**3.2 Radar remote sensing of biomass**

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 carbon pool in the forested ecosystems.

Remote Sensing of Biomass in

prediction in the low biomass miombo in this area.

been used in Miombo woodlands (Shugart *et al*., 2010).

the Miombo Woodlands of Southern Africa: Opportunities and Limitations for Research 89

saturation at high biomass levels. The relationship between radar and field biomass in this area was further explored by Mitchard *et al*. (2009) by using ALOS PALSAR, an L-band sensor. The authors also studied other forested areas in Africa (Cameroon and Uganda). For all sites they found an improved relationship between field biomass and HV-backscatter (r2biomass=0.61-0.76, p<0.0001), but with a clear saturation between 150-200 Mg/ha of biomass. Biomass prediction was done with fairly good accuracy of ±20% for plots with less the 150 Mg/ha. According to the authors these are partly because L-band SAR does not respond directly to aboveground biomass, but to aspects of vegetation structure, partially due to spatial variability in structure and partially due to radar calibration aorthorectification and field estimation errors propagating through the analysis. Also, backscatter responds differently to differing soil and vegetation moisture conditions, and the surface topography, adding to observed prediction errors. Despite these factors the analysis was able to predict, with fairly good accuracy, aboveground biomass from radar data for very different in vegetation types. This finding suggests that utilization of L-band data should be essential for projects involving the mapping and monitoring of woodland and savanna biomass, thus having important implications for carbon-credit projects, such as those under proposed REDD schemes (Mitchard *et al*., 2009). The comparison of Ribeiro *et al*. (2008)b and Mitchard *et al*. (2009) study for Niassa National Reserve in northern Mozambique indicate clearly that L-band is an improvement over C-band for biomass

Light Detection and Ranging (LiDAR) measurements provide the most direct estimates of canopy height and the vertical structure of canopy foliage. Together, these measurements enable ecologists to quantify the 3D distribution of vegetation at a landscape scale, to understand processes of carbon accumulation and forest succession and to improve the state of ecosystem models (Chambers *et al*., 2007). LiDAR systems incorporate a laser altimeter to measure accurately the distance from the sensor to the canopy top and bottom elevations. The energy returned from distances between the canopy and ground provides evidence of the vertical distribution of sub-canopy strata. One application of LiDAR technology has been to map variations in canopy height with meter-level accuracy. Canopy heights can then be translated into estimates of aboveground biomass based on the allometric relationships between height, basal area and biomass (Shugart *et al*., 2010). This airborne system has not

Other approaches for biomass estimation from radar sensors develop regression equations to use different channels of radar imagery for average height and basal diameter estimation and then use these parameters to estimate biomass (Kellndorfer *et al*., 2004; Simard *et al*., 2006). Kellndorfer *et al.* (2004) used the Shuttle Radar Topography Mission (SRTM) data in conjunction with a National Elevation Dataset (NED) to estimate pine canopy height in Southeast Georgia, USA. The study indicates that SRTM can be successfully correlated via linear regression modeling with ground-measured mean canopy height (r2=0.79-0.86). Mean canopy height can *a posteriori* be used in allometric equation to estimate landscape biomass. In a recent study Simard *et al.* (2006) calibrated the SRTM data using Light Detection and Range (LiDAR) data and high resolution Digital Elevation Model (DEM) for the mangrove forests in Everglades National Park. The resulting mangrove tree height map (error 2.0 m) was then used in combination field data to map the spatial distribution of biomass for the entire area. This kind of approach has not been used yet for the miombo or similar woodlands but the results above indicate that the technique can be applied to miombo with some caution, due to

The microwaves in a specific wavelength and polarization is emitted from the sensor to the Earth, where it interacts with vegetation (and other objects) and is then backscattered to the sensor. Three different bands are commonly used in radar remote sensing of vegetation: Pband (30 cm) L-band (λ=23.5 cm), C-band (λ=5.8 cm) and X-band (λ=3.1 cm). The sensitivity of backscatter measurements at different wavelengths and polarization (horizontal and/or vertical) to the size and orientation of woody components and their density makes the radar sensors suitable for direct measurements of aboveground woody biomass (carbon stock) and structural attributes such as volume and basal area.

