from satellites are strongly dependent on tree canopy phenology, forest types, and
local environmental conditions. As a result, each study region should design its
own models to match its own condition and to suit related satellite data, or to use
the same variables but different coefficients.
Among various regression models, it is difficult to conclude outright that one
model is superior to another. Model applicability is dependent upon the locations
that a model applied and the criteria (such as RMS, R
2 , variance, bias, and overall
accuracy) that a validation/assessment is used (Powell et al. 2010). The latter are
always inconsistent. Generally, the models including multiple variables produce
relatively better estimates than single variable, and vegetation index-based models
better than single band models. Because of the uncertainty in the samples and
model establishment, linear regression models are recommended since non-linear
models could not always enhance the reliability statistics (Heiskanen 2006a, b).
Because there is no widely acceptable satellite variables across a range of forest
conditions, efforts are needed on comparisons of possible models in practical
purpose in a given region.
The non-parametric approach is optimal and robust for biomass estimates using
large number of input variables. It can implement various datasets, such as annual
time series of MODIS data, radar data, lidar, and other climate parameters. The
level of precision of biomass map is strongly dependent on the details of training
samples (Labrecque et al. 2006). If the training samples represent a very detailed
and broad range of real biomass values, the resultant biomass could be highly
accurate. Like the spectral-based regression models (Labrecque et al. 2006), the
non-parameteric methods rely on image-specific relationships and the transferability of these relationships to other images is usually difficult. In other words,
sufficient field samples over a research area are required to train the non-parametric
models. Moreover, to implement the non-parametric methods effectively, it would
be important to conduct a feature selection to extract variables which are highly
sensitive to forest biomass.
Satellite-based allometric models based on forest structural variables are
applicable across biomes once the models are established. The challenge is to
retrieve forest canopy attributes appropriately from satellite data. From passive
optical satellite data, canopy reflectance models are promising in retrieving forest
structural parameters using little or no field data (Soenen et al. 2010). As a result,
the satellite-based generalized allometric models are particularly useful for the
regions where little field measurements are available (Zhang and Kondragunta
2006). Radar and lidar data have advantages in calculating forest structures for
allometric biomass models.
Different satellite instruments serve as a tool to estimate biomass in various
spatial resolution and coverage. Lidar data and high-resolution passive satellite
imagery, such as QuickBird and IKNOS data, are optimal for the generation of
forest inventory at an individual tree crown scale in local areas (Bauer et al. 1997;
Wu and Strahler 1994; Gougeon and Leckie 1999; Wulder et al. 2000; García et al.
2010). Optical data at a medium–high spatial resolution produce biomass
distributions in a spatial stratification of vegetation. Data such as 30 m Landsat
86
X. Zhang and W. Ni-meister
local environmental conditions. As a result, each study region should design its
own models to match its own condition and to suit related satellite data, or to use
the same variables but different coefficients.
Among various regression models, it is difficult to conclude outright that one
model is superior to another. Model applicability is dependent upon the locations
that a model applied and the criteria (such as RMS, R
2 , variance, bias, and overall
accuracy) that a validation/assessment is used (Powell et al. 2010). The latter are
always inconsistent. Generally, the models including multiple variables produce
relatively better estimates than single variable, and vegetation index-based models
better than single band models. Because of the uncertainty in the samples and
model establishment, linear regression models are recommended since non-linear
models could not always enhance the reliability statistics (Heiskanen 2006a, b).
Because there is no widely acceptable satellite variables across a range of forest
conditions, efforts are needed on comparisons of possible models in practical
purpose in a given region.
The non-parametric approach is optimal and robust for biomass estimates using
large number of input variables. It can implement various datasets, such as annual
time series of MODIS data, radar data, lidar, and other climate parameters. The
level of precision of biomass map is strongly dependent on the details of training
samples (Labrecque et al. 2006). If the training samples represent a very detailed
and broad range of real biomass values, the resultant biomass could be highly
accurate. Like the spectral-based regression models (Labrecque et al. 2006), the
non-parameteric methods rely on image-specific relationships and the transferability of these relationships to other images is usually difficult. In other words,
sufficient field samples over a research area are required to train the non-parametric
models. Moreover, to implement the non-parametric methods effectively, it would
be important to conduct a feature selection to extract variables which are highly
sensitive to forest biomass.
Satellite-based allometric models based on forest structural variables are
applicable across biomes once the models are established. The challenge is to
retrieve forest canopy attributes appropriately from satellite data. From passive
optical satellite data, canopy reflectance models are promising in retrieving forest
structural parameters using little or no field data (Soenen et al. 2010). As a result,
the satellite-based generalized allometric models are particularly useful for the
regions where little field measurements are available (Zhang and Kondragunta
2006). Radar and lidar data have advantages in calculating forest structures for
allometric biomass models.
Different satellite instruments serve as a tool to estimate biomass in various
spatial resolution and coverage. Lidar data and high-resolution passive satellite
imagery, such as QuickBird and IKNOS data, are optimal for the generation of
forest inventory at an individual tree crown scale in local areas (Bauer et al. 1997;
Wu and Strahler 1994; Gougeon and Leckie 1999; Wulder et al. 2000; García et al.
2010). Optical data at a medium–high spatial resolution produce biomass
distributions in a spatial stratification of vegetation. Data such as 30 m Landsat
86
X. Zhang and W. Ni-meister