TM/ETM+ data are most frequently used for calculating forest biomass in regional
areas.
MODIS data seem the best option for the investigation of forest biomass in
national and continental scales. Although the spatial resolution is relatively coarse
(250–1,000 m), time series of MODIS data contain forest phenological variation
which is an important variable in producing a trained model that generally represents forest properties well for biomass estimates. However, MODIS pixels are
generally a mixture of forests and non-forests, which affect the accuracy of results.
Indeed, models established from field plots of a single tree species are better than
those with multiple species (Eklundh et al. 2003) and most errors and unexplained
variation in the biomass models are from complex vegetation composition
(Heiskanen 2006a, b).
Using discrete small footprint lidar data, the individual tree-based approach
provides biomass estimates at different levels. This approach permits the estimation of parameters at the tree level rather than at the plot or stand level. It has
advantages in highly fragmented forests. However the individual tree-based
approach may not be able to separate individual trees in dense forests.
Large footprint full waveform lidar data, based on statistical regression models,
can provide accurate estimates of AGB at plot and stand levels. This method
predicts field-measured AGB with a large variation of accuracies and uncertainties
with the correlation coefficients ranging from 0.65 to 0.96 and with RMSEs from 4
to 80 Mg/ha (Lefsky et al. 1999a, b, c, 2002; Drake et al. 2002, 2003; Nelson et al.
1988; Popescu et al. 2004; Lim et al. 2003; Lim and Treitz 2004; Ni-Meister et al.
2010). The large variations come from using different lidar systems, different lidar
sensed vegetation structure parameters, and different site conditions. In general,
the higher the lidar point cloud density is, the better the accuracy is. Using
combination of different height metrics achieves better accuracies than using
maximum canopy height as an AGB predictor. However it remains challenging to
accurately estimate biomass over dense deciduous forests at present.
Rader data have the advantage of weather- and daylight-independency. These
data are quite useful in the investigation of biomass in tropical forests where
cloud-free satellite data are rare. The current use of radar sensors to measure
biomass is limited mainly in low biomass regions and biomass change due to
deforestation using L- and P-band backscattering.
Finally, saturation is a common issue in biomass estimates using passive
optimal satellite and radar data. Because of the saturation of reflectance values,
models tend to underestimate large biomass densities and overestimate small ones
(Cohen et al. 2003; Blackard et al. 2008). The threshold of saturation varies with
satellite data. From MODIS data, aboveground biomass is slightly underestimated
for the areas where the biomass is larger than *300 Mg/ha in tropical Africa
(Baccini et al. 2008) and 250 Mg/ha and it is over-predicted biomass values below
45Mg/ha in California (Baccini et al. 2004; Zhang and Kondragunta 2006).
3 Remote Sensing of Forest Biomass
87
areas.
MODIS data seem the best option for the investigation of forest biomass in
national and continental scales. Although the spatial resolution is relatively coarse
(250–1,000 m), time series of MODIS data contain forest phenological variation
which is an important variable in producing a trained model that generally represents forest properties well for biomass estimates. However, MODIS pixels are
generally a mixture of forests and non-forests, which affect the accuracy of results.
Indeed, models established from field plots of a single tree species are better than
those with multiple species (Eklundh et al. 2003) and most errors and unexplained
variation in the biomass models are from complex vegetation composition
(Heiskanen 2006a, b).
Using discrete small footprint lidar data, the individual tree-based approach
provides biomass estimates at different levels. This approach permits the estimation of parameters at the tree level rather than at the plot or stand level. It has
advantages in highly fragmented forests. However the individual tree-based
approach may not be able to separate individual trees in dense forests.
Large footprint full waveform lidar data, based on statistical regression models,
can provide accurate estimates of AGB at plot and stand levels. This method
predicts field-measured AGB with a large variation of accuracies and uncertainties
with the correlation coefficients ranging from 0.65 to 0.96 and with RMSEs from 4
to 80 Mg/ha (Lefsky et al. 1999a, b, c, 2002; Drake et al. 2002, 2003; Nelson et al.
1988; Popescu et al. 2004; Lim et al. 2003; Lim and Treitz 2004; Ni-Meister et al.
2010). The large variations come from using different lidar systems, different lidar
sensed vegetation structure parameters, and different site conditions. In general,
the higher the lidar point cloud density is, the better the accuracy is. Using
combination of different height metrics achieves better accuracies than using
maximum canopy height as an AGB predictor. However it remains challenging to
accurately estimate biomass over dense deciduous forests at present.
Rader data have the advantage of weather- and daylight-independency. These
data are quite useful in the investigation of biomass in tropical forests where
cloud-free satellite data are rare. The current use of radar sensors to measure
biomass is limited mainly in low biomass regions and biomass change due to
deforestation using L- and P-band backscattering.
Finally, saturation is a common issue in biomass estimates using passive
optimal satellite and radar data. Because of the saturation of reflectance values,
models tend to underestimate large biomass densities and overestimate small ones
(Cohen et al. 2003; Blackard et al. 2008). The threshold of saturation varies with
satellite data. From MODIS data, aboveground biomass is slightly underestimated
for the areas where the biomass is larger than *300 Mg/ha in tropical Africa
(Baccini et al. 2008) and 250 Mg/ha and it is over-predicted biomass values below
45Mg/ha in California (Baccini et al. 2004; Zhang and Kondragunta 2006).
3 Remote Sensing of Forest Biomass
87