Remote Sensing of Environment 140 (2014) 614624

Contents lists available at ScienceDirect

Remote Sensing of Environment

journal homepage: www.elsevier.com/locate/rse

Mapping tropical forest carbon: Calibrating plot estimates to a simple LiDAR metric
Gregory P. Asner , Joseph Mascaro
Department of Global Ecology, Carnegie Institution for Science, 260 Panama Street, Stanford, CA 94305, USA

a r t i c l e

i n f o

a b s t r a c t
Mapping aboveground carbon density (ACD) in tropical forests can enhance large-scale ecological studies and support CO2 emissions monitoring. Light Detection and Ranging (LiDAR) has proven useful for estimating carbon density patterns outside of eld plot inventory networks. However, the accuracy and generality of calibrations between LiDAR-assisted ACD predictions (EACDLiDAR) and estimated ACD based on eld inventory techniques (EACDeld) must be increased in order to make tropical forest carbon mapping more widely available. Using a network of 804 eld inventory plots distributed across a wide range of tropical vegetation types, climates and successional states, we present a general conceptual and technical approach for linking tropical forest EACDeld to LiDAR top-of-canopy height (TCH) using regional-scale inputs of basal area and wood density. With this approach, we show that EACDLiDAR and EACDeld reach nearly 90% agreement at 1-ha resolution for a wide array of tropical vegetation types. We also show that Lorey's Height a common metric used to calibrate LiDAR measurements to biomass is severely awed in open canopy forests that are common to the tropics. Our proposed approach can advance the use of airborne and space-based LiDAR measurements for estimation of tropical forest carbon stocks. 2013 Elsevier Inc. All rights reserved.

Article history: Received 4 April 2013 Received in revised form 10 September 2013 Accepted 26 September 2013 Available online 25 October 2013 Keywords: Aboveground carbon density Biomass Carbon stock estimation Carnegie Airborne Observatory LiDAR Lorey's height National forest inventory Rainforest

1. Introduction Over the past decade, estimation of tropical forest carbon stocks has evolved from an activity based largely on eld inventories (e.g., Malhi et al., 2006), to an effort assisted by airborne and spaceborne remote sensing (Asner et al., 2010; Baccini et al., 2012; Drake et al., 2002; Lefsky et al., 2002; Saatchi et al., 2011). As the more recent approach, remote sensing-based carbon estimates are usually compared to eld inventory-based assessments. However, as noted by Clark and Kellner (2012), virtually all eld-based carbon assessments also represent estimates. Inventory data (i.e., tree diameters, heights, wood densities) are passed into allometric models (e.g., Chave et al., 2005), previously developed by harvesting and weighing trees to determine their biomass (of which ~48% is carbon in tropical forests; Martin & Thomas, 2011), and the summation of each tree's carbon stock estimate within a plot is derived as the eld-estimated aboveground carbon density (EACDeld, units of Mg C ha 1). Uncertainty can be estimated as well (Chave et al., 2004), but true measurement of ACD will ultimately require whole-plot harvests of forest biomass, which are extremely labor intensive and thus rarely carried out (Colgan, Asner, & Swemmer, 2013). In the interim, tree allometry will continue to underlie EACDeld because allometry is one of the most conserved properties in nature (e.g., Niklas, 2006),
Corresponding author. E-mail addresses: gpa@carnegiescience.edu (G.P. Asner), jmascaro@carnegiescience.edu (J. Mascaro). 0034-4257/$ see front matter 2013 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.rse.2013.09.023

and remote-sensing approaches that can accurately predict EACDeld will be critical to carbon stock mapping and monitoring. LiDAR (light detection and ranging) has become a commonly used technology in the effort to remotely predict EACDeld in many forest types (e.g., Ene et al., 2012; Gobakken et al., 2012; McRoberts, Nsset, & Gobakken, 2013). Unlike passive optical techniques, LiDAR uses emitted laser pulses to derive metrics of forest structure in three dimensions (e.g., Omasa, Qiu, Watanuki, Yoshimi, & Akiyama, 2003). Whereas EACDeld assessments have applied tree allometry from the bottom-up to all trees encountered in a plot, most LiDAR-based efforts apply allometric equations at the whole plot or stand level. This approach is not strictly allometry, which refers to scaling at the organism level, but instead it can be thought of as plot-aggregate allometry. Plot-aggregate allometry posits that if forest structure and biomass organization follow consistent scaling patterns, simple plotlevel variables could capture as much information about EACDeld as full eld inventories. Several tropical studies have used LiDAR metrics, such as canopy prole height, to form plot-aggregate allometries to predict EACDeld (e.g., Asner et al., 2010; Drake et al., 2002; Lefsky et al., 2002). Other tropical studies have utilized a daisy-chain of plotaggregate allometries, for example, by linking LiDAR metrics to the plot variable Lorey's height the basal area-weighted average height of all trees and subsequently linking Lorey's height to EACDeld (Harris et al., 2012; Saatchi et al., 2011). However, the trade-offs between direct calibration of LiDAR metrics to EACDeld versus daisychain type approaches have not been communicated in the literature.

G.P. Asner, J. Mascaro / Remote Sensing of Environment 140 (2014) 614624

615

In a previous study, we found that relationships between simple plot-aggregate allometries and airborne LiDAR metrics are not consistent across tropical ecosystems (Asner et al., 2012). To address the additional variation, we devised a general approach to plotaggregate allometry and LiDAR calibration in tropical forests. We argued that because tree-level carbon stocks depend on diameter, wood density and height (Chave et al., 2005), EACDeld should similarly depend on basal area (the cross sectional area of all stems), basal-areaweighted wood density, and canopy height. However, our previous airborne LiDAR calibrations to plot-aggregate allometry for tropical forests were based on a particular 3-D metric termed Mean Canopy Prole Height, or MCH (sensu Lefsky et al., 2002). The MCH metric approximates the vertical centroid of all canopies within a plot, which in turn, is a proxy for the distance stretching from the ground to the top of the main stem of the trees. While this approach yielded consistent relationships with EACDeld when using a single LiDAR sensor (e.g., Asner, Hughes, Varga, Knapp, & Kennedy-Bowdoin, 2009; Asner et al., 2010, 2011, 2012; Mascaro et al., 2011), we recently found that different LiDAR sensors produce inconsistent MCH values based on instrument specications. In particular, laser beam divergence and power, and the sensitivity of the LiDAR receiver, causes differences among LiDAR measurements of the vertical distribution of the canopy tissues, upon which the MCH and similar metrics are calculated (see also Nsset, 2009). This issue affects all LiDAR metrics that are sensitive to the vertical prole of the vegetation, such as RH50, RH80, and many others (Ni-Meister et al., 2010). Fortunately however, our tests also indicate that top-of-canopy height (TCH) the distance from ground to the topmost point on the canopy is a much more consistent index among different modern LiDAR sensors, thereby offering a way to circumvent the highly variable, sensor-specic performance of vertical prole metrics. However, the generality of TCH-based approach to plot-aggregate carbon stock estimation has not been broadly examined for tropical vegetation. Another recent development is that LiDAR-based estimates of tropical forest ACD (EACDLiDAR) approach EACDeld when eld plots reach one hectare in size (Asner et al., 2010; Mascaro, Detto, Asner, & Muller-Landau, 2011; Zolkos, Goetz, & Dubayah, 2013). Several factors contribute to this: (1) errors caused by spatial misalignment of plots and LiDAR data are diminished with larger plots (Asner et al., 2009), (2) integrating measurements over larger plots provides a more representative average (Zolkos et al., 2013), and (3) disagreement in protocol between LiDAR and eld observations namely the effects of bisecting tree crowns in LiDAR data versus calling a tree in or out of the plot in eld data decreases to a manageable level (Mascaro, Detto, Asner, & Muller-Landau, 2011). Although such patterns may be consistent in sign across tropical vegetation types, the magnitude of the plot-size effect on error and particularly the improvements detected at 1 ha in size reect not plot size per se, but the size of the plot relative to the average crown size, which provides the bulk of the LiDAR return signal (e.g., contrast with boreal forest, where crowns are much smaller; Naesset et al., 2011). Because 1-ha plots entail a very large amount of labor to measure for biomass inventory (indeed, they are the upper limit in the size of most tropical forest inventory plots; Malhi et al., 2006), they represent a costly and labor-intensive trade-off with the number of plots that must be used for validation. This issue is central to determining the generality of LiDAR-based approaches for tropical forest carbon mapping, and thus additional research that might reduce the need for exhaustive plot-based calibration is warranted. Here, we use a network of 804 tropical forest inventory plots to assess the ability of a very simple LiDAR metric top of canopy height or TCH to predict EACDeld for a wide range of tropical vegetation types and ecological settings. We had three specic objectives: First, we assessed the effectiveness of calibrating LiDAR TCH to EACDeld using both regionally- and generically-constrained estimates of plotaggregate allometry. Second, we tested whether Lorey's Height can be

