“…Because we also wish to gauge the importance of the sill ratio variables when used in combination with other spatial and spectral information for modeling forest structure, we employ partial least squares (PLS) regression. While PLS has been used extensively with hyperspectral data (Ourcival et al, 1999;Smith et al, 2002Smith et al, , 2003Townsend et al, 2003;Coops et al, 2003;McDonald et al, 2003), Wolter et al (2008) demonstrated the capability of this approach to estimate forest BA and species composition using broad band, satellite sensor data. PLS regression is convenient in this regard as it allows simultaneous modeling of multiple continuous predictor variables, does not make unrealistic assumptions about spectral or ground measurement error, as in ordinary least-squares regression (Curran & Hay, 1986;Cohen et al, 2003), and addresses the problem of collinearity (dependence) among multiple independent and dependent variables (Helland, 1988 Semivariance and other texture measures applied to remote sensing data have been used extensively to identify unique forest structure (Woodcock & Strahler, 1987;Woodcock et al (1988), Cohen et al, 1990;Franklin et al, 2001;Song & Woodcock, 2002;Coburn & Roberts, 2004).…”