nodes (leaves). The splitting procedure stops either when the variability within a
node is considered sufficiently low (based on the deviance within the node) or
when a prescribed minimum number of cases is reached.
A regression tree approach has been widely used in biomass estimates from
time series MODIS data at large scales. Baccini et al. (2004) generated tree-based
models using MODIS Nadir BRDF (Bidirectional Reflectance Distribution
Function) Adjusted Reflectance (NBAR), climate, and topographic variables in
California. The model produced forest biomass in 1 km pixels with an RMSE of
44.4 Mg/ha. Similarly, time series of MODIS NBAR was employed to establish
regression tree models to predict AGB in Africa (Baccini et al. 2008), which
revealed that the model explained 82 % of the variance in AGB ranging from
0 to 545 Mg/ha, with a RMSE of 50.5 Mg/ha. Houghton et al. (2007) applied
bootstrapped regression trees to develop associations between mean MODIS
reflectance and biomass in Russia. After creating 500 regression trees using
different random samples of the data at 500 m resolution, forest biomass was
calculated with an error of *40 %.
The regression tree approach was also applied to calculate AGB at a spatial
resolution of 250 m across conterminous United States (CONUS) (Blackard et al.
2008). The variables used in the model were MODIS-derived land cover, Landsatderived National Land Cover Dataset (NLCD), topographic variables, monthly and
annual climate parameters, and time series of MODIS land surface reflectance and
vegetation index. The estimated biomass shows that pixel-based error ranges from
42 to 163 Mg/h and relative error from 0.51 to 0.92 in different regions where the
western regions had substantially better results than the eastern regions.
Powell et al. (2010) established RF models using a set of variables that were
Landsat TM Tasseled Cap indices, spectral indices, topographic variables, and
climate variables. The model was served to calculate biomass in Arizona and
Minnesota with a RMSE of 32 and 39 Mg/ha, respectively. A comparison indicates that the RF model consistently yields smaller RMSE than both multiple
regression models and a k-NN algorithm (Gradient Nearest Neighbor) do while RF
model produces relatively larger variance.
3.3.2.5 Satellite-Based Generalized Allometric Models
Generalized allometric model is a physically-based approach in forest biomass
determinations. Although the tree allometric models are generally species-specific
and site-specific, they are also generalized to estimate biomass in mixed species
across large regions (e.g., Jenkins et al. 2003; Wirth et al. 2004).
Foliage-based generalized allometric model can link remotely sensed data to
forest biomass over a continental scale (Zhang and Kondragunta 2006). Unlike
DBH, canopy leaf properties are sensitively reflected in passive optical remote
sensing and are widely measured from optical satellite data. Thus, a foliage-based
allometric model has been developed (Zhang and Kondragunta 2006):
3 Remote Sensing of Forest Biomass
77
node is considered sufficiently low (based on the deviance within the node) or
when a prescribed minimum number of cases is reached.
A regression tree approach has been widely used in biomass estimates from
time series MODIS data at large scales. Baccini et al. (2004) generated tree-based
models using MODIS Nadir BRDF (Bidirectional Reflectance Distribution
Function) Adjusted Reflectance (NBAR), climate, and topographic variables in
California. The model produced forest biomass in 1 km pixels with an RMSE of
44.4 Mg/ha. Similarly, time series of MODIS NBAR was employed to establish
regression tree models to predict AGB in Africa (Baccini et al. 2008), which
revealed that the model explained 82 % of the variance in AGB ranging from
0 to 545 Mg/ha, with a RMSE of 50.5 Mg/ha. Houghton et al. (2007) applied
bootstrapped regression trees to develop associations between mean MODIS
reflectance and biomass in Russia. After creating 500 regression trees using
different random samples of the data at 500 m resolution, forest biomass was
calculated with an error of *40 %.
The regression tree approach was also applied to calculate AGB at a spatial
resolution of 250 m across conterminous United States (CONUS) (Blackard et al.
2008). The variables used in the model were MODIS-derived land cover, Landsatderived National Land Cover Dataset (NLCD), topographic variables, monthly and
annual climate parameters, and time series of MODIS land surface reflectance and
vegetation index. The estimated biomass shows that pixel-based error ranges from
42 to 163 Mg/h and relative error from 0.51 to 0.92 in different regions where the
western regions had substantially better results than the eastern regions.
Powell et al. (2010) established RF models using a set of variables that were
Landsat TM Tasseled Cap indices, spectral indices, topographic variables, and
climate variables. The model was served to calculate biomass in Arizona and
Minnesota with a RMSE of 32 and 39 Mg/ha, respectively. A comparison indicates that the RF model consistently yields smaller RMSE than both multiple
regression models and a k-NN algorithm (Gradient Nearest Neighbor) do while RF
model produces relatively larger variance.
3.3.2.5 Satellite-Based Generalized Allometric Models
Generalized allometric model is a physically-based approach in forest biomass
determinations. Although the tree allometric models are generally species-specific
and site-specific, they are also generalized to estimate biomass in mixed species
across large regions (e.g., Jenkins et al. 2003; Wirth et al. 2004).
Foliage-based generalized allometric model can link remotely sensed data to
forest biomass over a continental scale (Zhang and Kondragunta 2006). Unlike
DBH, canopy leaf properties are sensitively reflected in passive optical remote
sensing and are widely measured from optical satellite data. Thus, a foliage-based
allometric model has been developed (Zhang and Kondragunta 2006):
3 Remote Sensing of Forest Biomass
77