used to enhance EACDeld predictions as is often practiced in LiDARassisted carbon mapping studies. Finally, we examined the effect of plot size on LiDAR predictions of EACDeld, with specic attention to the use of fewer 1-ha validation plots to save time and reduce project cost. 2. Methods 2.1. Field plots Greatly expanding on Asner, Mascaro, et al. (2012), our eld plot network includes plots in 14 distinct tropical ecoregions, with more than a half million trees measured, and across an enormous range of forest types, oristic composition, disturbance regimes and successional states. The network presented here has been updated to 804 plots in Colombia, Hawaii, Madagascar, Panama and Peru (50 of which are reserved for validation; Table 1). The plots are positioned from sea level in the Pacic, to more than 3500 m altitude in the Andes, and across a wide range of climate conditions (mean annual precipitation range: 18011,000 mm year 1; mean annual temperature range: 6 27 C). The types of vegetation included in the database range from dense, humid lowland forest to dry spiny woodlands, and from mature, closed canopy forests to open woodland-savanna physiognomies. A complete description of the plots can be found in the references provided in Table 1, and a plot-level listing of the data used for LiDARto-eld calibration is provided in the Online Electronic Supplement 1. EACDeld was assessed for each plot following a consistent protocol detailed by Asner, Mascaro, et al. (2012). The minimum diameter class for all stems was standardized to 5 cm. In Peru and Colombia, stems 510 cm in size were estimated by a subplot one-eighth the area of the main plot. In Madagascar, a similar scheme was used for stems 5 10 cm and 1020 cm in size (Asner, Clark, et al., 2012). Plot size and conguration differed across projects (Table 1), variables we consider in the Uncertainty section. However, tree-level allometries used to estimate biomass were consistent across all plots, and are based on the most local information rst, followed by generalized equations from Chave et al. (2005). Tree height for eld inventory data was measured for the three largest trees in each plot for all project areas, and for all trees in more recent projects, either using laser ranging hypsometers (Impulse 2000, Laser Technology, Durham NC), or clinometers. For remaining trees without a height measurement, height-diameter models were used to estimate individual tree height (sensu Chave et al., 2005). In all cases, we used height-diameter allometry at the species or regional level rather than defaulting to allometric equations that exclude a height parameter unless such equations were species-specic (thereby subsuming species-level diameter-to-height variation in the coefcients, discussed below). The inclusion of height in allometry has been found to be essential to preventing overestimation of EACDeld in nearly all tropical regions (Feldpausch et al., 2012). 2.2. LiDAR data The LiDAR data were collected using the Carnegie Airborne Observatory (CAO) Alpha (Asner et al., 2007) or AToMS (Asner et al., 2012) sensor packages, with data collection and analysis methods applied consistently across sites. Both the Alpha and AToMS scanning LiDAR sensors are full waveform, but the work presented here relied only on the discrete return data of up to four returns per pulse in order to make the results applicable to a much wider range of LiDARs currently in operation throughout the world. Over all eld plots listed in Table 1, the CAO LiDARs were operated at 2000 m above ground level (a.g.l.) with a 30o eld of view, pulse repetition frequency of 50 kHz, and a ground speed of 110 knots. Both LiDARs have a laser beam divergence set to 0.56 mrad (1/e), providing 1.12 m laser spot spacing from 2000 m a.g.l. However, 50% overlap between adjacent ight lines resulted in two laser

616

G.P. Asner, J. Mascaro / Remote Sensing of Environment 140 (2014) 614624

Table 1 Summary statistics for the LiDAR plot network of tropical forest sites in Colombia, Hawaii, Madagascar, Per, and Panam. The data span non-planted intact mature forests, secondary and degraded forests, logged forests, and open woodlands and savannas as described in the source references. Region Plot number and size (ha) Plot shape Taxonomic resolution Mean (SD) LiDAR TCH (m) Hawaii Wet Moist Dry Madagascar Northern Wet Southern Moist Southern Dry Per Northern Moist Northern Moist Southern Moist Southern Submon. Wet Southern Mont. Wet Panam Moist Moist Wet Dry Colombia Moist Ranges Totals 19 (0.28) 123 (0.28) 34 (0.28) 26 (0.28) 19 (0.28) 36 (0.13) 32 (0.28) 10 (1.0) 156 (0.28) 10 (0.28) 15 (0.28) 228 (0.10.36) 40 (0.81.0) 13 (0.41.0) 33 (0.1) 10 (0.28) 0.11.0 ha 804 plots Circular Circular Circular Circular Circular Circular Circular Square Circular Circular Circular Rectangular Rectangular Rectangular Rectangular Circular Species Species Species Species Species Species Genus Genus Genus Genus Genus Species Species Species Species Genus 13.2 (5.7) 10.5 (7.1) 3.4 (3.4) 20.6 (7.0) 13.0 (6.9) 3.4 (1.2) 16.8 (7.5) 27.2 (1.8) 21.1 (9.0) 22.8 (4.0) 17.4 (4.4) 18.4 (8.1) 21.1 (2.4) 21.8 (2.5) 5.3 (2.7) 19.8 (8.3) Basal area (m2 ha1) Wood density (g cm3) EACDeld (Mg ha1) 119.2 (52.7) Asner et al. (2011) 105.5 (82.2) Asner et al. (2011) 33.5 (30.6) Asner et al. (2011) 128.1 (66.7) Asner, Knapp, et al. (2012) 68.8 (43.8) Asner, Knapp, et al. (2012) 13.3 (7.3) Asner, Knapp, et al. (2012) 70.9 (51.2) 160.1 (9.7) 103.2 (67.8) 69.9 (21.5) Asner et al. unpublished Asner et al. unpublished Asner et al. (2010) Asner et al. unpublished Reference

33.7 (16.1) 0.62 (0.06) 28.5 (18.2) 0.64 (0.06) 11.9 (9.4) 0.69 (0.08) 34.9 (14.6) 0.57 (0.05) 27.1 (14.9) 0.57 (0.06) 15.3 (8.2) 0.32 (0.04) 20.1 (10.4) 30.4 (2.1) 23.8 (11.4) 26.7 (5.3) 26.8 (7.0) 21.9 (10.8) 20.9 (5.0) 29.4 (3.8) 5.9 (4.5) 22.6 (9.5) 0.58 (0.07) 0.60 (0.02) 0.53 (0.09) 0.50 (0.03) 0.52 (0.04) 0.53 (0.04) 0.56 (0.03) 0.54 (0.03) 0.52 (0.03) 0.61 (0.03)

64.2 (23.9) Asner et al. unpublished 74.7 (46.1) 84.5 (18.1) 91.1 (16.3) 12.1 (10.9) Mascaro, Asner, et al. (2011) CTFS/SIGEO CTFS/SIGEO Hall et al. unpublished

