Estimation of above‐ground biomass of large tropical trees with terrestrial LiDAR

Tropical forest biomass is a crucial component of global carbon emission estimations. However, calibration and validation of such estimates require accurate and effective methods to estimate in situ above‐ground biomass (AGB). Present methods rely on allometric models that are highly uncertain for large tropical trees. Terrestrial laser scanning (TLS) tree modelling has demonstrated to be more accurate than these models to infer forest AGB. Nevertheless, applying TLS methods on tropical large trees is still challenging. We propose a method to estimate AGB of large tropical trees by three‐dimensional (3D) tree modelling of TLS point clouds. Twenty‐nine plots were scanned with a TLS in three study sites (Peru, Indonesia and Guyana). We identified the largest tree per plot (mean diameter at breast height of 73.5 cm), extracted its point cloud and calculated its volume by 3D modelling its structure using quantitative structure models (QSM) and converted to AGB using species‐specific wood density. We also estimated AGB using pantropical and local allometric models. To assess the accuracy of our and allometric methods, we harvest the trees and took destructive measurements. AGB estimates by the TLS–QSM method showed the best agreement in comparison to destructive harvest measurements (28.37% coefficient of variation of root mean square error [CV‐RMSE] and concordance correlation coefficient [CCC] of 0.95), outperforming the pantropical allometric models tested (35.6%–54.95% CV‐RMSE and CCC of 0.89–0.73). TLS–QSM showed also the lowest bias (overall underestimation of 3.7%) and stability across tree size range, contrasting with the allometric models that showed a systematic bias (overall underestimation ranging 15.2%–35.7%) increasing linearly with tree size. The TLS–QSM method also provided accurate tree wood volume estimates (CV RMSE of 23.7%) with no systematic bias regardless the tree structural characteristics. Our TLS–QSM method accounts for individual tree biophysical structure more effectively than allometric models, providing more accurate and less biased AGB estimates for large tropical trees, independently of their morphology. This non‐destructive method can be further used for testing and calibrating new allometric models, reducing the current under‐representation of large trees in and enhancing present and past estimates of forest biomass and carbon emissions from tropical forests.


