**3. Remote sensing techniques for biomass estimation in the miombo woodlands**

Remote Sensing refers to the acquisition and recording of information about objects on Earth without any physical contact between the sensor and the subject of analysis. It provides spatial coverage by measurement of reflected and emitted electromagnetic radiation, across a wide range of wavebands. Remotely sensed data provide in many ways an enhanced and very feasible alternative to manual observation with a very short time delay between data collection and transmission.

Since the first launch of an earth observatory satellite, in July 1972 by the US called Earth Resources Technology Satellite (ERTS, later renamed Landsat), remote sensing has been increasingly used to acquire information about environmental processes and over the last three decades this use has had a substantial increase with the development of modern remote sensing technology as radar sensors (ERS, RADARSAT, etc), LiDAR systems and the new generation of optical sensors (IKONOS, SPOT-5, GeoEye, etc.). Thus, remote sensing techniques are a time-and-cost-efficient method of observing forest ecosystems.

The capabilities of remote sensing techniques have been lately expanded to serve monitoring efforts as well as to provide valuable information on degradation in specific areas. Remote sensing, due to its potential to estimate biophysical parameters with detention of temporal and spatial variability, becomes a powerful technique. These, in combination with field sampling methods may provide detailed estimations of forestry parameters. In forest ecosystems, estimation of biophysical parameters has been made by a succession of methods. However, the extension of these estimations in space and time has obvious limitations especially in highly variable ecosystems such as miombo. Moreover, the network of field plots across the miombo ecoregion is not sufficient enough to cover it, which imposes further limitations in the use of remote sensing for biomass estimation.

In this section we highlight the uses of remote sensing techniques to derive biomass information in miombo. Emphasis is given to optical remote sensing techniques because the majority of measurements have been concentrated on these techniques. Radar and microwave remote sensing in miombo have, in the last decade, become a useful source of information for biomass estimation, but has been barely applied in the region. Therefore, radar and microwave remote sensing are also briefly discussed in this section.

Remote Sensing of Biomass in

wavelength.

data.

photosynthetic rate (Shugart *et al*., 2010).

the Miombo Woodlands of Southern Africa: Opportunities and Limitations for Research 85

seasonally with phenological changes in vegetation. For example in the dry and sub-humid miombo most species lose their leaves for 4-5 months of the year in response to the dry season. The associated spectral reflectance spectrum within a year is exploited to extract information on woodland structure, type, age, condition, biomass, leaf area, and even

Differential vegetation reflectance has resulted in the development of numerous vegetation indices and biomass estimation techniques that utilize multiple measurements in the visible and NIR. Vegetation indices (VIs) are dimensionless metrics that function as indicators of relative abundance and activity of green vegetation, often including LAI [most VIs saturate at around LAI = 4, Huete *et al*. (2002)], percentage green cover, chlorophyll content, green biomass, and Absorbed Photosynthetically Active Radiation (APAR) (Jensen, 1996). The first

> Re *NIR SR*

Where ρNIR is the reflectance in the NIR wavelength and ρRed is the reflectance in the red

Rouse *et al.* (1974) developed what is now known as the Normalized Vegetation Index the Normalized Differentiated Vegetation Index (NDVI), a satellite metric used to detect

*NIR d NDVI*

(eq. 2. 17 Symbol definitions are the same as in eq. 1.) NDVI is broadly correlated with both chlorophyll and water content of canopies and consequently is linked both theoretically and experimentally to LAI, the Fraction of Photosynthetically Active Radiation (fPAR) and the biomass of plant canopies (Gamon *et al.*, 1995; Sellers, 1985). SR and NDVI have been further developed and improved in other related vegetation indices such as the Soil Adjusted Vegetation Index (SAVI, Huete, 1988; Huete & Liu, 1994) and Enhanced Vegetation Index

*NIR REd EVI G*

reflectance in the blue it is more sensible to variability of canopy structure.