95.5 (48.3) Asner, Clark, et al. (2012)

Validation plots. TCH = top-of-canopy height. EACDeld = eld-estimated aboveground carbon density.

shots per 1.25 m2 (1.12 m 1.12 m) for all plots in this study. Despite similar data collection settings, the two CAO LiDARs differ in several key specications including laser pulse width, peak diode power, and receiver sensitivity (Table A1). In particular, the AToMS LiDAR proved much more sensitive to the internal 3-D architecture of any given forest canopy, independent of beam divergence, spot spacing or scan frequency. Although Asner, Mascaro, et al. (2012) previously compared regions using LiDAR-derived mean canopy prole height index (MCH), we used top-of-canopy height (TCH) in this study. We did so for two specic reasons: (1) while TCH is available for most airborne and spaceborne LiDARs, MCH is not available in some instances, and (2) MCH is less generalized across sensor types than TCH due to canopy penetration differences (Nsset, 2009). TCH was determined by constructing digital surface and terrain models (Asner et al., 2009), and subtracting them to determine vegetation height at 1.12 m resolution. Laser ranges from the LiDAR were combined with embedded high resolution Global Positioning System-Inertial Measurement Unit (GPS-IMU) data to determine the 3-D locations of laser returns, producing a cloud of LiDAR data. The LiDAR data cloud consists of a very large number of georeferenced point elevation estimates (cm), where elevation is relative to a reference ellipsoid. The LiDAR data points were processed to identify which laser pulses penetrated the canopy volume and reached the ground surface. We used these points to interpolate a raster digital terrain model (DTM) for the ground surface. This was achieved using a 10 m x 10 m kernel passed over each ight block; the lowest elevation estimate in each kernel was assumed to be ground. Subsequent points were evaluated by tting a horizontal plane to each of the ground seed points. If the closest unclassied point was b 5.5o and b 1.5 m higher in elevation, it was classied as ground. This process was repeated until all points within the ight coverage were evaluated. The digital surface model (DSM) was based on interpolations of all rstreturn points. Measurement of the vertical difference between the DTM

and DSM yielded a model of canopy height above ground (digital canopy model, DCM). In each forest inventory plot, the average of all 1.12 m DCM pixels was used to estimate mean plot TCH. For plots of 0.1 to 1.0 ha in size, this equates to 8938,930 height measurements per eld plot. 2.3. Plot-aggregate allometry Our approach to plot-aggregate allometry for LiDAR calibration follows general tree allometric theory (Chave et al., 2005), along with several key observations on the application of tree allometry. An idealized tree allometry for a given species is: AGB aD
b

where AGB is aboveground tree biomass (kg, of which ~48% is carbon; Martin & Thomas, 2011), D is stem diameter (cm), and a and b are parameters to be estimated empirically (b is the allometric scaling coefcient; Niklas, 2006). The form of this power function has been central to tree allometric theory for nearly a century (West, Brown, & Enquist, 1999). However, the high biological diversity of tropical forests, variation among species (such as wood density, tree form, and the relative height attained per unit of diameter increment) leads to weaker performance of the power function for tropical tree allometry (e.g., Brown, 1997; Chave et al., 2005). Even particular tropical tree species are unlikely to follow an idealized power function because their biomass begins to diverge from the modeled prediction of diameter, as they get older, undergo stem rot or experience crown breakage (e.g., MullerLandau et al., 2006). This does not signal a breakdown in theory, but a decoupling of traditional allometric components (i.e., overall diameter) from actual allometric components (i.e., cross sectional area of functional xylem cells within the tree). Given these realities, a sigmoidal function or higher-order polynomial is often required.

G.P. Asner, J. Mascaro / Remote Sensing of Environment 140 (2014) 614624

617

In light of these limitations, current generalized tree allometric equations are tailored to multiple species, and are more robust to declining biomass relative to diameter in older trees (Chave et al., 2005): AGB aD H
b1 b2 b3

low-lying forest, often has the same Lorey's height but much lower EACDeld than a more densely stocked forest. Drawing from the form of Chave et al. (2005; model 2), Asner, Mascaro, et al. (2012) proposed the following general plot-aggregate allometry: EACDfield aTCH BA BA
b1 b2 b3

where D is stem diameter (cm), H is canopy height (m), is wood specic gravity (wood density, g cm 3). Note that the driving variables differ in their inuence on the nal output by way of separate scaling coefcients. Note also that, if Db1Hb2b3 is reduced to D2H, the equation collapses to the volume of a cylinder, and variation in a represents tree taper. Chave et al. (2005) found evidence for three different generalized equations based on persistent differences from dry, moist, and wet tropical forests, and these can be thought of as reecting predictable variation in tree taper (indeed, when Db1Hb2b3 is held to D2H, the parameter a is lower for dry and wet forests, where trees are thought to be more conical than in moist forests; Chave et al., 2005). LiDAR-based efforts to predict EACDeld have utilized many different sensor congurations as well as different indices, either alone or in concert (see reviews by Nsset, 2009; Zolkos et al., 2013). With fullwaveform or discrete-return LiDARs in the tropics, mean canopy prole height or various canopy prole quantiles have often been used (Asner, Mascaro, et al., 2012; Asner et al., 2009, 2010, 2011; Baccini et al., 2012; Lefsky et al., 2002, 2005; Mascaro, Asner, et al., 2011). Among the simplest approaches that can be envisioned is the prediction of EACDeld by LiDAR TCH (see LiDAR data above) using a power function (e.g., Drake et al., 2003): EACDfield aTCH
b

where TCH is LiDAR-derived top-of-canopy height (m; replacing MCH as discussed above), BA is plot-averaged basal area (m2 ha 1), BA is basal-area weighted wood density. As the Chave et al. (2005) generalized allometry allows for tree species-level variation, the Asner et al. plot-aggregate allometry allows for regional variation in wood density and stocking to inuence the tted model output rather than being subsumed into the model coefcients. Relating EACDeld to both LiDAR TCH and eld inventory data (BA, BA) appears circular at rst glance: tree diameter data enter into both the independent (through BA) and dependent variables (via allometric predictions of AGB for each tree that are summed to determine EACDeld). However, EACDeld incorporates the within-plot covariance of tree diameter, height, and wood densitythis covariance represents the potential benet of exhaustive eld inventory relative to three simple, whole-plot variables. Thus, the tting of Eq. (6) is a starting point for our method that is intended solely to assess the value added of complete inventory versus plot-aggregate variables: a strong t in Eq. (6) would suggest that exhaustive inventory provides a very low benet relative to simpler, plot-level inputs, while a poor t would suggest that full inventory is essential. Under Model application below, we demonstrate how BA and BA can be decomposed regionally via LiDAR TCH, eliminating the need for exhaustive inventory. 2.4. Model tting Models based on the power function (Eqs. (3), (5)) were t using nonlinear Maximum Likelihood Analysis, and by incorporating a third model parameter (k), which produces a geometric error term: yi a xi i xi   2 i eN 0; We used this approach to account for heteroskedasticity common to previously published LiDAR-to-EACDeld relationships. The technique is analogous to tting a linear model to log-transformed x and y data, thereby avoiding the need for log-transformation and backtransformation (Baskerville, 1972). With more complex models, such as Eqs. (2) and (6), the number of free parameters increases the possibility of arriving at local optima with maximum likelihood, and thus we relied on traditional linear models t to log-transformed data. In the case of the general equation (Eq. (6)), we followed Asner, Mascaro, et al. (2012) and t a model in the form: ln EACDi ln a b1 ln TCH i b2 ln BAi b3 ln BAi i   2 i eN 0; We then back-transformed the nal model and multiplied by a correction factor (CF) to account for the back-transformation of the regression error (Baskerville, 1972; Goldberger, 1968); the correction factor is given by exp(MSE/2), where MSE is the mean square error of the regression model. The tting of the model in Eq. (8) violated two assumptions: that the input variables were uncorrelated, and that the plots (some of which were adjacent) were spatially independent. In the rst case, the inclusion of correlated input variables is often employed in allometric models where the goal is not to determine the 8
b k

