We show that the number of KLD (Karhunen-Lo`eve decomposition) modes D_KLD(f)
needed to capture a fraction f of the total variance of an extensively chaotic
state scales extensively with subsystem volume V. This allows a correlation
length xi_KLD(f) to be defined that is easily calculated from spatially
localized data. We show that xi_KLD(f) has a parametric dependence similar to
that of the dimension correlation length and demonstrate that this length can
be used to characterize high-dimensional inhomogeneous spatiotemporal chaos.