Smoothed Versus Unsmoothed LiDAR in a Double-Sample Forest Inventory Robert C. Parker and A. Lee Mitchel, Department of Forestry, Forest and Wildlife Research Center, Mississippi State University, Mississippi State, MS 39762. ABSTRACT: Light detection and ranging (LiDAR) data at 0.5- and 1-m postings were used in a double-sample forest inventory on Louisiana State University���s Lee Experimental Forest, Louisiana. Phase 2 plots were established with differential global positioning system (DGPS). Tree dbh ( 4.5 in.) and two sample heights were measured on every 10th plot of the Phase 1 sample. Volume was computed for natural and planted pine and mixed hardwood species. LiDAR trees were selected with focal filter procedures from smoothed and unsmoothed LiDAR canopy surfaces. Dbh-height and ground-LiDAR height models were used to predict dbh from LiDAR height and compute Phase 2 estimates of ft2 basal area and ft3 volume. Phase 1 LiDAR estimates were computed by randomly assigning heights to species classes using the probability distribution from ground plots in each inventory strata. Phase 2 LiDAR estimates were computed by randomly assigning heights to species-product groups using a Monte Carlo simulation for each ground plot. Regression coefficients for Phase 2 estimates of ft2 and ft3 from the smoothed versus unsmoothed surfaces of high- and low-density LiDAR were computed by species group. Regression estimates for combined volume were partitioned by species-product distribution of Phase 2 volume. There was no statistical difference ( 0.05) between smoothed versus unsmoothed for high- and low-density LiDAR on adjusted mean volume estimates (sampling errors of 9.52 versus 8.46% for high-density and 9.25 versus 7.65% for low-density LiDAR). South. J. Appl. For. 29(1):40���47. Key Words: LiDAR, inventory, double-sample. Light detection and ranging (LiDAR) is a relatively new remote sensing tool that has the potential for use in the acquisition of measurement data for inventories of standing timber. LiDAR systems have been used in a variety of forestry applications (Nelson et al. 1988, Nilsson 1996, Magnussen and Boudewyn 1998, Lefsky et al. 1999, Means et al. 1999, Means et al. 2000) for the quantification of biomass, basal area, and tree and stand height estimates. Parker and Evans (2004) used small-footprint, multi-return LiDAR (nominal posting density of approximately 2 m 0.25 points per m2) in a double-sample application with a ground-based forest inventory in central Idaho and achieved a sampling error of 11.5% at the 95% level of confidence for mean cubic feet per acre (ft3/ac). Sampling error was cal- culated as one-half the confidence interval expressed as a percentage of the mean volume. Parker and Glass (2003) used high-density (4 points per m2) and low-density (1 point per m2) small-footprint, multi-return LiDAR in a double- sample, ground based forest inventory in central Louisiana and achieved sampling errors (at the 95% confidence level) for ft3/ac of 8.6% with low-density LiDAR using LiDAR- to-ground height adjustment and 7.6% with low-density LiDAR without using height adjustment. A linear regres- sion equation for ground height as a function of LiDAR height was used to adjust bias of the LiDAR heights. Low- density LiDAR with no height adjustment produced lower sampling errors for estimates of ft3/ac than high-density LiDAR with or without height adjustment. The authors believed the error associated with the height equation in- creased the sampling error of the double-sample volume estimate because of the high positive correlation between height and volume. Sampling errors associated with volume estimates from low-density LiDAR surfaces were lower than with high-density LiDAR because the fewer laser point coordinates decreased the ���noise��� during height determina- tion associated with false maxima. Small-footprint LiDAR will detect tree crowns and pen- etrate these surfaces recording elevation values for lower branches, main stem, understory vegetation, and ground. Because of this penetration, horizontal distribution of ele- vation values may not consistently represent the true canopy surface. Elevation change from point to point can be used to identify individual trees from the LiDAR surface. Because NOTE: Robert C. Parker can be reached at (662) 325-2775 Fax: (662) 325-8726 rparker@cfr.msstate.edu. Approved for publication as FWRC Publication No. FO255 of the Forest and Wildlife Research Center, Mississippi State University. Copyright �� 2005 by the Society of American Foresters. 40 SJAF 29(1) 2005

of the penetration capabilities of small-footprint LiDAR, the elevation values derived from the interpolated canopy sur- face may have enough variation to cause commission errors in tree recognition (McCombs et al. 2003). The primary objective of this study was to investigate the use of smoothed versus unsmoothed height surfaces from high- and low-density LiDAR in a double-sample forest inventory of pine and mixed hardwood stands with respect to predic- tions of ground ft3/ac with the LiDAR inventory variables ft3/ac and basal area per acre (ft2/ac). A secondary objective was to examine the partitioning of the per acre volume estimates into species-product classes by percent distribu- tion of volume and basal area on the ground plots. Study Area The study area (1,200 ac) was located on the Louisiana State University���s (LSU) Lee Experimental Forest near Bo- galusa in Washington Parish, Louisiana. Forests within this region are dominated by natural and planted stands of loblolly pine (Pinus taeda), natural shortleaf pine (P. echi- nata), longleaf pine (P. elliottii), and mixed hardwood stands of red oak (Quercus spp.), sweetgum (Liquidambar styraciflua), and hickory (Carya spp.). The cooperative project with LSU was funded by NASA Stennis Space Center appropriations through the Mississippi State Univer- sity, Remote Sensing Technology Center (MSU-RSTC). Methods LiDAR and Field Inventory Data Airborne1, Inc. acquired the small-footprint, multi-return LiDAR data of the study area in June 2002 with an Optech ALTM-1225 system to attain nominal posting spacings of 1 m (1 hit per m2, footprint size of 0.213 m, called low- density LiDAR) and 0.5 m (4 hits per m2, footprint size of 0.122 m, called high-density LiDAR) for two returns and intensity. Low-density data were obtained at an altitude of 1,067 m on a swath width of 609 m and high-density data from 610 m on a 189-m swath. Each return consisted of a UTM (Zone 15N, NAD 83) x, y, and z coordinate, where z was height above ellipsoid (HAE) in meters. The elevation of a global positioning system (GPS) point (i.e., HAE) is computed as height above a mathematical function or ellip- soid used to represent the shape of the earth rather than elevation above mean sea level. LiDAR data sets were surfaced to produce first return canopy and last return digital elevation models (DEM) with 0.2-m cell sizes using a linear interpolation technique. Tree locations and heights were determined with algorithms and focal filter procedures developed by McCombs et al. (2003) that used a variable search window radius based on relative density. These procedures used moving 2.5-, 4.0-, or 5.5-ft radius search windows to identify each tree peak as the point that is higher than 85% of the surrounding maxima from one of the three search window radius files. A spatial filtering technique derived from image analysis called smoothing was used to reduce commission errors by mini- mizing the abrupt elevation changes in the initial canopy surface. The Focal Analysis option in ERDAS��� Imagine software performed smoothing based on user-defined inputs for window size and preferred statistical procedure. A 5 5-pixel window was used to create a 1-m2 filter that would avoid removal of small peaks in the canopy surface (small trees), while maximizing the smoothing function. The filter moved across the LiDAR canopy surface, pixel by pixel, averaged the values within the window, and placed the result in the center pixel. Results of the smoothing process are shown in Figure 1. Because tree volumes would be ���adjusted��� with the double-sample protocol, the resulting minor change in tree height values from the averaging process should not affect the regression estimates of mean Figure 1. Identified trees (black squares) on unsmoothed (A) versus smoothed (B) LiDAR surfaces where gray scale values range from low (dark) to high (light) elevation. SJAF 29(1) 2005 41