Eq. (3) is mathematically identical to Eq. (1), but as discussed, it is based on plot-aggregate allometry rather than true allometry, because it reects whole-plot properties of forest structure in aggregate rather than the properties of each particular organism (e.g., trees, lianas, etc.). Just as the power function in Eq. (1) fails to explain general tropical tree allometry, Eq. (3) is also unlikely to be consistent across tropical forests, which may differ in plot-average wood density (Baker et al., 2004; Stegen, Swenson, Valencia, Enquist, & Thompson, 2009; ter Steege et al., 2006), or in stocking levels (e.g., basal area or the amount of stem wood cross sectional area) at a given canopy height (Asner, Mascaro, et al., 2012). In the latter case, Asner et al. (2011) found that Hawaiian forests had almost twice the basal area at a given canopy height than mainland Neotropical forests (Asner et al., 2010). Because of its widespread use in LiDAR remote sensing of biomass (e.g., Saatchi et al., 2011), we also considered plot-aggregate allometry based on Lorey's height (LOR), which is the basal area-weighted height of all trees in a plot: LOR BAi Hi = BAi 4

where BAi is the basal area and Hi is the height of the ith tree in the plot. Saatchi et al. (2011), using eld data alone, showed that EACDeld was related to Lorey's height according to: EACDfield aLOR
b

However, the estimated parameters a and b differed substantially across tropical forests. As a result, EACDeld was predicted to be lower for secondary and degraded forests at the same Lorey's height. These differences might be explained by differences in other aggregate properties such as plot-average wood density or stocking. Stocking may be particularly important because the Lorey's height index specically removes variation caused by stocking differences, so a plot with a handful of large trees, such as might characterize logged or

618

G.P. Asner, J. Mascaro / Remote Sensing of Environment 140 (2014) 614624

most parsimonious model but to estimate ACD with greatest condence (e.g., consider the inclusion of both tree diameter and tree height in many allometric model tting exercises; Chave et al., 2005). In the second case, we carefully evaluated the inclusion of adjacent plots (by excluding them and re-tting the model in Eq. (8)) and found that they had a minimal effect on our conclusions (see Uncertainty, below). Interrelationships between LiDAR-derived TCH, basal area (BA), and basal-area-weighted wood density (BA) were assessed by ordinary least squares (OLS) regression, applying a Bonferroni correction to values within a region. 2.5. Model application The general equation (Eq. (6)) only achieves its true utility when it is applied without eld inventory data. To facilitate this, we modeled variation in BA and BA according to TCH for each region using regression by ordinary least squares. This is analogous to applying Chave et al. (2005) tree-level allometry using predicted height inputs from diameter, or literature-based wood density estimates. To apply the general model in a new region, exhaustive inventory would not be needed. Instead, spatially-explicit point-based estimates of BA (by relascope or prism method) and BA (by recording dominant species) would be collected within the coverage of a LiDAR TCH dataset. By regressing BA and BA onto TCH and substituting these regressions into the general model, regionally-tailored predictions could be generated. In the simplest form, Asner, Mascaro, et al. (2012) proposed a single stocking coefcient (the ratio of BA to TCH) and a single wood density constant for each broad tropical region based on the literature, including many new maps of wood density emerging today (ter Steege et al., 2006). Here, we used simple linear models to capture the variation in BA and BA for each region. In cases in which the resulting equation had a positive y intercept (i.e., predicted positive BA at a TCH of zero), we t new models forced through the origin. This assumption is ecologically based (e.g., that at a TCH of zero, BA must not be greater than zero) and was applied because several regions included few plots of low TCH, risking unstable or spurious model outputs. However, we did allow resulting equations with negative y intercepts to accommodate the effects of our minimum stem size of 5 cm diameter at breast height (i.e., consider low tropical vegetation which commonly has positive LiDAR TCH, but for which no/few stems meet the minimum size class). Negative predicted values of BA were set to zero. 2.6. Uncertainty For two study areas in the Peru Amazon and Central Panama, we set aside 50 inventory plots totaling 48.4 ha of forest for validation of the general Eq. (6). In the Allpahuayo region of northern Peru, we established ten separate 1-ha plots specically targeted to the maximum canopy height in the region (where uncertainty in aboveground biomass was likely to be greatest). For each these square plots, we subdivided them into eight sections of equal area, allowing detailed analyses of the effects of plot size on uncertainty. In Central Panama, we partitioned a previously existing 38.4-ha plot into 40 units (0.81.0ha each) consisting largely of old-growth and old secondary forests with greater biomass (Wright et al., 2011), where again we anticipated high uncertainty. Previous work in this area showed low spatial autocorrelation among adjacent plots (Mascaro, Detto, Asner, & Muller-Landau, 2011). This permanent plot includes trees mapped to 10 cm spatial resolution, similarly allowing an analysis of spatial resolution on uncertainty. We utilized quantile regression to improve our understanding of the effects of plot size on scatter and heteroskedasticity (Cade & Noon, 2003). To consider the possible effects of spatial autocorrelation due to the inclusion of adjacent plots, we carried out our analysis with and without the 136 plots taken from a single large plot in Panama. We determined that the effects on the general model results (see Fig. 5b) were minimal: R2 declined less than 1% and RMSE increased by 1.6 Mg Cha 1. Spatial

autocorrelation of carbon stocks vis--vis LiDAR prediction errors are considered in greater detail in Mascaro, Detto, Asner, and MullerLandau, (2011). 3. Results 3.1. Contrasting plot-aggregate allometries LiDAR TCH alone failed to explain cross-regional variation in tropical forest carbon estimates derived from detailed eld inventory measurements (Fig. 1a). In particular, TCH explained just 56% of the variation in EACDeld using the power-law approach (Eq. (3); Table 2; R2 in all cases determined as % variation in the observed data explained by the prediction, tted model R2 values were adjusted for degrees of freedom), with major differences observed among contrasting forest physiognomies from the western Amazon (blue symbols) to the Hawaiian Islands (red symbols; Fig. 1a). Using a simple power-law model between EACDeld and Lorey's height (LOR; Eq. (5)), we found that only 39.3% of the EACDeld variation was explained (Table 2, Fig. 1b). The model demonstrated substantial lack-of-t by region, with predicted EACDeld in forests of all types in Hawaii and Madagascar departing from those in the Neotropics. Similarly, EACDeld in a large number of plots in the Peruvian Amazon and Panama were over-predicted by an average of 100 Mg C ha 1. Lorey's height was only broadly related to LiDAR TCH across all regions, with very important differences discussed later (Fig. 2; R2 = 0.84, RMSE = 2.55 m). Our plot-aggregate approach including LiDAR TCH, BA and BA explained 92.3% of the variation among 754 plots (Table 2), and was consistent for all regions (Fig. 1c; RMSE = 17.1 Mg C ha 1). We emphasize that the analysis represented in Fig. 1c quantitatively establishes that plot-aggregate properties of tropical forests capture

(a)

(b)

(c)

Fig. 1. (a) Estimated values of aboveground carbon density using LiDAR top-of-canopy height (EACDLiDAR) compared to plot-estimated aboveground carbon density using exhaustive eld measurements and allometric equations (EACDeld). (b) Similar to panel (a) but using Lorey's height instead of top-of-canopy height. (c) A model incorporating plot-aggregate variation in LiDAR TCH, basal area, and wood density compared to EACDeld. In all panels, the solid one-to-one line is provided for reference.