*d*

( Re ) ( Re )

1 Re 2

Where C1 = 6.0, atmosphere resistance red correction coefficient; C2 = 7.5, atmosphere resistance blue correction coefcient; L = 1.0, canopy background brightness correction factor; G = 2.5, gain factor. Given that EVI considers the canopy background (L) and

These indexes might be calculated with any sensor [ETM+, MODerate Resolution Imaging Spectroradiometer (MODIS)], except for EVI, which is calibrated exclusively for MODIS

Most optical remote sensing techniques for biomass estimation essentially explore the relationship between field measurements of vegetation parameters with spectral VIs to estimate large-scale distribution of biomass. The results vary according to the woodland structure and diversity derived from site-specific conditions such as the disturbance regime

*NIR C d C Blue L*

 

 

*NIR d*

 

 

(1)

(2)

(3)

true VI was the Simple Ratio (eq. 1.) described by Birth & McVey (1968).

changes in pixel-scale vegetation greenness (Asner, 2004).

presented in eq. 3 (EVI, Huete *et al*., 2002).

### **3.1 Remote sensing of biomass using optical sensors**

Passive optical remote sensing, the process of imaging or sampling the interactions between electromagnetic energy and matter at selected wavelengths, has the ability to monitor terrestrial ecosystems at various temporal and spatial scales and has been widely tested for land cover mapping and various forestry applications (Patenaude *et al*., 2004).

Remote sensing of vegetation depends on the optical properties of plant tissues (foliage, wood, litter, etc.) that determine the landscape-scale reflectance of ecosystems (Asner, 2004; Jensen, 1996). Thus, plant species composition, stand structures and associated canopy shadows and vegetation vigor (leaves, branches and bark) are important factors affecting vegetation reflectance (Lu, 2001). Figure 4 is an example of the spectral signatures of two different types of vegetation in comparison with sand and water. In this example is clear the discrepant reflectance of vegetation and other targets. The shape of the curves of the two vegetation types is similar, but they have dissimilar percent of reflectance, which is due to differences in vegetation structure and composition.

Fig. 4. Differential spectral signatures of vegetation, sand and water (adapted from http://rst.gsfc.nasa.gov/Intro/Part25.html). The Red Edge represents the basis for optical remote sensing of vegetation.

The basis for satellite measurements of vegetation from optical sensors is the differential reflection of light in the visible (0.4 to 0.7 μm) and Near-Infrared (NIR~0.8 μm) regions of the electromagnetic spectrum originated by the various leaf components (chlorophylls and βcarotene) and structure. For instance, chlorophyll strongly absorbs light in the red (~0.6 μm) while leaf's structural components highly reflect in the NIR (Jensen, 1996). In the visible part of the spectrum (Figure 4) the dominant signal detected by the sensor is the direct scattering. Multiple bounces of photons from the forest back to the sensor are small in comparison to direct scattering because the chlorophyll and other leaf pigments within the forest foliage are highly absorbing at visible wavelengths. At NIR wavelengths, however, the foliage is more transparent and multiple scattering becomes more important. Hence, this striking difference in light reflectance, known as the *red edge* (Figure 4), is sensitive to vegetation changes over a wider range of vegetation densities (Jensen, 1996; Shugart, *et al*., 2010).

Within the same forest stand the sensor measures the aggregate reflectance as a function of the wavelength of the tree canopies and understory within a pixel. The reflectance changes

Passive optical remote sensing, the process of imaging or sampling the interactions between electromagnetic energy and matter at selected wavelengths, has the ability to monitor terrestrial ecosystems at various temporal and spatial scales and has been widely tested for

Remote sensing of vegetation depends on the optical properties of plant tissues (foliage, wood, litter, etc.) that determine the landscape-scale reflectance of ecosystems (Asner, 2004; Jensen, 1996). Thus, plant species composition, stand structures and associated canopy shadows and vegetation vigor (leaves, branches and bark) are important factors affecting vegetation reflectance (Lu, 2001). Figure 4 is an example of the spectral signatures of two different types of vegetation in comparison with sand and water. In this example is clear the discrepant reflectance of vegetation and other targets. The shape of the curves of the two vegetation types is similar, but they have dissimilar percent of reflectance, which is due to

land cover mapping and various forestry applications (Patenaude *et al*., 2004).