Radar remote sensing has been widely used to map vegetation worldwide, but it is not commonly used in miombo. But, given the spatial variability of vegetation in this ecosystem, radar is a promising technique to estimate biomass. Another advantage of the radar system for miombo is its insensitivity to weather conditions, making wet season measurements possible. During that time of the year some areas are inaccessible. The constraint though is that radar imagery does not cover large areas in the miombo region nor it is being collected frequently, causing difficulties for large-scale frequent monitoring of biomass.

Pierce *et al.* (2003) refer that Synthetic Aperture Radar (SAR) is known to have a response directly related to the amount of living material that it interacts with. The strong relationship between field biomass and radar image intensity and the development of approaches to estimate aboveground biomass represents one of the unique applications of SAR data in ecology (Kasischke *et al.*, 2004). SAR sensor offers the potential of rapid, accurate, high resolution and low cost mapping of the lower biomass density vegetation of Africa. Moreover, the 46-day repeat cycle of ALOS/PALSAR allow sufficient images to be captured during the year to negate any effects of seasonality and soil moisture, and allow the monitoring of landscapes for any changes in aboveground biomass (Mitchard *et al*., 2009).

Interferometric SAR (InSAR), which uses the phase information in signals received taken at multiple times to derive very precise measurements of movements on Earth, has been widely used. Recent innovations in orbital designs for repeat pass radar interferometry (InSAR) will allow the sensor to measure height of forest and provide a vertical dimension for accurately resolving the vegetation biomass of forests globally (Chambers *et al*., 2007).

The most basic approach to estimating aboveground biomass in forests is to develop a multiple-linear regression equation that estimates total biomass as a function of a combination of SAR channels or ratio of different channels (Kasischke *et al*., 2004). For example Pierce *et al.* (2003) studied the relationship between radar L-band (25 cm wavelength) and C-band (5.2 cm wavelength) combined and separately, in the regrowth forest of the Amazon Basin. The study found that L- and C-band together improve the accuracy compared to using only one frequency channel. In a recent study Saatchi *et al.* (2007) found high correlation between radar L-band and field biomass (r2=0.68) in woodlands and savannas of the Amazon Basin. The authors were able to produce a multiple linear regression model that includes microwave and optical data to estimate the biomass of this vegetation type.

Ribeiro *et al*. (2008)b correlated contemporary 30-m C-band (5.8 cm) RADARSAT backscatter and field woody biomass and LAI data (rbiomass= 0.65 and rLAI=0.57, p<0.0001) to, in combination with optical data (30-m Landsat ETM+), produce the aboveground biomass map for the Niassa National Reserve (NNR) in northern Mozambique. The results were satisfactory but expected to underestimate biomass in dense woodlands due to radar

The microwaves in a specific wavelength and polarization is emitted from the sensor to the Earth, where it interacts with vegetation (and other objects) and is then backscattered to the sensor. Three different bands are commonly used in radar remote sensing of vegetation: Pband (30 cm) L-band (λ=23.5 cm), C-band (λ=5.8 cm) and X-band (λ=3.1 cm). The sensitivity of backscatter measurements at different wavelengths and polarization (horizontal and/or vertical) to the size and orientation of woody components and their density makes the radar sensors suitable for direct measurements of aboveground woody biomass (carbon stock)

Radar remote sensing has been widely used to map vegetation worldwide, but it is not commonly used in miombo. But, given the spatial variability of vegetation in this ecosystem, radar is a promising technique to estimate biomass. Another advantage of the radar system for miombo is its insensitivity to weather conditions, making wet season measurements possible. During that time of the year some areas are inaccessible. The constraint though is that radar imagery does not cover large areas in the miombo region nor it is being collected

Pierce *et al.* (2003) refer that Synthetic Aperture Radar (SAR) is known to have a response directly related to the amount of living material that it interacts with. The strong relationship between field biomass and radar image intensity and the development of approaches to estimate aboveground biomass represents one of the unique applications of SAR data in ecology (Kasischke *et al.*, 2004). SAR sensor offers the potential of rapid, accurate, high resolution and low cost mapping of the lower biomass density vegetation of Africa. Moreover, the 46-day repeat cycle of ALOS/PALSAR allow sufficient images to be captured during the year to negate any effects of seasonality and soil moisture, and allow the monitoring of landscapes for any changes in aboveground biomass (Mitchard *et al*.,