G.P. Asner, J. Mascaro / Remote Sensing of Environment 140 (2014) 614624 Table 2 Plot-aggregate allometry equations and corresponding model parameter estimates. Input data are 754 eld plots (all but the 50 validation plots in the network) across ve tropical countries and many vegetation types. EACDeld is eld-estimated aboveground carbon density (Mg ha1), TCH is LiDAR top-of-canopy height, BA is basal area (i.e., crosssectional area of all stems, m2 ha 1), BA is basal area-weighted wood density (g cm3). Estimated values of a reect a correction factor for back-transformation, and adjusted R2 values are the percent variation in observed EACDeld (on the original scale) by the prediction. All estimated model parameters are signicant at P b 0.0001. Predicted versus observed EACDeld values are shown in Fig. 1. Equation Model parameter estimates a EACDeld = aTCHb1 EACDeld = aLORb1 EACDeld = aTCHb1BAb2b3 BA b1 b2 b3 RMSE R2 df

619

contributing regions was minimal, although Neotropical wet forests were slightly and consistently over-predicted by the general model (Fig. 6). 3.3. Uncertainty For 48.4 ha of forest set aside for validation in Central Panama and the Peruvian Amazon (n = 50 plots of 0.81.0 ha per plot), the general approach yielded an adj-R2 = 0.86 and RMSE = 13.2 Mg C ha 1 (Fig. 7a). The site-level RMSE relative to mean EACDeld was slightly greater in Central Panama (11.3%) than in the Amazon plots (10.8%). Given that all 40 validation plots in Central Panama were partitioned from a single large plot, our results should be interpreted with caution. However, the slightly greater RMSE in Central Panama suggests that the inuence of spatial autocorrelation on the 40 adjacent Panama plots used for validation was minimal. We also considered how plot size affects the relationship between EACDLiDAR and EACDeld for the validation plots (Fig. 7ad; Online Electronic Supplement 2). While plot size imparted little to no lack-oft on the relationship (which declined from 2.1 Mg C ha 1 to b 1 Mg C ha 1), the scatter and heteroskedasticity of the data were strongly coupled to plot size: RMSE declined from 40.1 to 13.2, while R2 increased from 0.46 to 0.86. Thus, plots of less than 0.12 ha (Fig. 7b) were prone to greater uncertainties. In contrast, plots reaching 1.0 ha in size (Fig. 7a) were closely aligned with LiDAR-based predictions of EACDeld. The validation data displayed a steady decline in the uncertainty between EACDLiDAR and EACDeld as plot size increased to 1 ha (Fig. 8). At 1-ha scale, errors between LiDAR- and eld-based estimates of carbon stock were in the 1014% range. Lack-of-t between the theoretical curve and the empirical data was considered in detail by Mascaro, Detto, Asner, and Muller-Landau, (2011). 4. Discussion LiDAR is highly sensitive to the three-dimensional structure of vegetation, and therefore primary LiDAR data can be thought of as resembling the primary eld inventory measurements used to estimate aboveground carbon density (EACDeld) in one critical respect: both LiDAR and eld measurements contain information on plant architecture that scales with plant biomass according to generalized allometric theory. EACDeld within a given tropical forest region has been shown to scale with simple LiDAR metrics according to a power law, suggesting that forest EACDeld scales similarly in aggregate (i.e., within a whole plot) to the scaling of individual tree biomass (Niklas, 2006). Based on this underlying concept, we explored different approaches to predicting EACDeld throughout an 804-plot network spanning many tropical vegetation types (50 of which were set aside for validation). We found that LiDAR estimates of top-of-canopy height (TCH) alone explain just a portion of EACDeld across tropical regions and conditions (Fig. 1a). However, we also found that the remaining variation was almost entirely explained by including plot-aggregate estimates of basal area and basal-area weighted wood density (Fig. 1c). This suggests that the application of the general plot-aggregate allometry can be accomplished with rapid surveys of basal area (such as by relascope or prism) and by cataloging of dominant species for wood density estimation within the LiDAR swath. We demonstrate this in principle in Fig. 5b, where we show only a very small decline in predictive power relative to exhaustive inventory (Fig. 5a). Basal area is well correlated with TCH in all of the tropical forests we have surveyed (Fig. 3). What varies among regions is the slope of the relationship between TCH and basal area, a factor we have termed the stocking coefcient (SC). However, we also nd the SC to be highly conserved within regions and often across wide-ranging environmental conditions. For example, SC values vary from just 1.13 to 1.34 among all lowland Neotropical forest plots spread across Panama, Colombia, and the Peruvian Amazon (Fig. 3,

6.8500 0.9520 42.2160 0.557 752 0.1507 2.1991 49.4028 0.393 752 3.8358 0.2807 0.9721 1.3763 17.1281 0.923 750

nearly all the EACDeld variation estimated by exhaustively detailed plot inventories. This is critically important for setting up the remaining analyses to develop regional and general calibrations of LiDAR TCH to plot-aggregate EACD in the tropics. 3.2. Applied plot-aggregate allometry Analysis of the eld plots by region indicated linear scaling of basal area with LiDAR TCH (Fig. 3). Critically, note that the estimated slopes of the regression models varied, particularly when comparing disparate tropical vegetation types and regions such as Hawaii, Madagascar and the Neotropics. This strongly suggests that regional LiDAR-to-EACDeld calibrations must take into account variation in the relationship between TCH and BA, a term referred to as the stocking coefcient or SC (Asner, Mascaro, et al., 2012). In contrast, most regions exhibited little to no relationship between basal area-weighted wood density (BA) and LiDAR TCH (Fig. 4). Therefore, an average BA can, in principle, be used to incorporate wood density in the regional application of plot-aggregate allometry to LiDAR measurements. Fig. 5 reports the ability of LiDAR to predict EACDeld when including regionally-derived estimates of BA and BA from Figs. 3 to 4. In Fig. 5a, we provide results of separate calibration equations for each of the 14 ecoregions (Table 3); combined, these yielded an ensemble R2 = 0.85 and RMSE = 24.7 Mg C ha 1. Compared to the combined performance of 14 separate ecoregional plot-aggregate allometries in Fig. 5a, a single general approach incorporating all regions into a single model gave an R2 = 0.83 and a RMSE = 26.3 Mg C ha 1 (Fig. 5b). Lack-of-t in

Fig. 2. Relationship between Lorey's heighta eld metric of height weighted by basal area, and LiDAR top-of-canopy height. Colors correspond to the ecoregions listed in Fig. 1, and the solid one-to-one line is provided for reference.

620

G.P. Asner, J. Mascaro / Remote Sensing of Environment 140 (2014) 614624

Fig. 3. Relationships between basal area (BA) and LiDAR top-of-canopy height (TCH) for 14 ecoregions included in a 804-plot network. Equations, regression lines and p-values are provided in each panel. Bias reects mean errors, or lack-of-t between the predicted and observed data.

Table 3). Similarly, from the lowlands to treeline, and from very wet to near desert forest conditions in Hawaii, the SC values range from just 2.5 to 3.0. Finally, among regions, we nd generally weak relationships between TCH and basal area-weighted wood density, although there is a slightly positive relationship between these variables in Peru (Fig. 4). This is extremely useful information, suggesting that the added value of exhaustive eld inventory for EACDeld is small. We believe that our plot network covers a wide range of tropical vegetation conditions. For example, we have replicate plots in extremely dry, spiny forests of Madagascar and in wet montane cloud forests of the Andean Amazon. Forest composition ranges from near monotypic stands of endemic species evolved on the most isolated islands on the planet Hawaii to high-diversity lowland Mesoamerican forests in Central Panama. Climate, soils, topography, as well as disturbance regimes from ancient forests to heavily logged forest to nearly fully deforested cattle pastures are well represented in our LiDAR-eld plot network. Nonetheless, we recognize that we are lacking some important and unique tropical canopies such as the Dipterocarpaceae forests of IndoMalaysia, well known for their steep diameter-to-height relationships (Feldpausch et al., 2012), although 60-m tall forests in the Osa Peninsula in Costa Rica conform to our general model predictions (P. Taylor, Univ. of Colorado, unpub. data). More importantly, the general concepts and technical approach we have presented should be tested in other ecosystems by acquiring the airborne LiDAR models of top-of-canopy height and combining them with spatially explicit estimates of basal area and wood density (e.g., Chave et al., 2009; ter Steege et al., 2006).