**Red Edge**

Fig. 4. Differential spectral signatures of vegetation, sand and water (adapted from

wider range of vegetation densities (Jensen, 1996; Shugart, *et al*., 2010).

http://rst.gsfc.nasa.gov/Intro/Part25.html). The Red Edge represents the basis for optical

The basis for satellite measurements of vegetation from optical sensors is the differential reflection of light in the visible (0.4 to 0.7 μm) and Near-Infrared (NIR~0.8 μm) regions of the electromagnetic spectrum originated by the various leaf components (chlorophylls and βcarotene) and structure. For instance, chlorophyll strongly absorbs light in the red (~0.6 μm) while leaf's structural components highly reflect in the NIR (Jensen, 1996). In the visible part of the spectrum (Figure 4) the dominant signal detected by the sensor is the direct scattering. Multiple bounces of photons from the forest back to the sensor are small in comparison to direct scattering because the chlorophyll and other leaf pigments within the forest foliage are highly absorbing at visible wavelengths. At NIR wavelengths, however, the foliage is more transparent and multiple scattering becomes more important. Hence, this striking difference in light reflectance, known as the *red edge* (Figure 4), is sensitive to vegetation changes over a

Within the same forest stand the sensor measures the aggregate reflectance as a function of the wavelength of the tree canopies and understory within a pixel. The reflectance changes

**3.1 Remote sensing of biomass using optical sensors** 

differences in vegetation structure and composition.

remote sensing of vegetation.

seasonally with phenological changes in vegetation. For example in the dry and sub-humid miombo most species lose their leaves for 4-5 months of the year in response to the dry season. The associated spectral reflectance spectrum within a year is exploited to extract information on woodland structure, type, age, condition, biomass, leaf area, and even photosynthetic rate (Shugart *et al*., 2010).

Differential vegetation reflectance has resulted in the development of numerous vegetation indices and biomass estimation techniques that utilize multiple measurements in the visible and NIR. Vegetation indices (VIs) are dimensionless metrics that function as indicators of relative abundance and activity of green vegetation, often including LAI [most VIs saturate at around LAI = 4, Huete *et al*. (2002)], percentage green cover, chlorophyll content, green biomass, and Absorbed Photosynthetically Active Radiation (APAR) (Jensen, 1996). The first true VI was the Simple Ratio (eq. 1.) described by Birth & McVey (1968).

$$SR = \frac{\rho \text{NIR}}{\rho \text{Red}} \tag{1}$$

Where ρNIR is the reflectance in the NIR wavelength and ρRed is the reflectance in the red wavelength.

Rouse *et al.* (1974) developed what is now known as the Normalized Vegetation Index the Normalized Differentiated Vegetation Index (NDVI), a satellite metric used to detect changes in pixel-scale vegetation greenness (Asner, 2004).

$$\text{NDVI} = \frac{(\rho \text{NIR} - \rho \text{Re} \, d)}{(\rho \text{NIR} + \rho \text{Re} \, d)} \tag{2}$$

(eq. 2. 17 Symbol definitions are the same as in eq. 1.) NDVI is broadly correlated with both chlorophyll and water content of canopies and consequently is linked both theoretically and experimentally to LAI, the Fraction of Photosynthetically Active Radiation (fPAR) and the biomass of plant canopies (Gamon *et al.*, 1995; Sellers, 1985). SR and NDVI have been further developed and improved in other related vegetation indices such as the Soil Adjusted Vegetation Index (SAVI, Huete, 1988; Huete & Liu, 1994) and Enhanced Vegetation Index presented in eq. 3 (EVI, Huete *et al*., 2002).

$$EVI = G \left[ \frac{\rho \text{NIR} - \rho \text{RED}}{\left( \rho \text{NIR} + \text{C1} \, \rho \text{Red} - \text{C2} \, \rho \text{Blue} + \text{L} \right)} \right] \tag{3}$$

Where C1 = 6.0, atmosphere resistance red correction coefficient; C2 = 7.5, atmosphere resistance blue correction coefcient; L = 1.0, canopy background brightness correction factor; G = 2.5, gain factor. Given that EVI considers the canopy background (L) and reflectance in the blue it is more sensible to variability of canopy structure.

These indexes might be calculated with any sensor [ETM+, MODerate Resolution Imaging Spectroradiometer (MODIS)], except for EVI, which is calibrated exclusively for MODIS data.