| INTRODUCTION
The above-ground carbon in tropical forests represents 40% of the total carbon stocked in forests globally (Gibbs, Brown, Niles, & Foley, 2007). However, the estimation of tropical forest carbon stocks presents large uncertainties (Mitchard et al., 2013(Mitchard et al., , 2014. Forest carbon stocks are not measured directly, but derived either from interpolation or extrapolation of point estimates of the above-ground biomass (AGB) contained in forest inventory plots, or from measurements of remote sensing proxies calibrated with plot-based AGB estimates (Gibbs et al., 2007).
The only way to truly and directly measure forest AGB implies cutting and weighing the mass of all trees in the plot, which is costly and causes a negative impact, and is thus seldom executed (Clark & Kellner, 2012). Instead, plot AGB is estimated from aggregation of individual tree AGB estimates. These tree AGB estimates are indirectly derived from easily measured tree parameters (diameter at breast height [DBH], height and wood density derived from tree species identification) by means of allometric models, which relate these tree parameters with real tree AGB measured in destructive sampling studies (Chave et al., 2005). This indirect estimation approach introduces an error propagation chain. The biggest source of error is derived from the allometric models, hence its appropriate selection is the most important aspect to improve the accuracy of AGB estimates (Molto, Rossi, & Blanc, 2013).
The uncertainty in the tree AGB estimation is even greater for large tropical trees (DBH >70 cm) because AGB in large trees varies more than in small trees (Chave et al., 2005;Goodman, Phillips, & Baker, 2014;Ploton et al., 2016;Slik et al., 2013), and due to the presence of buttresses is prone to larger measurement error (Chave et al., 2014). Moreover, it is particularly relevant to accurately estimate AGB of large trees because of their major influence on the tropical forest AGB variation (Slik et al., 2013;Stegen et al., 2011).
As an alternative, remote sensing systems can be used to estimate tropical forest carbon stocks. One of the most promising remote sensing approaches to estimate forest AGB is via light detection and ranging (LiDAR), either via spaceborne platforms (e.g. ICESat), airborne laser scanning or terrestrial laser scanning (TLS). Laser pulses from LiDAR instruments can penetrate the forest canopy providing good estimates of forest canopy heights and structure, from which AGB along the vertical profile and canopy cover can be estimated (Goetz & Dubayah, 2011). TLS data provide the highest level of three-dimensional (3D) detail of forest and tree structure . Currently, TLS data are being used to model 3D structure of individual trees allowing direct measurements of forest and tree structural parameters such as DBH (Bauwens, Bartholomeus, Calders, & Lejeune, 2016), tree height (Király & Brolly, 2007), crown dimensions (Holopainen, Vastaranta, & Kankare, 2011) and individual branches (Raumonen, Kaasalainen, Kaasalainen, & Kaartinen, 2011). Several review articles provide additional information about the characteristics of TLS and its use for forestry surveying .
Several approaches estimate forest AGB by exploiting the capability of TLS data to characterize forest structure at tree level. A simple approach is to measure tree structural parameters from a TLS 3D point cloud and apply allometric models to relate the measured parameters with AGB (e.g. Yao et al. (2011)). However, this method still relies on allometric models. A different kind of approach has been developed to reconstruct the complete 3D tree architecture from TLS data rather than a single or few structural parameters. Quantitative structure models (QSMs; Delagrange, Jauvin, & Rochon, 2014;Hackenberg, Wassenberg, Spiecker, & Sun, 2015;Raumonen et al., 2013) are architectural tree models reconstructed from the TLS point cloud of individual trees and allow volume measurements. The estimated tree volume is converted to tree AGB by multiplying it by the specific wood density Hackenberg et al., 2015).
Thus, this method estimates AGB based on the biophysical modelling of specific tree structure rather than the allometric models which are based on empirical relationships from a sample of trees and rely on a limited number of tree structural parameters.
The QSM reconstruction method developed by Raumonen et al. (2013) has been applied for wood volume estimation and AGB size. The TLS-QSM method also provided accurate tree wood volume estimates (CV RMSE of 23.7%) with no systematic bias regardless the tree structural characteristics. 4. Our TLS-QSM method accounts for individual tree biophysical structure more effectively than allometric models, providing more accurate and less biased AGB estimates for large tropical trees, independently of their morphology. This nondestructive method can be further used for testing and calibrating new allometric models, reducing the current under-representation of large trees in and enhancing present and past estimates of forest biomass and carbon emissions from tropical forests.

K E Y W O R D S
above-ground biomass, allometric models, LiDAR, terrestrial laser scanning, tree volume, tropical trees, 3D modeling estimation in boreal and temperate forest  and in more structurally complex tropical forests in Gabon (Disney et al., 2014). AGB estimates derived from this approach in Australia showed a higher agreement with reference values from destructive sampling (coefficient of variation of root mean square error [CV RMSE] = 16.1%) compared to AGB estimates derived by allometric models (CV RMSE = 46.2%-57%) . However, the accuracy of AGB estimates in tropical forest trees has not been investigated yet with reference data.
Several challenges arise when one wants to estimate tree AGB in a tropical forest using QSM. First, for very large and complex trees there is a lack of reference data to validate the 3D reconstruction models from TLS. Furthermore, the structural complexity of a tropical forest can potentially have a large influence on acquired TLS data. This requires careful design of an appropriate scanning pattern to diminish vegetation occlusion and to allow accurate reconstruction of the 3D structure of trees (Wilkes et al., 2016).
Here, we assess the potential and accuracy of volume reconstruction using QSMs for estimating AGB of large tropical forest trees. For this, 29 plots were scanned with TLS and one large tree per plot was destructively sampled afterwards. With the TLS data acquired, we (1) optimized the QSM tree volume reconstruction method based on a subsample of nine of the 29 trees. After each tree was scanned and harvested, we (2) performed in situ destructive measurements to independently estimate tree volume for comparison with model estimates and calculate their accuracy. Finally, using the independent tree dataset (remaining 20 trees non-used in point 1), we (3) compared the accuracy of the AGB estimates based on QSMs with the accuracy of the AGB estimates based on pantropical and local allometric models.