Only through continued testing can the plot-aggregate based approach be improved and evaluated in new environments, both tropical and non-tropical. In comparison to the general model based on TCH (Eq. (6)), we found that Lorey's height is quite ineffective as an intermediate link between LiDAR and EACDeld. Although LiDAR TCH and Lorey's height are broadly related to one another (Fig. 2) as has been found in previous studies (Goerndt, Monleon, & Temesgen, 2010; Gonzalez et al., 2010; Lefsky, 2010; Magnussen, Naesset, Gobakken, & Frazer, 2012), Lorey's height explained no variation in EACDeld, BA, or BA that was not already directly explained by LiDAR TCH alone (Fig. 1, Table 2). Furthermore, the plots departing most strongly from the Lorey's height relationship contained large, scattered trees indicative of degraded and secondary forests, and/or site conditions such as inundation that create naturally open-canopy forest (Fig. 9). These outliers are not bad data, rather they indicate a critically important distinction between LiDAR TCH and Lorey's height: while TCH is an integrated measurement of the top-of-canopy height across an entire plot, Lorey's height is an indexed height based on the largest trees. Thus, it is possible for two plots to have the same Lorey's height but quite different BA and EACDeld, particularly in disturbed landscapes. This likely explains the different EACDeld-to-Lorey's height models that are required to accommodate structural variation in secondary, open-canopy, and degraded forests (Saatchi et al., 2011). It is also important to recognize that both the calibration and validation accuracy of the general modeling approach may depend on

G.P. Asner, J. Mascaro / Remote Sensing of Environment 140 (2014) 614624

621

Fig. 4. Relationships between basal area-weighted wood density (BA) and LiDAR top-of-canopy height (TCH) for 14 ecoregions included in an 804-plot network. Equations, regression lines and p-values are provided in each panel. Bias reects mean errors, or lack-of-t between the predicted and observed data.

the number and size of the eld plots. Our results suggest that larger plots of high quality (e.g., spatially mapped trees or partitioned subplots) may lead to a more robust understanding of calibration uncertainties. Even with perfect GPS accuracy, eld inventory protocols assign trees to particular x and y coordinates in space. Yet the same trees are, in reality, three-dimensional objects in the LiDAR data. In the LiDAR calibration phase, use of small plots relative to typical crown sizes will always lead to inated scatter and thus increased RMSE between LiDAR TCH (or any metric) and EACDeld (Fig. 7). As a result, a large number of small plots will be needed to calibrate. In contrast, fewer large plots (e.g., N 0.5 ha) reduce scatter and RMSE to produce a better calibration, as long as the large plots cover a wide range of carbon stock levels. The reasons for this are straightforward: (1) the accuracy

Table 3 Fitted models used to apply the regional (LiDAR top-of-canopy height or TCH; e.g., as in Eq. (3)). Each regional TCH equation predicts EACDeld based on calibration by exhaustive eld inventories. For the general Eq. (6), BA and BA are estimated based on regional adjustments that require eld assessments of BA and BA within the LiDAR data, but in principle without the need for exhaustive eld inventories. Fig. 6 shows the resulting output of the regional versus universal approaches. Region Regional approach Universal stocking (estimates ACD) adjustment (estimates BA) 8.0043*TCH1.0988 12.0759*TCH0.9269 14.7846*TCH0.7421 2.169*TCH1.3411 5.0504*TCH1.0137 3.6126*TCH1.0553 2.4949*TCH 2.5795*TCH 2.9979*TCH 1.6822*TCH 1.9255*TCH 4.3686*TCH Universal wood density adjustment (estimates BA)

(a)

(b)

Fig. 5. Performance of (a) separate ecoregional models and (b) the general plot-aggregate allometry approach to predicting plot-estimated aboveground carbon density (EACDeld). In each panel, the solid one-to-one line is provided for reference.

Hawaii Wet Moist Dry Madagascar Northern Wet Southern Moist Southern Dry Peru Northern Moist Southern Moist Southern Submon. Wet Southern Mont. Wet Panama Moist Wet Dry Colombia Moist

0.0050*TCH + 0.6891 0.0024*TCH + 0.6695 0.0002*TCH + 0.6917 0.0017*TCH + 0.5310 0.0007*TCH + 0.5838 0.0106*TCH + 0.2831

0.2601*TCH1.9337 0.4356*TCH1.7551 0.6481*TCH1.4888

1.2666*TCH1.1340 1.1623*TCH0.7048 1.1636*TCH

0.0044*TCH + 0.5056 0.0057*TCH + 0.4113 0.0067*TCH + 0.3487

1.6917*TCH1.2811

1.5056*TCH

0.0037*TCH + 0.5881

0.7373*TCH1.5560 1.9183*TCH1.2511 0.9385*TCH1.4858 2.1011*TCH1.2682

1.2343*TCH0.8673 1.3424*TCH 1.4967*TCH1.9618 1.1340*TCH

0.0016*TCH + 0.5556 0.0018*TCH + 0.5010 0.0006*TCH + 0.5201 0.0011*TCH + 0.6357

622

G.P. Asner, J. Mascaro / Remote Sensing of Environment 140 (2014) 614624

Fig. 6. Performance of ecoregional models (solid lines) versus the general plot-aggregate approach (dashed lines) to predicting plot-estimated aboveground carbon density (EACDeld) from LiDAR top-of-canopy height (TCH). Fig. 5 contains a global analysis. Bias reects mean errors, or lack-of-t between the predicted and observed data.

of plot-aggregate allometry appears to increase when averaging over more vegetation in larger plots; and (2) larger plots minimize the edge effects related to uncertainty in including or excluding a tree in the

eld survey. Optimizing plot size is likely to be dependent not on plot size per se, but on the ratio of typical crown sizes to the plot size. Thus, smaller plot sizes may be more tractable in boreal forests for example

(b) 0.08 - 0.12 ha (a) 0.80 - 1.00 ha

(c) 0.24 - 0.25 ha

(d) 0.48 - 0.50 ha

Fig. 7. (a) Validation of the general plot-aggregate allometry approach in fty independently measured 0.81.0 ha plots in Panamanian and Peruvian forests. The performance of the general plot-aggregate allometry is also shown for partitions at (b) 0.080.12 ha; (c) 0.2425 ha; (c) 0.480.50 ha. RMSE = root mean squared error, bias reects mean errors or lack-of-t between the predicted and observed data. Solid lines are the result of quantile regression, with -values of 0.975 and 0.025 (blue), 0.68 and 0.32 (red), and 0.50 (black).

G.P. Asner, J. Mascaro / Remote Sensing of Environment 140 (2014) 614624

623

Fig. 8. Effects of plot size on the uncertainty between LiDAR-predicted and EACDeld using the general approach for validation sites in Panama and Peru. The theoretical line is specic to a given error level at a given plot size, in this case, the error observed in a regional calibration in Panama by Mascaro, Asner, et al. (2011) at a plot size of 0.36 ha. Based on this, the extrapolation of the error according to the inverse square root of the plot size is shown by the theoretical line. The t is a 2nd-order polynomial applied to log-transformed data for the three sites shown. The t departs particularly at the extremes, and this is likely due to increasing spatial autocorrelation of errors at both small and large scales (Mascaro, Detto, Asner, & Muller-Landau, 2011).