Interferometric SAR (InSAR), which uses the phase information in signals received taken at multiple times to derive very precise measurements of movements on Earth, has been widely used. Recent innovations in orbital designs for repeat pass radar interferometry (InSAR) will allow the sensor to measure height of forest and provide a vertical dimension for accurately resolving the vegetation biomass of forests globally (Chambers *et al*., 2007). The most basic approach to estimating aboveground biomass in forests is to develop a multiple-linear regression equation that estimates total biomass as a function of a combination of SAR channels or ratio of different channels (Kasischke *et al*., 2004). For example Pierce *et al.* (2003) studied the relationship between radar L-band (25 cm wavelength) and C-band (5.2 cm wavelength) combined and separately, in the regrowth forest of the Amazon Basin. The study found that L- and C-band together improve the accuracy compared to using only one frequency channel. In a recent study Saatchi *et al.* (2007) found high correlation between radar L-band and field biomass (r2=0.68) in woodlands and savannas of the Amazon Basin. The authors were able to produce a multiple linear regression model that includes microwave and optical data to estimate the biomass of

Ribeiro *et al*. (2008)b correlated contemporary 30-m C-band (5.8 cm) RADARSAT backscatter and field woody biomass and LAI data (rbiomass= 0.65 and rLAI=0.57, p<0.0001) to, in combination with optical data (30-m Landsat ETM+), produce the aboveground biomass map for the Niassa National Reserve (NNR) in northern Mozambique. The results were satisfactory but expected to underestimate biomass in dense woodlands due to radar

frequently, causing difficulties for large-scale frequent monitoring of biomass.

and structural attributes such as volume and basal area.

2009).

this vegetation type.

saturation at high biomass levels. The relationship between radar and field biomass in this area was further explored by Mitchard *et al*. (2009) by using ALOS PALSAR, an L-band sensor. The authors also studied other forested areas in Africa (Cameroon and Uganda). For all sites they found an improved relationship between field biomass and HV-backscatter (r2 biomass=0.61-0.76, p<0.0001), but with a clear saturation between 150-200 Mg/ha of biomass. Biomass prediction was done with fairly good accuracy of ±20% for plots with less the 150 Mg/ha. According to the authors these are partly because L-band SAR does not respond directly to aboveground biomass, but to aspects of vegetation structure, partially due to spatial variability in structure and partially due to radar calibration aorthorectification and field estimation errors propagating through the analysis. Also, backscatter responds differently to differing soil and vegetation moisture conditions, and the surface topography, adding to observed prediction errors. Despite these factors the analysis was able to predict, with fairly good accuracy, aboveground biomass from radar data for very different in vegetation types. This finding suggests that utilization of L-band data should be essential for projects involving the mapping and monitoring of woodland and savanna biomass, thus having important implications for carbon-credit projects, such as those under proposed REDD schemes (Mitchard *et al*., 2009). The comparison of Ribeiro *et al*. (2008)b and Mitchard *et al*. (2009) study for Niassa National Reserve in northern Mozambique indicate clearly that L-band is an improvement over C-band for biomass prediction in the low biomass miombo in this area.

Light Detection and Ranging (LiDAR) measurements provide the most direct estimates of canopy height and the vertical structure of canopy foliage. Together, these measurements enable ecologists to quantify the 3D distribution of vegetation at a landscape scale, to understand processes of carbon accumulation and forest succession and to improve the state of ecosystem models (Chambers *et al*., 2007). LiDAR systems incorporate a laser altimeter to measure accurately the distance from the sensor to the canopy top and bottom elevations. The energy returned from distances between the canopy and ground provides evidence of the vertical distribution of sub-canopy strata. One application of LiDAR technology has been to map variations in canopy height with meter-level accuracy. Canopy heights can then be translated into estimates of aboveground biomass based on the allometric relationships between height, basal area and biomass (Shugart *et al*., 2010). This airborne system has not been used in Miombo woodlands (Shugart *et al*., 2010).