| Study area
We acquired field data from 29 plots across three tropical forest sites in Peru, Indonesia and Guyana. Table 1 shows the description of each site.

| TLS sampling and field data collection
Plots were established around a tree to be harvested after the laser scanning. Plot spatial design and tree selection are detailed in Appendix S1. Once the plots were set up, we scanned the plot with TLS, performed a forest inventory, harvested the selected tree and measured the geometric structure of the harvested tree.

| TLS data acquisition
TLS datasets were acquired using a RIEGL VZ-400 3D ® terrestrial laser scanner (RIEGL Laser Measurement Systems GmbH, Horn, Austria). This scanner is a discretized multiple-return LiDAR scanner and its specifications are shown in Table 2. Details of the sampling design are described in Appendix S2.

| Forest inventory data collection
For each tree, we measured DBH (or diameter above buttresses), tree height, height of first branch and crown width. We measured DBH with a forestry tape and tree height with a Nikon "Forestry-Pro" (Hayama, Japan) laser hypsometer with precisions of 0.01 and 0.2 m respectively. An experienced taxonomist (specialist of the local flora) identified the trees at species level.

| Harvested tree reference measurements
We measured the geometry of the stem, buttresses and branches of each harvested tree. As in Figure 1(1), tree stem diameters (1a) were  Kankare et al. (2013). For trees with buttresses or major irregularities, we measured as in Figure 1(2). Finally, we measured all branches until tapered diameter ≤10 cm by measuring each internode independently as in Figure 1(3).

| Tree wood volume estimation from 3D QSM
We co-registered each individual TLS scan into a single plot point cloud using RiScan PRo software (version 2.0; RIEGL Laser Measurement Systems GmbH, www.riegl.com) and the accuracy of our co-registration was kept below 1 cm.
We reconstructed the woody structure of trees using the QSM method developed by Raumonen et al. (2013) and further developed by  and Raumonen et al. (2015). The We filtered out cylinders with diameter <10 cm from resulting QSMs to be consistent with the reference volume estimation and we calculated the total tree volume by summing the volume of all remaining cylinders. Due to the random generation of the QSM patches (point cloud partition into small segments) Raumonen et al., 2015), for each parameter set used we reconstructed 20 QSMs and averaged the volume of the 20 model realizations.