Most optical remote sensing techniques for biomass estimation essentially explore the relationship between field measurements of vegetation parameters with spectral VIs to estimate large-scale distribution of biomass. The results vary according to the woodland structure and diversity derived from site-specific conditions such as the disturbance regime

Remote Sensing of Biomass in

disturbance (Chambers *et al*., 2007).

(logging, grazing, etc).

LAI, PAR and carbon uptake.

miombo in particular.

**3.2 Radar remote sensing of biomass** 

carbon pool in the forested ecosystems.

the Miombo Woodlands of Southern Africa: Opportunities and Limitations for Research 87

Fractional cover of photosynthetic vegetation (PV), NPV and bare substrate are the principal determinants of ecosystem composition, physiology, structure, biomass and biogeochemical stocks. These also capture the biophysical impact of disturbances caused by both natural and human drivers (Asner *et al*., 2005). In this case, SMA provides the fraction of per-pixel reflectance that is accounted for by the selected target. Results from SMA can be directly evaluated in field campaigns, including endmember abundance change over time. The temporal or spatial change in the NPV fraction can provide a quantitative measure of

Up to date SMA has not been applied to miombo but due to its capacity to proportionally decompose several endmembers in a pixel it is a promising remote sensing technique in this ecosystem given the spatial heterogeneity referred previously. Asner *et al.* (2005) applied SMA for the Brazilian Amazon and bordering Cerrado (an ecosystem structurally similar to miombo) to estimate vegetation attributes for a large-scale and multi-temporal Landsat ETM+ scene spanning the years 1999-2001. Their results showed that PV fractional covers for the cerrado range from 74% to 85% depending on the degree of woody cover, while the NPV cover varied from ~9% and 12% and bare soil between 2% and 10%. The study was able to demonstrate the capacity of this technique to separate vegetation types (forests from savannas) as well as different kinds of disturbances

In a particular interesting application Asner *et al*. (1998) applied SMA to invert a geometrical-optical model to estimate overstory stand-density and crown dimensions. The study applied SMA to a Landsat image to provide estimates of woody cover, herbaceous, bare soil and shade fractions. The estimated cover was then used to unmix the contribution of woody cover and herbaceous canopies to AVHRR multiangle reflectance data. The angular reflectance was used with the radiative transfer (RT) model inversions to estimate

There are two drawbacks of biomass estimation using optical remote sensing techniques: (i) eld plots are rarely designed to be related to spaceborne data and (ii) saturation at dense leaf canopies restricts estimates to low biomass levels when passive sensor data is used (Gibbs *et al*., 2007). Also, biomass is a three-dimension feature of vegetation. However, the ability of optical sensors is limited to two dimensions only, i.e. the upper layers of vegetation (Anaya *et al*., 2009). Remote-sensing systems relying on optical data (visible and infrared light) are further limited in the tropics by cloud cover. In view of these drawbacks microwave data has become an attractive technique to estimate biomass worldwide and in

The term radar remote sensing is used to denote the active remote sensing where microwave or radio frequency radiation is transmitted to the surface. Microwaves have higher wavelengths (3 to 75 cm for vegetation studies) than solar radiation (400 to 2,500 nm) that is used by optical remote sensing techniques. Unlike optical sensors, the microwave energy penetrates clouds, rain, dust, or fog and allows collection of images, regardless of solar illumination, so that the radar images can be generated at any time under the most varied weather conditions. The microwaves penetrate into the forest canopy and scatter from large woody components (stems and branches) that constitute the bulk of biomass and

and seasonality. Thus, we do not attempt to generalize the relationships between field measurements and VIs for miombo, but give some examples that reflect this variation.

Ribeiro *et al*. (2008)b worked in northern Mozambique in an area where fires occur annually during the dry season with higher incidence in the eastern side, where elephant density is also higher. In this area, the combination of fires and elephants imposes high variability of miombo structure (Ribeiro *et al*., 2008a). Under these conditions, the authors found relatively low but significant relationships between woody biomass,Leaf Area Index (LAI) measured in the field and, NDVI and SR derived from Landsat ETM+ [NDVI (rbiomass =0.30, rLAI = 0.35; p < 0.0001) and SR (rbiomass = 0.36, rLAI = 0.40, p < 0.0001)].

