrum (e.g. Gao, 1996; Zarco-Tejada et al., 2003). Those indices seem particularly interesting for LAI estimation considering the importance of the SWIR spectral

There is growing evidence that imaging spectroscopy data can improve LAI estimates in comparison to broadband data by reducing the saturation effects. Depending on the vegetation type and range of LAI, different types of VIs have been found useful. However, the red edge indices have been most effective in estimating LAI of crops (Wu et al., 2010), grasslands (Mutanga & Skidmore, 2004) and thicket shrubs (Brantley et al., 2011). On the other hand, indices based on NIR and SWIR bands have been successful in broadleaved (le Maire et al., 2008) and coniferous forests (Gong et al., 2003; Schlerf et al., 2005; Pu et al., 2008). The importance of the SWIR spectral region in estimating boreal forest LAI has also been emphasized by multivariate regression analysis (e.g. Lee et al., 2004). However, broadband sensors can also have advantages over narrowband sensors in LAI estimation, for example, by being less sensitive to noise due to the sensor, atmosphere and background (e.g. Broge & Leblanc, 2000). Although there are case studies from different biomes, the performance of narrowband VIs has been poorly assessed over European boreal forests.

The aim of the study is to establish the extent to which vegetation indices can be used to measure variation in LAI based on a test site in southern boreal forest in Finland. We explore different VIs in LAI estimation during full leaf development. We compare the performance of narrowband VIs to traditional broadband VIs. The objective is to identify VIs, which are least sensitive to species composition and, on the other hand, perform well in

The study area, Hyytiälä, is located in the southern boreal zone in central Finland (61° 50'N, 24°17'E) and has an annual mean temperature of 3°C and precipitation of 700 mm. Dominant tree species in the Hyytiälä forest area are Norway spruce (*Picea abies* (L.) Karst), Scots pine (*Pinus sylvestris* L.) and Silver birch (*Betula pendula* Roth). Understory vegetation, on the other hand, is composed of two layers: an upper understory layer (low dwarf shrubs

correlation with empirical data (e.g. Mutanga & Skidmore, 2004; Schlerf et al., 2005).

region in estimating LAI using broadband indices.

**2. Case study** 

coniferous stands.

**2.2.1 Study area** 

**2.2 Materials and methods** 

**2.1 Aims** 

in a narrow range and shows poor relationships with canopy LAI (Chen & Cihlar, 1996; Stenberg et al., 2004). The reason for this is the green understory, which results in a noncontrasting background in the visible part of the spectrum (Nilson & Peterson, 1994; Myneni et al., 1997).

Many modifications of basic VIs have been suggested to give better sensitivity to LAI. Typical modifications use other visible bands than red (e.g. the green vegetation index, GNDVI, Gitelson et al., 1996), try to reduce soil effects based on the soil line concept (e.g. the soil adjusted vegetation index, SAVI, Huete, 1988), or include short wave infrared (SWIR) bands. Many modifications also attempt to reduce atmospheric effects (e.g. the enhanced vegetation index, EVI, Huete et al., 2002). The soil line is based on the observation that soil reflectances fall in a line in the red-NIR spectral space (e.g. Huete, 1988). Many VIs utilize the parameterized soil line in their calculation, but these VIs have not been successful in boreal forests as bare soil is rarely visible (e.g. Chen, 1996).