| Sensitivity analysis and independent estimation of QSM accuracy
We split our tree population into two independent sub-datasets using stratified random sampling without replacement: a tree dataset of nine trees (three from each study area) for the sensitivity analysis of a QSM parameter value, and a second tree dataset of 20 trees (the remaining six trees for Peru and seven for Guyana and Indonesia) for independent estimation of tree volume and AGB estimates accuracy.
The reconstruction of the QSMs requires a few input parameters, of which the size of the point cloud segments-expressed by the "surface patches diameter" (hereafter "PatchDiam")-had the most influence on the outcome . A detailed explanation of the QSM parameters and QSM sensitivity to them is provided in the Supporting Information and in Raumonen et al. (2013Raumonen et al. ( , 2015 and .
Our sensitivity analysis consisted of the evaluation of the QSMs optimal PatchDiam value, which gives the most accurate volume estimate among the different PatchDiam values tested (1, 2.5, 5, 7.5, 10 and 15 cm). For each tree in the sensitivity analysis tree dataset, we compared the mean estimated volume (from the 20 QSM realizations per PatchDiam) against the tree volume obtained from the destructive measurements. We computed tree volume estimation RMSE. The optimal PatchDiam was chosen as the one that minimized the RMSE.
Once the optimal PatchDiam was found, we assessed the stability of the optimization procedure. We replicated the stratified F I G U R E 1 Tree geometry measurements.
(1) Stem diameter (1a) every metre (1b) until start of first branch. For trees with buttresses (2): diameter in two orthogonal directions (2a) and for each buttress horizontal length (from the furthest point to the stem) (2b); width (mean width between the tip and the buttress intersection with the stem) (2c); and height (from the ground to the highest insertion point of the buttress into the stem) (2d). For branches (3): proximal diameter at the base of each internode and above flaring (3a), distal diameter at the tip of each internode and below flaring of the next node (3b) and branch length from the base to the tip of each internode (3c) PatchDiam's obtained (the one providing the smallest RMSE in each of the 1,000 samples) as well as the variability of the RMSE results (range, mean and standard deviation) for all samples with a given optimal PatchDiam.
Finally, the optimized PatchDiam was used to run QSM for the independent estimation dataset (20 trees) and to calculate the tree volume following the same procedure described above. We used matlab (The MathWorks Inc. 2014) for QSM reconstruction and "R" (R Core Team 2013) for further calculations.

| Tree volume estimation from reference measurements
We used the reference geometric measurements (Section 2.2.3) from each harvested tree to determine the tree reference volume. We applied the Smalian formula as in Nogueira, Nelson, and Fearnside (2005) to estimate volume of stem sections and individual branches until 10 cm diameter, while for the buttresses we applied a general prism volume formula. Detailed information can be found in Appendix S4. Total tree wood volume was calculated as the sum of volumes of main stem, large branches (>10 cm diameter) and buttresses.
As in Berger, Gschwantner, McRoberts, and Schadauer (2014), any misrepresentation of the main stem and branches volumes by the Smalian approximation and any measurement error taken were considered negligible and ignored. Furthermore, the sum of all cylinders was assumed to represent the true tree volume with no error and that the wood volume was measured without error.

| Tree AGB estimation from volume models and wood density
We calculated individual tree AGB by multiplying individual tree wood volume estimates by the specific basic wood density (ρ). Values of ρ were assigned to the finest taxonomic level possible (species, genus or family) according to the Global Wood Density Database Zanne et al., 2009) and tree species identified in the field. We applied an expansion factor accounting for small branches (≤10 cm diameter). The expansion factor related the volume of small branches to the one of the large branches (>10 cm diameter). We calculated an expansion factor of 0.255 using data from biomass destructive sampling of 51 trees in a nearby Peruvian Amazon forest site (Goodman, Phillips, & Baker, 2013;Goodman et al., 2014). We used the same value for Peru and Guyana (0.255), while we calculated the expansion factor for Indonesia (0.28) from our own collected data. The final contribution of small branches to tree volume was 10%, 14% and 7% for Guyana, Peru and Indonesia respectively.

| Tree AGB estimation from allometric models
We estimated AGB using 12 allometric models, of which eight were locally calibrated and four pantropical (see Appendix S5).
The pantropical allometric models used were developed by Chave et al. (2005), which have been recently improved (Chave et al., 2014).
The local allometric models used for the Peruvian trees were developed by Goodman et al. (2014), while allometric models for Indonesian trees were developed by Manuri et al. (2014) and Jaya, Siregar, Daryono, and Suhartana (2007). No suitable local allometric model could be found for Guyana. The details of the allometric models used to estimate AGB for the harvested trees are described in the Supporting Information.

