where Y is forest biomass, a and b are regression coefficients, X is a independent
parameter including a vegetation index, spectral reflectance, or satellite-based
forest attribute, and e is the residual.
The variables of either X or Y in this simple format of model could be a
logarithm transformation value. In this way, the non-linear correlation between
biomass and satellite-derived variables are easily established. The model coefficients (a and b) are commonly determined using the ordinary least-squares (OLS)
approach with the assumption that the independent variable is accurately
measured. The regression slope from OLS will be biased if spectral bands and
vegetation indices are measured with errors. Alternatively, the Reduced Major
Axis (RMA) regression, an orthogonal regression technique that minimizes error
in both the X and Y directions (Larsson 1993), is believed to be more appropriate
at modeling forest biomass (Powell et al. 2010).
Either vegetation index or spectral reflectance has been correlated to AGB in
numerous simple linear equations in a local region (Foody et al. 2003; Lu et al.
2004; Rhaman et al. 2005; Roy and Ravan 1996; Heiskanen 2006a, b). The model
significance varies greatly with the type of spectral variables and local environment conditions (Foody et al. 2003; Lu 2006). Indeed, spectral biomass models
developed using shortwave infrared bands are more reliable as compared to the
visible bands which are more sensitive to atmospheric changes (Roy and Ravan
1996). Vegetation indices involving the red spectral band correlates poorly to
forest biomass in Brazilian but strongly in Malaysia (Lu et al. 2004). AGB in
Canadian forests has no relation to red, NIR, and SWIR reflectance (R
2 = 0.01,
0.05, and 0.09, respectively) or to the NDVI (R
2 = 0.03) and it is also weakly
associated with the SWVI (short-wave vegetation index) computed from the NIR
and SWIR (short-wave infrared) channels (R
2 = 0.25) (Fraser and Li 2002).
Simple band ratios produce higher correlation with AGB than complex vegetation
indices do (Lu et al. 2004). TC brightness and wetness parameters show very
strong relationship with the biomass values (Roy and Ravan 1996).
SF from QuickBird HRSI linearly correlates to tree biomass and is able to
produce a biomass map effectively in a high spatial resolution (Leboeuf et al.
2007) while the RMSE varies from 11 to 18 Mg/ha and bias from 2 to 5 Mg/ha in
different test sites. In contrast, SF from IKONOS is related to biomass in a logarithmic form and exhibits a saturation level in large biomass values (Hall et al.
1995; Jasinski and Crago 1999). SF at sub-pixel of SPOT imagery from spectral
mixture modeling produces AGB at both deciduous and conifer plots with a RMSE
of 32.6 Mg/ha (Soenen et al. 2010).
LAI is also strongly correlated to AGB as demonstrated by regression equations. The linear relationships are generally significant in the deciduous forests in
the Western Ghats of Karnataka, India (R
2 = 0.63) (Madugundu et al. 2008) and
the logarithmic relationship works well in low-density forests and savanna
woodlands during the dry season (R
2 = 0.66) (Saatchi et al. 2007a, b).
3 Remote Sensing of Forest Biomass
73
parameter including a vegetation index, spectral reflectance, or satellite-based
forest attribute, and e is the residual.
The variables of either X or Y in this simple format of model could be a
logarithm transformation value. In this way, the non-linear correlation between
biomass and satellite-derived variables are easily established. The model coefficients (a and b) are commonly determined using the ordinary least-squares (OLS)
approach with the assumption that the independent variable is accurately
measured. The regression slope from OLS will be biased if spectral bands and
vegetation indices are measured with errors. Alternatively, the Reduced Major
Axis (RMA) regression, an orthogonal regression technique that minimizes error
in both the X and Y directions (Larsson 1993), is believed to be more appropriate
at modeling forest biomass (Powell et al. 2010).
Either vegetation index or spectral reflectance has been correlated to AGB in
numerous simple linear equations in a local region (Foody et al. 2003; Lu et al.
2004; Rhaman et al. 2005; Roy and Ravan 1996; Heiskanen 2006a, b). The model
significance varies greatly with the type of spectral variables and local environment conditions (Foody et al. 2003; Lu 2006). Indeed, spectral biomass models
developed using shortwave infrared bands are more reliable as compared to the
visible bands which are more sensitive to atmospheric changes (Roy and Ravan
1996). Vegetation indices involving the red spectral band correlates poorly to
forest biomass in Brazilian but strongly in Malaysia (Lu et al. 2004). AGB in
Canadian forests has no relation to red, NIR, and SWIR reflectance (R
2 = 0.01,
0.05, and 0.09, respectively) or to the NDVI (R
2 = 0.03) and it is also weakly
associated with the SWVI (short-wave vegetation index) computed from the NIR
and SWIR (short-wave infrared) channels (R
2 = 0.25) (Fraser and Li 2002).
Simple band ratios produce higher correlation with AGB than complex vegetation
indices do (Lu et al. 2004). TC brightness and wetness parameters show very
strong relationship with the biomass values (Roy and Ravan 1996).
SF from QuickBird HRSI linearly correlates to tree biomass and is able to
produce a biomass map effectively in a high spatial resolution (Leboeuf et al.
2007) while the RMSE varies from 11 to 18 Mg/ha and bias from 2 to 5 Mg/ha in
different test sites. In contrast, SF from IKONOS is related to biomass in a logarithmic form and exhibits a saturation level in large biomass values (Hall et al.
1995; Jasinski and Crago 1999). SF at sub-pixel of SPOT imagery from spectral
mixture modeling produces AGB at both deciduous and conifer plots with a RMSE
of 32.6 Mg/ha (Soenen et al. 2010).
LAI is also strongly correlated to AGB as demonstrated by regression equations. The linear relationships are generally significant in the deciduous forests in
the Western Ghats of Karnataka, India (R
2 = 0.63) (Madugundu et al. 2008) and
the logarithmic relationship works well in low-density forests and savanna
woodlands during the dry season (R
2 = 0.66) (Saatchi et al. 2007a, b).
3 Remote Sensing of Forest Biomass
73