Huete *et al*. (2002) found that the NDVI saturated in high biomass regions like the Amazon (150 Mg/ha or over 15 years of age) while the EVI was sensitive to canopy variations. NDVI (r=0.15-0.25) and SR (0.12-0.24) showed to be weakly correlated with biomass in dense stands, while the correlation coefficients were considerable (rNDVI=0.46 and rSR=0.50) for the less dense site. Recently a study conducted in the savannas and woodlands of the Amazon Basin detected quite good relationships between field biomass and MODIS derived NDVI with an r2=0.45 (Saatchi *et al*., 2007).

A major constraint on the use of spectral VIs in dry and sub-humid ecosystems for vegetation studies, is that VIs are likely to underestimate live biomass due to their insensitivity to nonphotosynthetic vegetation (NPV) and sensitivity to low organic matter content in soils (Okin & Roberts, 2004). The effect of NPV is particularly important in the dry miombo where tree density is low (sometimes less than 10%) and disturbances may create several open patches in which NPV (such as litter, dry woody material, dry grass in the dry season) and soils are a significant component of the surface. In these areas, the effect of the exposed soil may be pronounced, while the low organic matter content on soils makes them bright and mineralogically heterogeneous.

The combination of those factors, creates vertical and horizontal structural differences that may not be depicted using VIs techniques, especially when using low to medium resolution sensors (e.g. 30-m Landsat, 250-m and greater MODIS, etc). In this situation, high spatial resolution optical sensors such as 1-4 m IKONOS, 2.5-10 m SPOT-5, among others, can detect variations in the spatial distribution of biomass density. This new generation of sensors also offer systematic observations at scales ranging from local to global and improves the monitoring of inaccessible areas (Rosenqvist *et al*., 2003). However, these high spatial resolution instruments are only available for small geographic areas in the miombo ecoregion and thus they are not suitable for large-scale high-resolution ecological assessments in this ecosystem.

The use of Spectral Mixture Analysis (SMA) has become an important technique to overcome the spatial and temporal heterogeneity in miombo. SMA uses a linear mixture technique to estimate the proportion of each ground's pixel area that belongs to different cover type (Okin & Roberts, 2004). SMA is based on the assumption that spectra of materials in a pixel combine linearly with proportion given by their relative abundance. For example, an NPV value of 0.50 from SMA indicates that woody tissues, surface litter and other dead vegetation occupy 50% of the surface area of the pixel. A combined spectrum can thus be decomposed in a linear mixture of its spectral endmembers, spectra of distinct material within the pixel. The weighting coefficients of each spectral endemember, which must sum up to 1, are then interpreted as the relative area occupied by each material in a pixel.

and seasonality. Thus, we do not attempt to generalize the relationships between field measurements and VIs for miombo, but give some examples that reflect this variation. Ribeiro *et al*. (2008)b worked in northern Mozambique in an area where fires occur annually during the dry season with higher incidence in the eastern side, where elephant density is also higher. In this area, the combination of fires and elephants imposes high variability of miombo structure (Ribeiro *et al*., 2008a). Under these conditions, the authors found relatively low but significant relationships between woody biomass,Leaf Area Index (LAI) measured in the field and, NDVI and SR derived from Landsat ETM+ [NDVI (rbiomass =0.30, rLAI = 0.35;

Huete *et al*. (2002) found that the NDVI saturated in high biomass regions like the Amazon (150 Mg/ha or over 15 years of age) while the EVI was sensitive to canopy variations. NDVI (r=0.15-0.25) and SR (0.12-0.24) showed to be weakly correlated with biomass in dense stands, while the correlation coefficients were considerable (rNDVI=0.46 and rSR=0.50) for the less dense site. Recently a study conducted in the savannas and woodlands of the Amazon Basin detected quite good relationships between field biomass and MODIS derived NDVI

A major constraint on the use of spectral VIs in dry and sub-humid ecosystems for vegetation studies, is that VIs are likely to underestimate live biomass due to their insensitivity to nonphotosynthetic vegetation (NPV) and sensitivity to low organic matter content in soils (Okin & Roberts, 2004). The effect of NPV is particularly important in the dry miombo where tree density is low (sometimes less than 10%) and disturbances may create several open patches in which NPV (such as litter, dry woody material, dry grass in the dry season) and soils are a significant component of the surface. In these areas, the effect of the exposed soil may be pronounced, while the low organic matter content on soils makes