| AGB estimation models accuracies and uncertainty assessment
We used the 20 trees in the dataset reserved for the independent estimation to compare the accuracy of AGB estimates from our TLS-QSM approach (against reference AGB) vs. the accuracy obtained from allometric models (against reference AGB). As general indicators of model accuracy, RMSE (in m 3 and Mg), CV RMSE (in %) and mean relative error (in %) were calculated. Slope and intercept values of orthogonal regression models between AGB modelled and reference values were used to identify departure from the 1:1 line, and the R-squared (hereafter R 2 ) was used to judge the fitting of these regressions. Finally, the concordance correlation coefficient (CCC) was calculated to compare agreement of AGB model estimates with AGB reference and to previously reported agreement using the QSM method .
To assess the uncertainty in the tree AGB estimations, we used the error propagation approach (Equation 4) to account for the uncertainties in the models components. We combined them and assumed that the uncertainties were statistically independent (not correlated and with a Gaussian distribution). We used Equation 4  (1) For AGB estimations from QSM volume models, the model uncertainty components considered were the wood volume and wood density.
The uncertainty in tree wood volume by QSM is provided by the standard deviation of the 20 QSM realizations per tree. For the estimation of wood densities uncertainties, we assumed for all species the same standard deviation of 10% of the mean as used by Chave et al. (2004). Likewise, to assess the uncertainty in the tree AGB estimation from allometric models, we used the uncertainties reported for each model (see Appendix S5). To assess the uncertainty in the tree AGB estimation from reference volume estimates, we considered two components: wood density (as described for QSM) and expansion factor. For the expansion factor, we assumed an error of 12.5% as reported in Segura and Kanninen (2005).

| Tree volume estimation with QSM
The results of the tree volume modelling with the TLS-QSM approach are divided into two steps: (1) QSM sensitivity analysis with nine trees to determine QSM optimal parameters and then (2) an independent assessment of the tree volume estimation accuracy with an independent sample of 20 trees.

| Sensitivity analysis of QSM tree volume modelling
The TLS-QSM tree volume estimation error (RMSE) when compared with the reference volume measurements decreased with decreasing PatchDiam (Table 3) until it reached a minimum error for PatchDiam of 2.5 cm, and then it increased again for smaller PatchDiam. This is in line with the results of the sensitivity analysis in  and . Therefore, 2.5 cm was considered the optimal PatchDiam, and thus selected for the tree volume estimation of the remaining tree dataset.
The stability assessment of PatchDiam optimization procedure showed that in 75% of the 1,000 random sampling replicates the optimal PatchDiam was 2.5 cm. Despite the relatively small sample reserved for the sensitivity analysis (9 out of 29 trees), the optimal PatchDiam was relatively stable regardless of the characteristics of the randomly selected trees.

| Independent assessment of tree volume estimation from TLS-QSM
To assess the accuracy of the tree wood volume estimation by the TLS-QSM, we compared the volume estimates by the TLS-QSM with the reference volume estimates from destructive measurements ( Figure 3).
The R 2 of the linear model describing the agreement of both datasets (Figure 3 blue line) was 0.9. Its slope was 0.93 indicating that the QSMs slightly underestimated the tree volume for the largest trees. The RMSE was 3.29 m 3 , compared with the mean tree volume of 15.13 m 3 , leading to a CV RMSE of 23.7%. Figure 3 shows that the TLS-QSM performed similarly throughout the three different sites, despite the three study areas contained different tree species, sizes and shapes. Results differ between "small trees" (DBH ≤ 70 cm, corresponding approximately with 9 Mg, hereafter small trees) and "large trees" (DBH > 70 cm, hereafter large trees). For small treeswhich were mostly part of the Indonesian dataset-TLS-QSM models showed less uncertainty and less deviation from the reference compared to large trees.
On the other hand, the analysis of the residuals (Figure 4) reveals that for small trees and large trees the model did not systematically tend to overestimate nor underestimate the volume. Despite the larger uncertainty in the volume estimation for large trees, there was no large systematic bias for larger tree size (Figure 4).
Buttresses were predominately absent in small trees, which had a better agreement with the reference data than trees with buttresses.  Our QSM modelling did not perform a detailed buttress modelling, but a cylinder fitting, which might be the cause of the higher residuals in the trees with buttresses.   Figure 6 displays the agreement between the AGB-modelled based on the TLS-QSM approach and local allometric models (y-axis) against AGB-reference (x-axis) for the sites where local allometric models were available.