Other approaches for biomass estimation from radar sensors develop regression equations to use different channels of radar imagery for average height and basal diameter estimation and then use these parameters to estimate biomass (Kellndorfer *et al*., 2004; Simard *et al*., 2006). Kellndorfer *et al.* (2004) used the Shuttle Radar Topography Mission (SRTM) data in conjunction with a National Elevation Dataset (NED) to estimate pine canopy height in Southeast Georgia, USA. The study indicates that SRTM can be successfully correlated via linear regression modeling with ground-measured mean canopy height (r2=0.79-0.86). Mean canopy height can *a posteriori* be used in allometric equation to estimate landscape biomass. In a recent study Simard *et al.* (2006) calibrated the SRTM data using Light Detection and Range (LiDAR) data and high resolution Digital Elevation Model (DEM) for the mangrove forests in Everglades National Park. The resulting mangrove tree height map (error 2.0 m) was then used in combination field data to map the spatial distribution of biomass for the entire area. This kind of approach has not been used yet for the miombo or similar woodlands but the results above indicate that the technique can be applied to miombo with some caution, due to

Remote Sensing of Biomass in

conservation areas.

**limitations for research** 

by Hufkens *et al.,* 2008; Haining, 1990).

temporal and spatial variations in biomass.

region.

the Miombo Woodlands of Southern Africa: Opportunities and Limitations for Research 91

This technique is promising in estimating biomass and other parameters for miombo, due its high spatial variability associated to disturbances across the region. Relatively to herbivory by elephants (tree debarking and debranching) and fires (killing of juvenile trees and natural regeneration, soil degradation, killing of adult debarked trees, etc.) data fusion from these instruments may have a significant contribution in addressing the impacts of these disturbances on biomass production in the region. The immediate result will be an improved an informed management of the woodlands in particular for

Future satellite missions plan to make frequent measuring of standing vegetation feasible. Houghton *et al*. (2009) laid out the required specications for these and other future satellite missions to signicantly reduce the nowadays existing uncertainties. They claimed, that in order to reduce the uncertainty in the land-atmosphere uxes to those of the next uncertain term (which is the net carbon uptake of the ocean with an uncertainty of ± 18%) a measurement error of less than 2 MgC (or an aboveground biomass uncertainty of about 4 Mg) per ha is required. It is furthermore argued, that disturbances (deforestation, fires and herbivory) are patchy on spatial scales of 100 m and less and only if remote sensing is operating on a similar scale, one can clearly identify changes in carbon storage over time

and minimize sampling errors due to averaging (Kohler & Huth, 2010).

**4. Remote sensing of biomass in miombo woodlands: Opportunities and** 

Miombo display complex vegetation patterns in which dense vegetation alternates with sparsely populated or bare soil in response to environmental and disturbance (deforestation, fires and herbivory) factors. Low vegetation cover, in some places, and small-scale variations in others, can produce unpredictable errors in the quantication of biophysical and ecological properties of the vegetation. Ignoring this spatial variation can produce inaccurate results, even in fairly homogeneous environments (Aubry & Debouzie, 2001 cited

The use of remote sensing for estimations of vegetation biomass has proved to be of great importance to fill up data gaps and to estimate large-scale variations, especially in low accessible places. The technique lacks of precision when compared to detailed forest inventories. But, for miombo the lack of detailed field data and uncertainty of biomass stocks associated to disturbances, make remote sensing one important technique to address

Although repeated satellite imaging has improved in resolution over the years, it is still limited in detecting fine patterns within savanna vegetation. Some forms of remote sensing, such as the sub-meter resolution IKONOS, GeoEye and QUICKBIRD satellite sensors, allow individual trees to be recognized. Other, such as microwave remote sensing from radar (RADARSAT, ALOS-PALSAR, JERS-1, etc.) and LiDAR provide a three dimensional representation of vegetation, which is an improvement over optical remote sensing. However, both low-resolution optical and microwave scenes are currently too expensive for large-scale or regional studies and they require a substantive amount of processing capacities. These represent some of the major limitations for its use in the southern Africa

Discriminating between subtypes of savanna vegetation, even simply looking at structural differences, has proved a taxing undertaking, especially in places where field data is limited.

its heterogeneity. This may imply for example that the canopy height estimation may only apply to canopy species, while the understory is not completely measured. This should be considered a topic of research for miombo, before any conclusion is raised.

