Using airborne & space lidars for large-area inventory

Boudreau et al (2008) reports on a three-phase statistical framework for estimating forest biomass and carbon using a space LiDAR – ICESat/GLAS – as a large-area sampling tool to inventory the Province of Quebec. An airborne profiling LiDAR was used to tie ground plot observations to GLAS pulses by overflying both plots and pulses to facilitate development of two sets of regression equations. The first equation (or set of equations if the area is stratified) related ground-measured dry biomass to airborne LiDAR measurements. This first equation was then used to predict biomass on individual GLAS pulses overflown by the profiler in order to develop a GLAS-based equation to predict biomass regionally. Results indicated that, in the southern half of the 1.27 million km 2 study area, GLAS-based estimates were within 10% of the ground estimates. But the associated variance estimates, a sum of between-flight line sampling error and a covariance term that accounted for correlations between land cover strata, were unstable. This instability was manifest in variances which were occasionally negative due to a negative covariance term that overwhelmed the sampling error. A study to estimate forest basal area, volume, and biomass in Hedmark County, Norway (27,000 km 2) was undertaken to develop statistical techniques whereby airborne laser profilers and scanners could be used as sampling tools to inventory large areas. A central objective of this study was to derive variance estimators that could logically be applied to a two-phase or two-stage (air, ground) sampling scenario, a more tractable problem than the three-level situation faced in Quebec. Two potentially useful sampling frameworks were formulated, the first a model-based, two-phase design (the GS estimator, Ståhl et al. 2010) and the second a model-assisted two-stage design (the TG estimator, Gregoire et al. 2010). The same linear models were used to predict total aboveground dry biomass in both sampling frameworks. The GS estimator assumes that the laser model(s) used to predict biomass in particular strata are correctly specified, and a unique model must be specified for each stratum reported. Attempts were made to develop models for each of the four productive forest strata, but the best models in two of the four classes were unacceptably weak, exhibiting R 2 < 0.4. The four productive forest classes, then, were concatenated into one productive forest class. The design-based TG estimator does not rest on an assumption of a correctly specified model, but rather corrects the …