(e.g., Gobakken et al., 2012; Naesset et al., 2011). There are practical trade-offs as well: large plots are more time-consuming to establish and census, so obtaining enough of them can be challenging and expensive. Finding a good balance between plot size and number remains an important issue ripe for further study. The approach we presented here may facilitate improved or easier LiDAR-assisted tropical forest carbon monitoring by allowing for inputs of simple stocking and wood density information, both of which can be acquired quickly in the eld and/or via current published databases provided they are co-registered with LiDAR data. Strong evidence for this rests in results shown in Fig. 7. We left out of the general calibration fty plots comprising 48.4 ha of forest. We then tested the method on these validation plots to derive a RMSE of 13 Mg C ha 1, which is slightly greater than 10% uncertainty from the calibration phase. By re-examining these eld plot data in different congurations, we found that uncertainties decline toward 10% at 1-ha resolution (Fig. 8). A recent compilation of published work from tropical forests also demonstrated that uncertainties approach 10% as plot sizes increase to 1 ha (Zolkos et al., 2013). We emphasize that although our general modeling approach simplies the way allometric theory is applied to estimate aboveground carbon density, true measurements of aboveground carbon density can only be determined by harvesting and weighing the real carbon stock of whole forests (Clark & Kellner, 2012). Given how conserved tree allometry is across the tropics (Chave et al., 2005), whole-plot harvests are likely to be in close agreement with allometric estimationsbut any differences will affect eld-inventory based and remote sensing-based carbon mapping in a similar fashion. In the meantime, our results suggest that the general LiDAR approach can be applied in any tropical vegetation type with relatively few a priori inputs, potentially achieving up to 90% agreement with the traditional eld inventory approach. If consistent, this general LiDAR approach could improve carbon stock mapping and carbon emissions monitoring by covering far more geography with airborne (and eventually space-based) LiDARs than can be acquired on the ground. This would, in turn, greatly decrease the labor, extensive

Fig. 9. Examples of 30-m radius (0.28 ha) eld plots from lowland Amazonian forest in Peru. The upper image contains scattered large trees on deforested land, whereas the lower image shows a naturally occurring bottomland. Note nearly identical Lorey's height values despite major differences in top-of-canopy height (TCH) and eld-estimated aboveground carbon density (EACDeld).

time periods and high per-hectare cost of maintaining ground-based forest inventory networks.

Acknowledgements We thank R. Tupayachi, F. Sinca, N. Jaramillo, P. Martinez, L. Carranza, and A. Escudero for eld assistance. We thank T. Kennedy-Bowdoin, C. Anderson, D. Knapp, and R. Martin for LiDAR data collection and processing. We thank S. Goetz, R. Dubayah, and ve additional reviewers for comments and criticisms that greatly improved the manuscript. This study was supported by the John D. and Catherine T. MacArthur Foundation. Data acquisition for the Azuero dry forest plots was supported by a grant from the Grantham Foundation for the Protection of the Environment to J. Hall. The Carnegie Airborne Observatory is made possible by the Gordon and Betty Moore Foundation, the John D. and Catherine T. MacArthur Foundation, Avatar Alliance Foundation, W. M. Keck Foundation, the Margaret A. Cargill Foundation, Grantham Foundation for the Protection of the Environment, Mary Anne Nyburg Baker and G. Leonard Baker Jr., and William R. Hearst III.

624

G.P. Asner, J. Mascaro / Remote Sensing of Environment 140 (2014) 614624 remote sensing: Factors affecting the generality of relationships. Global Ecology and Biogeography, 12, 147159. Ene, L. T., Naesset, E., Gobakken, T., Gregoire, T. G., Stahl, G., & Nelson, R. (2012). Assessing the accuracy of regional LiDAR-based biomass estimation using a simulation approach. Remote Sensing of Environment, 123, 579592. Feldpausch, T. R., Lloyd, J., Lewis, S. L., Brienen, R. J. W., Gloor, M., Monteagudo Mendoza, A., et al. (2012). Tree height integrated into pantropical forest biomass estimates. Biogeosciences, 9, 33813403. Gobakken, T., Naesset, E., Nelson, R., Bollandsas, O. M., Gregoire, T. G., Stahl, G., et al. (2012). Estimating biomass in Hedmark County, Norway using national forest inventory eld plots and airborne laser scanning. Remote Sensing of Environment, 123, 443456. Goerndt, M. E., Monleon, V. J., & Temesgen, H. (2010). Relating forest attributes with areaand tree-based Light Detection and Ranging metrics for Western Oregon. Western Journal of Applied Forestry, 25, 105111. Goldberger, A. S. (1968). The interpretation and estimation of Cobb-Douglas functions. Econometrica, 36, 464472. Gonzalez, P., Asner, G. P., Battles, J. J., Lefsky, M.A., Waring, K. M., & Palace, M. (2010). Forest carbon densities and uncertainties from Lidar, QuickBird, and eld measurements in California. Remote Sensing of Environment, 114, 15611575. Harris, N. L., Brown, S., Hagen, S.C., Saatchi, S. S., Petrova, S., Salas, W., et al. (2012). Baseline map of carbon emissions from deforestation in tropical regions. Science, 336, 15731575. Lefsky, M.A. (2010). A global forest canopy height map from the Moderate Resolution Imaging Spectroradiometer and the Geoscience Laser Altimeter System. Geophysical Research Letters, 37. http://dx.doi.org/10.1029/2010GL043622. Lefsky, M.A., Cohen, W. B., Harding, D. J., Parker, G. G., Acker, S. A., & Gower, S. T. (2002). Lidar remote sensing of above-ground biomass in three biomes. Global Ecology and Biogeography, 11, 393399. Lefsky, M.A., Harding, D. J., Keller, M., Cohen, W. B., Carabajal, C. C., Espirito-Santo, F. D., et al. (2005). Estimates of forest canopy height and aboveground biomass using ICESat. Geophysical Research Letters, 32, L22S02. Magnussen, S., Naesset, E., Gobakken, T., & Frazer, G. (2012). A ne-scale model for area-based predictions of tree-size-related attributes derived from LiDAR canopy heights. Scandinavian Journal of Forest Research, 27, 312322. Malhi, Y., Wood, D., Baker, T. R., Wright, J., Phillips, O. L., Cochrane, T., et al. (2006). The regional variation of aboveground live biomass in old-growth Amazonian forests. Global Change Biology, 12, 11071138. Martin, A.R., & Thomas, S.C. (2011). A reassessment of carbon content in tropical trees. PLoS ONE, 6, e23533. Mascaro, J., Asner, G. P., Muller-Landau, H. C., Van Breugel, M., Hall, J., & Dahlin, K. M. (2011). Controls over aboveground forest carbon density on Barro Colorado Island, Panama. Biogeosciences, 8, 16151629. Mascaro, J., Detto, M., Asner, G. P., & Muller-Landau, H. C. (2011). Evaluating uncertainty in mapping forest carbon with airborne LiDAR. Remote Sensing of Environment, 115, 37703774. McRoberts, R. E., Nsset, E., & Gobakken, T. (2013). Inference for lidar-assisted estimation of forest growing stock volume. Remote Sensing of Environment, 128, 268275. Muller-Landau, H. C., Condit, R. S., Chave, J., Thomas, S.C., Bohlman, S. A., Bunyavejchewin, S., et al. (2006). Testing metabolic ecology theory for allometric scaling of tree size, growth and mortality in tropical forests. Ecology Letters, 9, 575588. Nsset, E. (2009). Effects of different sensors, ying altitudes, and pulse repetition frequencies on forest canopy metrics and biophysical stand properties derived from small-footprint airborne laser data. Remote Sensing of Environment, 113, 148159. Nsset, E., Gobakken, T., Solberg, S., Gregoire, T. G., Nelson, R., Stahl, G., et al. (2011). Model-assisted regional forest biomass estimation using LiDAR and InSAR as auxiliary data: A case study from a boreal forest area. Remote Sensing of Environment, 115. Niklas, K. J. (2006). A phyletic perspective on the allometry of plant biomasspartitioning patterns and functionally equivalent organ-categories. New Phytologist, 171, 2740. Ni-Meister, W., Lee, S., Strahler, A. H., Woodcock, C. E., Schaaf, C., Yao, T., et al. (2010). Assessing general relationships between aboveground biomass and vegetation structure parameters for improved carbon estimate from lidar remote sensing. Journal of Geophysical Research, 115, G00E11. Omasa, K., Qiu, G. Y., Watanuki, K., Yoshimi, K., & Akiyama, Y. (2003). Accurate estimation of forest carbon stocks by 3-D remote sensing of individual trees. Environmental Science & Technology, 37, 11981201. Saatchi, S. S., Harris, N. L., Brown, S., Lefsky, M., Mitchard, E. T. A., Salas, W., et al. (2011). Benchmark map of forest carbon stocks in tropical regions across three continents. Proceedings of the National Academy of Sciences, 108, 98999904. Stegen, J. C., Swenson, N. G., Valencia, R., Enquist, B. J., & Thompson, J. (2009). Above-ground forest biomass is not consistently related to wood density in tropical forests. Global Ecology and Biogeography, 18, 617625. ter Steege, H., Pitman, N. C. A., Phillips, O. L., Chave, J., Sabatier, D., Duque, A., et al. (2006). Continental-scale patterns of canopy tree composition and function across Amazonia. Nature, 443, 444447. West, G. B., Brown, J. H., & Enquist, B. J. (1999). A general model for the structure and allometry of plant vascular systems. Nature, 400, 664667. Wright, S. J., Yavitt, J. B., Wurzburger, N., Turner, B.L., Tanner, E. V. J., Sayer, E. J., et al. (2011). Potassium, phosphorus, or nitrogen limit root allocation, tree growth, or litter production in a lowland tropical forest. Ecology, 92, 321330. Zolkos, S. G., Goetz, S. J., & Dubayah, R. (2013). A meta-analysis of terrestrial aboveground biomass estimation using lidar remote sensing. Remote Sensing of Environment, 128, 289298.