The sensitivity of shortwave infrared (SWIR) reflectance to forest biophysical variables has been recognized for a long time (e.g. Butera, 1986; Horler & Ahern, 1986) and several VIs utilizing the SWIR band have been designed. Rock et al. (1986) showed that the moisture stress index (MSI), i.e. the ratio of SWIR reflectance to NIR reflectance, was an indicator of forest damage. Later, the ratio has commonly been referred to as the infrared simple ratio (ISR, Chen et al., 2002; Fernandes et al., 2003). The SWIR reflectance has also been used for adjusting NDVI (Nemani et al., 1993) and SR (Brown et al., 2000). The reduced simple ratio (RSR) has been used specifically for estimating LAI (Brown et al., 2000; Stenberg et al., 2004) and has been employed also in regional and global-scale operational algorithms (Chen et al., 2002; Deng et al., 2006). RSR seems to reduce the sensitivity to the type and amount of understory vegetation, because background reflectance varies less in SWIR than in visible and NIR (Brown et al., 2000; Chen et al., 2002). RSR has also some capability to unify coniferous and broadleaved forest types, which reduces the need for land cover type specific LAI algorithms. However, in comparison to ISR, the use of red band makes RSR sensitive to atmospheric effects (Fernandes et al., 2003). However, although inclusion of SWIR reflectance increases the sensitivity of VIs to LAI, these indices also have a tendency to saturate at high levels of LAI (e.g. Brown et al., 2000; Heiskanen et al., 2011).

Imaging spectroscopy provides much narrower spectral bands than typical multispectral sensors. Due to the more detailed sampling of the vegetation spectra, such data can detect specific absorption features of vegetation and therefore improve the estimation of vegetation biochemical properties. For example, the SPOT 5 HRG sensors capture a spectral range from 500 nm to 1750 nm with four broad bands, in comparison to Hyperion's 242 (10 nm wide) bands between 400 nm and 2500 nm. At the canopy scale, the contents of biochemical components and LAI are highly inter-related (e.g. Asner, 1998; Roberts et al., 2004). Therefore, imaging spectroscopy could potentially improve LAI estimates. Furthermore, there is potentially complementary information outside the typical spectral bands of broadband sensors.

One way to utilize imaging spectroscopy data is to calculate narrow-band VIs in a similar fashion as for broadband data but using narrower bands. The aim is to improve the sensitivity of the VI to a specific vegetation biochemical property. For example, Ustin et al. (2009) give a comprehensive review on VIs used as indicators of plant pigments (chlorophyll, carotenoids and anthocyanin). The methods of estimating the non-pigment biochemical composition of vegetation (water, nitrogen, cellulose and lignin), on the other hand, are reviewed by Kokaly et al. (2009). Many of the developed indices have been designed to work at leaf level and do not necessarily upscale to canopy level, because of the high sensitivity to canopy structure, background, solar and view geometry. Another approach is to find iteratively the simple combinations of bands that give the best correlation with empirical data (e.g. Mutanga & Skidmore, 2004; Schlerf et al., 2005).

Most chlorophyll indices exploit the information in the red edge around 700 nm (Ustin et al., 2009). Imaging spectroscopy data also enables the estimation of the red edge position (REP), which is particularly sensitive to changes in chlorophyll content (e.g. Dawson & Curran, 1998). Water indices, on the other hand, utilize the water absorbing regions in the SWIR region of the spectrum (e.g. Gao, 1996; Zarco-Tejada et al., 2003). Those indices seem particularly interesting for LAI estimation considering the importance of the SWIR spectral region in estimating LAI using broadband indices.

There is growing evidence that imaging spectroscopy data can improve LAI estimates in comparison to broadband data by reducing the saturation effects. Depending on the vegetation type and range of LAI, different types of VIs have been found useful. However, the red edge indices have been most effective in estimating LAI of crops (Wu et al., 2010), grasslands (Mutanga & Skidmore, 2004) and thicket shrubs (Brantley et al., 2011). On the other hand, indices based on NIR and SWIR bands have been successful in broadleaved (le Maire et al., 2008) and coniferous forests (Gong et al., 2003; Schlerf et al., 2005; Pu et al., 2008). The importance of the SWIR spectral region in estimating boreal forest LAI has also been emphasized by multivariate regression analysis (e.g. Lee et al., 2004). However, broadband sensors can also have advantages over narrowband sensors in LAI estimation, for example, by being less sensitive to noise due to the sensor, atmosphere and background (e.g. Broge & Leblanc, 2000). Although there are case studies from different biomes, the performance of narrowband VIs has been poorly assessed over European boreal forests.