Radar and LiDAR sensors provide complementary information about the forest structure. LiDAR is sensitive to eaf material and radar to structural features, which can be combined to increase accuracy of biomass and the forest structure estimates (Figure 5). However, the signatures have some level of commonality because of biophysical and structural nature of forest stands. The vertical distribution of reflective surfaces that can be inferred from LiDAR reveals the bole and branch structure supporting the leaves within a near vertical volume of the vegetation. LiDAR sensors measure this vertical profile by sampling the forest stand along its orbital tracks. Radar provides imaging capability to estimate forest height through InSAR configuration or forest volume and biomass through polarimetric backscatter power. However, the vegetation signature from radar measurements is from a slanted volume and is sensitive to both vertical and horizontal arrangement of vegetation components (leaves, branches, and stems). The combination of the two sensors has the capability of providing the vegetation three‐dimensional structure at spatial resolutions suitable for ecological studies (25–100 m). However, there are limited studies to explore the fusion of the two measurements because of the lack of data over the same study areas as well as welldeveloped ground data (Shugart *et al*., 2010). According to the authors this will be an ongoing research area for several years to come.

Fig. 5. Radar imager maps of observed radar energy returned at various polarization transmit-receive combinations (HH, HV, VV, where H is horizontal polarization and V is vertical polarization) which are related to the volume and biomass of forest components. The multibeam lidar sampler measures the vertical variation in the strength of the scattered laser signal, which is related to forest vertical structure profile and biomass. The two signals are combined using "fusion" algorithms to improve the accuracy of radar estimates of biomass and to extend lidar measurements of structure in both space and time (Source: Shugart *et al*., 2010).

its heterogeneity. This may imply for example that the canopy height estimation may only apply to canopy species, while the understory is not completely measured. This should be

Radar and LiDAR sensors provide complementary information about the forest structure. LiDAR is sensitive to eaf material and radar to structural features, which can be combined to increase accuracy of biomass and the forest structure estimates (Figure 5). However, the signatures have some level of commonality because of biophysical and structural nature of forest stands. The vertical distribution of reflective surfaces that can be inferred from LiDAR reveals the bole and branch structure supporting the leaves within a near vertical volume of the vegetation. LiDAR sensors measure this vertical profile by sampling the forest stand along its orbital tracks. Radar provides imaging capability to estimate forest height through InSAR configuration or forest volume and biomass through polarimetric backscatter power. However, the vegetation signature from radar measurements is from a slanted volume and is sensitive to both vertical and horizontal arrangement of vegetation components (leaves, branches, and stems). The combination of the two sensors has the capability of providing the vegetation three‐dimensional structure at spatial resolutions suitable for ecological studies (25–100 m). However, there are limited studies to explore the fusion of the two measurements because of the lack of data over the same study areas as well as welldeveloped ground data (Shugart *et al*., 2010). According to the authors this will be an

Fig. 5. Radar imager maps of observed radar energy returned at various polarization transmit-receive combinations (HH, HV, VV, where H is horizontal polarization and V is vertical polarization) which are related to the volume and biomass of forest components. The multibeam lidar sampler measures the vertical variation in the strength of the scattered laser signal, which is related to forest vertical structure profile and biomass. The two signals are combined using "fusion" algorithms to improve the accuracy of radar estimates of biomass and to extend lidar measurements of structure in both space and time (Source:

considered a topic of research for miombo, before any conclusion is raised.

ongoing research area for several years to come.

Shugart *et al*., 2010).

This technique is promising in estimating biomass and other parameters for miombo, due its high spatial variability associated to disturbances across the region. Relatively to herbivory by elephants (tree debarking and debranching) and fires (killing of juvenile trees and natural regeneration, soil degradation, killing of adult debarked trees, etc.) data fusion from these instruments may have a significant contribution in addressing the impacts of these disturbances on biomass production in the region. The immediate result will be an improved an informed management of the woodlands in particular for conservation areas.

Future satellite missions plan to make frequent measuring of standing vegetation feasible. Houghton *et al*. (2009) laid out the required specications for these and other future satellite missions to signicantly reduce the nowadays existing uncertainties. They claimed, that in order to reduce the uncertainty in the land-atmosphere uxes to those of the next uncertain term (which is the net carbon uptake of the ocean with an uncertainty of ± 18%) a measurement error of less than 2 MgC (or an aboveground biomass uncertainty of about 4 Mg) per ha is required. It is furthermore argued, that disturbances (deforestation, fires and herbivory) are patchy on spatial scales of 100 m and less and only if remote sensing is operating on a similar scale, one can clearly identify changes in carbon storage over time and minimize sampling errors due to averaging (Kohler & Huth, 2010).