| Overall accuracy within study sites: TLS-QSM vs. local allometric models
For the Peruvian study area the TLS-QSM approach is the closest to the 1:1 line, whereas the deviation from the 1:1 line is clearly larger for the three local allometric models tested, which systematically underestimate the AGB of large trees. The TLS-QSM approach showed 10% and 50% lower RMSE and 80% and 85% lower bias than the most-and least-accurate local allometric models. The agreement between TLS-QSM estimates and reference values expressed as CCC is higher (0.96) compared to the most-and least-accurate allometric models (0.76-0.92; Table 5).
For the Indonesian study area, unlike for the Peruvian site, the local allometric models showed lower RMSE and bias than the TLS-QSM for this particular subset of trees. The best local allometric model had a 44% smaller RMSE than the TLS-QSM, was closer to the 1:1 line and had a higher agreement with reference values (CCC = 0.96) than our approach (0.92) ( Table 6).
F I G U R E 4 Analysis of volume estimation residuals. Trees with diameter at breast height (DBH) ≤ 70 cm were classified as small size trees (red colour) and trees with DBH > 70 cm were classified as large trees (blue colour). Coefficient "a" denotes tree with buttresses while coefficient "b" denotes absence of tree buttresses Chave14.eq.7 C have14.eq.4

| Consistent and accurate AGB estimation of tropical trees from QSMs
We found that the TLS-QSM approach can provide reliable and accurate AGB estimates for large tropical trees (DBH > 70 cm), outperforming the accuracy of all the pantropical allometric models tested.
To the best of our knowledge, this is the first study assessing the accuracy of tropical trees AGB estimates using QSMs from TLS point clouds of trees across different tropical forest regions. A previous study by Disney et al. (2014) presented a proof of concept for the use of TLS-QSM for tree AGB estimation of tropical trees in Gabon, but in their research no tropical trees were harvested, thus the accuracy of its AGB estimates could not be assessed but only compared to the AGB estimates provided by allometric models. Our study showed that AGB estimations by allometric models often are not a reliable indicator of AGB for large tropical trees. This issue was also addressed by Clark and Kellner (2012),  and Ploton et al. (2016). Clark and Kellner (2012) and  both noted that large trees are under-represented in calibration T A B L E 4 Accuracies of AGB estimations across sites by the TLS-QSM approach and by pantropical allometric models

| AGB estimations by TLS-QSM vs. pantropical allometric models
Across the three sites the TLS-QSM method to estimate AGB was more accurate than the most accurate pantropical allometric model evaluated (Chave05 m1.3, in Appendix S5), with an absolute improvement of 7.2% less CV RMSE (Table 4). This accuracy improvement was even more pronounced in terms of bias reduction. Moreover, TLS-QSM showed a higher agreement with reference values (CCC = 0.95) compared to the most accurate pantropical allometric model (CCC = 0.89).  found a comparable trend of higher accuracy for their TLS-QSM method in relation to allometric models for estimating AGB of eucalyptus trees in Australia. The accuracy of the AGB estimates by TLS-QSM in our study was lower than the accuracy reported by , and our agreement (CCC = 0.95) was lower than the agreement found by method achieved an absolute improvement of 16% and 27% lower CV RMSE, which is comparable to the error decrease reported by .
It should be noted that the models accuracies were estimated by comparing each model AGB estimates with AGB reference estimates derived from destructive geometric measurements, rather than with AGB weighted. The uncertainties introduced in measuring stems, buttresses and branches volumes were taken into account, but-as in Kankare et al. (2013) and Berger et al. (2014)-the uncertainty due to the use of Smalian formula for estimating true volume was assumed to be negligible. Furthermore, the uncertainty introduced in the correction factor for small branches volume and in the application of a single species-specific wood density value for each tree instead of discriminating wood density for different woody fractions, both were not measured but taken from literature. Moreover, models uncertainties increasing with tree size indicates heteroscedasticity effects, which should be considered with caution when developing allometric models. This reinforces the need for improved methods for estimating large trees biomass, and for further research with larger datasets to assess the uncertainty on large trees biomass estimation.