The combination of those factors, creates vertical and horizontal structural differences that may not be depicted using VIs techniques, especially when using low to medium resolution sensors (e.g. 30-m Landsat, 250-m and greater MODIS, etc). In this situation, high spatial resolution optical sensors such as 1-4 m IKONOS, 2.5-10 m SPOT-5, among others, can detect variations in the spatial distribution of biomass density. This new generation of sensors also offer systematic observations at scales ranging from local to global and improves the monitoring of inaccessible areas (Rosenqvist *et al*., 2003). However, these high spatial resolution instruments are only available for small geographic areas in the miombo ecoregion and thus they are not suitable for large-scale high-resolution ecological

The use of Spectral Mixture Analysis (SMA) has become an important technique to overcome the spatial and temporal heterogeneity in miombo. SMA uses a linear mixture technique to estimate the proportion of each ground's pixel area that belongs to different cover type (Okin & Roberts, 2004). SMA is based on the assumption that spectra of materials in a pixel combine linearly with proportion given by their relative abundance. For example, an NPV value of 0.50 from SMA indicates that woody tissues, surface litter and other dead vegetation occupy 50% of the surface area of the pixel. A combined spectrum can thus be decomposed in a linear mixture of its spectral endmembers, spectra of distinct material within the pixel. The weighting coefficients of each spectral endemember, which must sum up to 1, are then interpreted as the relative area occupied

p < 0.0001) and SR (rbiomass = 0.36, rLAI = 0.40, p < 0.0001)].

with an r2=0.45 (Saatchi *et al*., 2007).

assessments in this ecosystem.

by each material in a pixel.

them bright and mineralogically heterogeneous.

Fractional cover of photosynthetic vegetation (PV), NPV and bare substrate are the principal determinants of ecosystem composition, physiology, structure, biomass and biogeochemical stocks. These also capture the biophysical impact of disturbances caused by both natural and human drivers (Asner *et al*., 2005). In this case, SMA provides the fraction of per-pixel reflectance that is accounted for by the selected target. Results from SMA can be directly evaluated in field campaigns, including endmember abundance change over time. The temporal or spatial change in the NPV fraction can provide a quantitative measure of disturbance (Chambers *et al*., 2007).

Up to date SMA has not been applied to miombo but due to its capacity to proportionally decompose several endmembers in a pixel it is a promising remote sensing technique in this ecosystem given the spatial heterogeneity referred previously. Asner *et al.* (2005) applied SMA for the Brazilian Amazon and bordering Cerrado (an ecosystem structurally similar to miombo) to estimate vegetation attributes for a large-scale and multi-temporal Landsat ETM+ scene spanning the years 1999-2001. Their results showed that PV fractional covers for the cerrado range from 74% to 85% depending on the degree of woody cover, while the NPV cover varied from ~9% and 12% and bare soil between 2% and 10%. The study was able to demonstrate the capacity of this technique to separate vegetation types (forests from savannas) as well as different kinds of disturbances (logging, grazing, etc).

In a particular interesting application Asner *et al*. (1998) applied SMA to invert a geometrical-optical model to estimate overstory stand-density and crown dimensions. The study applied SMA to a Landsat image to provide estimates of woody cover, herbaceous, bare soil and shade fractions. The estimated cover was then used to unmix the contribution of woody cover and herbaceous canopies to AVHRR multiangle reflectance data. The angular reflectance was used with the radiative transfer (RT) model inversions to estimate LAI, PAR and carbon uptake.

There are two drawbacks of biomass estimation using optical remote sensing techniques: (i) eld plots are rarely designed to be related to spaceborne data and (ii) saturation at dense leaf canopies restricts estimates to low biomass levels when passive sensor data is used (Gibbs *et al*., 2007). Also, biomass is a three-dimension feature of vegetation. However, the ability of optical sensors is limited to two dimensions only, i.e. the upper layers of vegetation (Anaya *et al*., 2009). Remote-sensing systems relying on optical data (visible and infrared light) are further limited in the tropics by cloud cover. In view of these drawbacks microwave data has become an attractive technique to estimate biomass worldwide and in miombo in particular.