Appendix A

Table A1 Basic LiDAR specications for the Carnegie Airborne Observatory (CAO) Alpha (Asner et al., 2007) and AToMS (Asner, Knapp, et al., 2012) sensor packages. PRF = pulse repetition frequency; FWHM = full width at half maximum; ns = nanoseconds; kW = kilowatts. CAO alpha PRF (kHz) 50 70 100 Pulse width (FWHM, ns) 11.5 15.3 20.4 Peak power (kW) 9.8 5.5 2.9 CAO AToMS PRF (kHz) 50 100 200 Pulse width (FWHM, ns) 7.2 6.9 6.9 Peak power (kW) 7.6 7.7 6.8

Appendix B. Supplementary data Supplementary data to this article can be found online at http://dx. doi.org/10.1016/j.rse.2013.09.023. References
Asner, G. P., Clark, J. K., Mascaro, J., Vaudry, R., Chadwick, K. D., Vieilledent, G., et al. (2012). Human and environmental controls over aboveground carbon storage in Madagascar. Carbon Balance and Management, 7. http://dx.doi.org/10.1186/1750-0680-1187-1182. Asner, G. P., Hughes, R. F., Mascaro, J., Uowolo, A., Knapp, D. E., Jacobson, J., et al. (2011). High-resolution carbon mapping on the million-hectare Island of Hawai'i. Frontiers in Ecology and the Environment, 9, 434439. Asner, G. P., Hughes, R. F., Varga, T. A., Knapp, D. E., & Kennedy-Bowdoin, T. (2009). Environmental and biotic controls over aboveground biomass throughout a tropical rain forest. Ecosystems, 12, 261278. Asner, G. P., Knapp, D. E., Boardman, J., Green, R. O., Kennedy-Bowdoin, T., Eastwood, M., et al. (2012). Carnegie Airborne Observatory-2: Increasing science data dimensionality via high-delity multi-sensor fusion. Remote Sensing of Environment, 124, 454465. Asner, G. P., Knapp, D. E., Kennedy-Bowdoin, T., Jones, M.O., Martin, R. E., Boardman, J., et al. (2007). Carnegie Airborne Observatory: In-ight fusion of hyperspectral imaging and waveform light detection and ranging (LiDAR) for three-dimensional studies of ecosystems. Journal of Applied Remote Sensing, 1. http://dx.doi.org/10.1117/1111. 2794018. Asner, G. P., Mascaro, J., Muller-Landau, H. C., Vieilledent, G., Vaudry, R., Rasamoelina, M., et al. (2012). A universal airborne LiDAR approach for tropical forest carbon mapping. Oecologia, 168, 11471160. Asner, G. P., Powell, G. V. N., Mascaro, J., Knapp, D. E., Clark, J. K., Jacobson, J., et al. (2010). High-resolution forest carbon stocks and emissions in the Amazon. Proceedings of the National Academy of Sciences, 107, 1673816742. Baccini, A., Goetz, S. J., Walker, W. S., Laporte, N. T., Sun, M., Sulla-Menashe, D., et al. (2012). Estimated carbon dioxide emissions from tropical deforestation improved by carbon-density maps. Nature Climate Change, 2, 182185. Baker, T. R., Phillips, O. L., Malhi, Y., Almeida, S. A. S., Arroyo, L., Di Fiore, A., et al. (2004). Variation in wood density determines spatial patterns in Amazonia forest biomass. Global Change Biology, 10, 545562. Baskerville, G. (1972). Use of logarithmic regression in the estimation of plant biomass. Canadian Journal of Forest Research -Revue Canadienne De Recherche Forestiere, 2, 4953. Brown, S. (1997). Estimating biomass and biomass change of tropical forests. Forestry Papers. Rome: FAO. Cade, B.S., & Noon, B. R. (2003). A gentle introduction to quantile regression for ecologists. Frontiers in Ecology and the Environment, 1, 412420. Chave, J., Andalo, C., Brown, S., Cairns, M.A., Chambers, J. Q., Eamus, D., et al. (2005). Tree allometry and improved estimation of carbon stocks and balance in tropical forests. Oecologia, 145, 8799. Chave, J., Chust, G., Condit, R., Aguilar, S., Hernandez, A., Lao, S., et al. (2004). Error propagation and scaling for tropical forest biomass estimates. In Y. Malhi, & O. Phillips (Eds.), Tropical forests and global atmospheric change (pp. 155166). London: Oxford University Press. Chave, J., Coomes, D., Jansen, S., Lewis, S. L., Swenson, N. G., & Zanne, A. E. (2009). Towards a worldwide wood economics spectrum. Ecology Letters, 12, 351366. Clark, D. B., & Kellner, J. R. (2012). Tropical forest biomass estimation and the fallacy of misplaced concreteness. Journal of Vegetation Science, 23, 11911196. Colgan, M. S., Asner, G. P., & Swemmer, T. (2013). Harvesting tree biomass at the stand-level to assess the accuracy of eld and airborne biomass estimation in savannas. Ecological Applications, 23, 11701184. Drake, J. B., Dubayah, R. O., Clark, D. B., Knox, R. G., Blair, J. B., Hofton, M.A., et al. (2002). Estimation of tropical forest structural characteristics using large-footprint lidar. Remote Sensing of Environment, 79, 305319. Drake, J. B., Knox, R. G., Dubayah, R. O., Clark, D. B., Condit, R., Blair, J. B., et al. (2003). Above-ground biomass estimation in closed canopy neotropical forests using lidar