| AGB estimations by QSM models vs. local allometric models in Indonesia and Peru
The TLS-QSM method also produced AGB estimates more accurate than the local allometric models for the Peruvian dataset, with a higher agreement (CCC = 0.96) with reference data than the local

| Reconstructing 3D woody structure of tropical forest trees using QSMs
We showed that the TLS-QSM method can be used to accurately estimate volume from 3D reconstructed structure of large tropical trees from scans in very dense forest with leaf-on conditions. The tree structure reconstructions for these large tropical trees contained larger uncertainty (higher variance on the QSM outcomes) than in previous studies Calders et al., 2013;Raumonen et al., 2015) which evaluated smaller trees and were located in more open forest conditions and less occluded trees. For the smallest trees in our study, the 3D reconstruction uncertainty values were closer to those previously reported by .
Consistent with previous QSM studies Calders et al., 2013;Disney et al., 2014;Raumonen et al., 2013), we optimized the reconstruction process based on the PatchDiam parameter, which was reported to be the most influential parameter (Calders et al., 2013). The main difference compared to  is in the method for judging the optimal reconstruction.
Our sample of tropical trees was characterized by being among the most challenging conditions for a 3D tree reconstruction method because the target trees were among the tallest trees in each plot and having the largest crown size and complexity. The combination of these limiting factors contributes to increased occlusion, in combination with very dense understorey, resulting in under-sampled areas in the tree crowns and larger uncertainties in the QSM reconstructions.
For these low-density point cloud areas the QSMs presented some unrealistic branching reconstructions. The low-density point cloud issue was also addressed by Raumonen et al. (2011Raumonen et al. ( , 2013. They stated that the reconstruction method was quite sensitive to low point cloud density and therefore, reliability of cylinders reconstructing small branches could be very low. Therefore, we discarded all branches with a diameter <10 cm and applied the expansion factor to account for their volume. Alternatively,  recently proposed an automated method for QSM parameterization. This method optimized the PatchDiam value based on the maximum match of QSM cylinders diameter with point cloud circle fitting diameter at four different heights along the main trunk. This approach focuses on comparing the reconstructed main trunk, regardless of the quality of the reconstructed tree crown. However, recent studies (Goodman et al., 2014;Ploton et al., 2016) showed the important contribution of the crown biomass to the total tree biomass for large tropical trees. Similarly, for the trees in our study, the crown contribution to the total tree biomass was 50% on average and even larger for the trees above 10 Mg (60% of the total tree biomass). Therefore, we decided not to implement the method of  for our study.
Future research should focus on developing an automated QSM optimization which optimizes the reconstruction of the entire tree and does not focus on the tree trunk alone. Automated optimization of this sort might enable to improve even further the accuracy of tree volume and AGB estimates of tropical trees from TLS data at large scale without harvesting trees.

| CONCLUSIONS
We present an approach to estimate tree wood volume and AGB

DATA ACCESSIBILITY
The individual trees TLS point cloud, QSM cylinder models, forest inventory and destructive sampling measurement data used for this research can be accessed in the LUCID repository (http://lucid.wur.nl/datasets/terrestrial-lidar-of-tropical-forests).
These datasets are owned by the CIFOR and Wageningen University.
The datasets are free to download and available for any use as long as the proper reference, as specified in the portal, is applied. For collaborations or questions please contact: jose.tanago@gmail.com, alvaro.lausarmiento@wur.nl or harm.bartholomeus@wur.nl.