Airborne LiDAR sensors are operated using an aircraft-mounted scanning laser altimeter, which emits short burst laser pulses. The round trip travel time between the emitted laser pulse, its interaction with a target, and its return to the sensor is measured, thus enabling a range (distance) measurement between the aircraft and the surface. Precise geographic coordinates and attitude (pitch, roll, yaw) of the aircraft are captured using a Differential Global Positioning System (DGPS) and an inertial measurement unit (IMU), respectively (Wehr and Lohr 1999). The exact location and pointing direction of the laser instrument are thus known. Each LiDAR return is represented by a (x;y;z) coordinate, where z refers to height or range; a large collection of such (x;y;z) returns is referred to as a “point cloud”. Resulting LiDAR height point clouds can be used to create detailed three-dimensional models of the area of interest, along with estimates of the vertical distribution of sub-canopy strata (Haala and Brenner 1999; Zimble et al. 2003). For example, several studies have shown that the height estimates derived from LiDAR height point clouds are as accurate and sometimes more accurate than traditional methods (Holmgren 2004). In fact, the 3D nature of LiDAR height data makes this technology especially suitable for assessing forest volume and biomass. Such LiDAR-based forest measurements are not only useful to forest inventory and canopy structure modelling (Lefsky et al. 2002a, b; Nssset 2002; Popescu et al. 2002,2004), but also

to estimation of forest fuel loads (Riano et al. 2003; Seielstad and Queen 2003), and extraction of digital elevation models (DEMs) (Popescu et al. 2002; Hodgson et al. 2003); all of these aspects are essential to forest management and site mapping.

LiDAR volume modeling has focused on both tree — and plot-based approaches (Nilsson 1996; Popescu et al. 2004), as well as stand-level assessments (van Aardt et al. 2006). These are based on either concrete surface modeling, i. e., top-of-canopy and ground, or LiDAR height distributional approaches, based on extraction of height distributional parameters (e. g., median, mode, percentiles) from all LiDAR returns within spatial units, such as grid cells or segments (stands) (Means et al. 2000; Nssset 2002; van Aardt et al. 2006, 2008). Whichever approach is selected, either tree-, plot-, or stand-level, or the use of surface height models vs. LiDAR distributions, the user should always keep in mind that the scale and geography will dictate the approach. For instance, homogenous even-aged stands are well-suited to LiDAR surface modeling (Fig. 2.5), while mixed stands with uneven ages are often best assessed by exploiting the complexity of LiDAR distributions throughout the canopy layer. Some specific examples are discussed next.

The use of locally varying window sizes, applied to height grids, is one popular approach for tree crown detection, height estimation, and determining average LiDAR values within a predefined area. Popescu et al. (2002) assessed both static and variable window size approaches for determining plot-level tree height. The authors reported that the maximum height within a predefined area returned an R2 of between 0.85 and 0.90, while an R2 of 0.85 was obtained for plot-level tree height when the local varying window size filters were applied. The local filter algorithm produced improved results in the case of co-dominant trees when compared to the maximum height approach. An alternative method includes the use of quartile-based LiDAR height metrics (Magnussen and Boudewyn 1998; Andersen et al. 2005; Nssset and Gobakken 2005), which makes use of distributional measures of LiDAR point height as opposed to absolute LiDAR height values (Holmgren et al. 2003; van Aardt et al. 2006,2008). Segmentation and regression modelling (Coops et al. 2004; van Aardt et al. 2008) and statistical analyses using field enumerated data have also been successfully employed (Nilsson 1996). van Aardt et al. (2008), for instance, showed that LiDAR distributions can be used to (i) segment a complex coniferous, deciduous, and mixed stand environment into homogenous structural stands, (ii) assess volume and biomass for these stands, and even (iii) map taxonomic types (coniferous vs. deciduous), for a wall-to-wall forest inventory by forest type. However, an estimate is of little use if not accompanied by an associated precision.

Most LiDAR-based forest inventory studies report high R2 values, except in highly variable or complex forest environments. And while most studies report root mean square error (RMSE) values that approach field inventory efforts, e. g., in the region of 10 % of the mean estimate, LiDAR is by no means a panacea. The calibration-validation approach of using a small sample of locally measured variables when transporting LiDAR models across regions or species, remains an absolute requirement. Many efforts have been made to assess LiDAR accuracy and precision when it comes to forest inventory, and while most such studies report these values, a couple of efforts or implications deserve special mention:

Duncanson et al. (2010) used an airborne LiDAR dataset in combination with forest inventory data to explore the relationship between their model error and canopy height, aboveground biomass (AGB), stand age, canopy rugosity, mean diameter at breast height (DBH), canopy cover, terrain slope, and dominant species type. The authors found that fusion of LiDAR and space borne imagery exhibited high associated error when applied to areas with greater than 200 Mg ha_1 of AGB; it therefore is recommended that practitioners always evaluate bias and precision using a field-based sub-sample. Generally speaking, however, more than 150 studies exist that have shown the usefulness of LiDAR sensing when it comes to forest volume or biomass estimation. The question thus becomes one of not necessarily proving the utility of the approach, but rather of constraining the estimation errors. Over and above biomass, it should be noted that there is one variable where a typical, consistent bias is often present, namely tree height. Tree height underestimation is primarily related to sensing characteristics, namely LiDAR point density and at-target beam diameter, dictated by sensor beam divergence (Baltsavias 1999). Current small-footprint discrete return LiDAR sensors have at-target beam diameters of approximately 0.5 m. These small footprints are less likely to interact with the top of the canopy, especially when the survey point density is less than the average crown size, which is the case with crown apexes of many conifer trees (Clark et al. 2004). Another reason for height underestimation is overestimations in the DTM, or digital elevation model (DEM). In old-growth high forests, where a thick understorey is usually present, it is difficult to differentiate between ground and non-ground LiDAR returns. As such, many DTM’s overestimate ground height; this results in underestimated tree height (Maltamo et al. 2004). One example of this is a universal LiDAR canopy height indicator that has been developed by Hopkinson et al. (2006). The authors predicted plot-level canopy height of various vegetation types using the standard deviation of de-trended (topographically normalised) first and second return LiDAR points (Fig. 2.5). The method returned a correlation coefficient of 0.80 when compared to field enumerated heights. However, the authors noted that when the survey was conducted over homogenous vegetation types, the local maximum LiDAR metric returned improved results.

LiDAR sensors are widely regarded as the future of forest inventory, but their application in operational environments remains limited. This is largely due to the costs associated with a LiDAR survey, the documented underestimation of tree heights (Gaveau and Hill 2003; Suarez et al. 2005), and the computational requirements of the LiDAR data processing and analysis. In addition, the application of LiDAR can be impaired by heavily undulating terrain, e. g., in mountainous areas, because of shading effects of the terrain. However, this problem applies to most airborne and satellite-based remote sensing techniques. Even given all of these caveats, LiDAR is bound to play an increasingly important role in forest inventory. It is evident that an increase in markets and vendors will drive costs down, while improved algorithms are increasingly able to address accuracy-precision concerns. These observations are borne out by the importance and investment that large international forest companies, e. g., Sappi, Mondi, Weyerhaeuser, Georgia-Pacific,

etc., are ascribing to LIDAR. This technology arguably will become more adopted, even if forest practitioners will never be able to completely circumvent field measurements for model validation purposes.