Direct Estimation of Tree Volume and Aboveground Biomass Using Deep Regression with Synthetic Lidar Data

Habib Pourdelan; Zhengkang Xiang; Hugh Stewart; Cam Nicholson; Martin Tomko; Kourosh Khoshelham

cs.LGMar 4, 2026

Direct Estimation of Tree Volume and Aboveground Biomass Using Deep Regression with Synthetic Lidar Data

Authors: Habib Pourdelan, Zhengkang Xiang, Hugh Stewart, Cam Nicholson, Martin Tomko, Kourosh Khoshelham

Organizations: The University of Melbourne · Private Forestry Service Victoria, Australia · Nicon Rural Services, Australia

Abstract

Accurate estimation of forest biomass is crucial for monitoring carbon sequestration and informing climate change mitigation strategies. Existing methods often rely on allometric models, which estimate individual tree biomass by relating it to measurable biophysical parameters, e.g., trunk diameter and height. This indirect approach is limited in accuracy due to measurement uncertainties and the inherently approximate nature of allometric equations, which may not fully account for the variability in tree characteristics and forest conditions. This study proposes a direct approach that leverages synthetic point cloud data to train a deep regression network, which is then applied to real point clouds for plot-level wood volume and aboveground biomass (AGB) estimation. We created synthetic 3D forest plots with ground truth volume, which were then converted into point cloud data using a lidar simulator. These point clouds were subsequently used to train deep regression networks based on PointNet, PointNet++, DGCNN, and PointConv. When applied to synthetic data, the deep regression networks achieved mean absolute percentage error (MAPE) values ranging from 1.69% to 8.11%. The trained networks were then applied to real lidar data to estimate volume and AGB. When compared against field measurements, our direct approach showed discrepancies of 2% to 20%. In contrast, indirect approaches based on individual tree segmentation followed by allometric conversion, as well as FullCAM, exhibited substantially large underestimation, with discrepancies ranging from 27% to 85%. Our results highlight the potential of integrating synthetic data with deep learning for efficient and scalable forest carbon estimation at plot level.

Figures & tables

Figure 1: Schematic comparison of conventional indirect AGB estimation (segmentation + allometry + aggregation) versus the proposed direct plot-level regression approach.

Figure 2: Overview of the proposed methodolog

Model (domain in	Diameter	$\ln(a)$	$b$	Correction
parentheses)	measurement			factor (c)
Eucalypt ( $D_{130}<169$ cm)	$D_{130}$	-2.016	2.375	1.0668
Multi ( $D_{10}<62$ cm)	$D_{10}$	-2.757	2.474	1.0775
Shrub ( $D_{10}<50$ cm)	$D_{10}$	-3.007	2.428	1.1281
Other ( $D_{130}<102$ cm)	$D_{130}$	-1.693	2.220	1.0436

Table 1: Coefficients for allometric models for AGB using a predictor of diameter (D) measured at either 10 cm or 130 cm height.

Figure 3: Lidar data and field measurement of AGB were collected in two study areas, Knewleave and Jigsaw Farms, both located in Victoria, Australia.

Panel of Figure 4 — Figure 4: Synthetic data generation process. (a) 3D eucalyptus tree model, (b) Synthetic forest plot built in Blender, (c) Simulated point cloud.

Setting	Riegl VUX-1UAV22	HOVERMAP ST
Altitude (m)	35	30–40
Flight speed (m/s)	2	Not available
Pulse repetition frequency (kHz)	200	300
Scan frequency (Hz)	80	20
Max. returns per pulse	5	2
Scan angle (°)	180	360

Table 2: Acquisition settings of the real (HOVERMAP ST) and simulated (Riegl VUX-1UAV22) laser scanners

Hyperparameter	Value
Initial learning rate	0.001
Batch size	16
Number of epochs	200
Optimizer	AdamW
Weight decay	$1\times 10^{-4}$
Learning rate scheduler	Cosine annealing warm restarts ( Loshchilov and Hutter, 2016 )

Table 3: Training hyperparameters used for all deep regression models.

Metric	Synthetic		MF1995.1		MF1999.2		MF2007.2		HP2010.2		Knewleaves20
	RS	FPS	RS	FPS	RS	FPS	RS	FPS	RS	FPS	RS	FPS
Plot area (m 2 )	300.32 $\pm$ 50.38	315.15 $\pm$ 52.51	301.74 $\pm$ 53.89	313.33 $\pm$ 55.40	279.73 $\pm$ 44.02	288.20 $\pm$ 44.99	339.92 $\pm$ 69.24	353.64 $\pm$ 75.94	277.85 $\pm$ 38.52	284.60 $\pm$ 39.26	335.20 $\pm$ 49.72	341.85 $\pm$ 42.10
Point density (m -2 )	7.03 $\pm$ 1.28	6.69 $\pm$ 1.21	7.05 $\pm$ 1.53	6.79 $\pm$ 1.48	7.53 $\pm$ 1.37	7.30 $\pm$ 1.32	6.45 $\pm$ 2.35	6.23 $\pm$ 2.30	7.50 $\pm$ 0.97	7.32 $\pm$ 0.94	6.33 $\pm$ 1.58	6.12 $\pm$ 1.07
Point density (m -3 )	0.59 $\pm$ 0.14	0.45 $\pm$ 0.08	0.45 $\pm$ 0.12	0.42 $\pm$ 0.12	0.50 $\pm$ 0.15	0.47 $\pm$ 0.14	0.46 $\pm$ 0.30	0.43 $\pm$ 0.28	0.50 $\pm$ 0.13	0.48 $\pm$ 0.12	0.41 $\pm$ 0.17	0.38 $\pm$ 0.13
Average point spacing (cm)	0.33 $\pm$ 0.03	0.71 $\pm$ 0.06	0.26 $\pm$ 0.04	0.59 $\pm$ 0.10	0.28 $\pm$ 0.04	0.60 $\pm$ 0.10	0.32 $\pm$ 0.05	0.72 $\pm$ 0.12	0.32 $\pm$ 0.04	0.71 $\pm$ 0.10	0.41 $\pm$ 0.08	1 $\pm$ 0.19

Table 4: Comparison of downsampled dataset characteristics using random sampling (RS) and farthest point sampling (FPS) for synthetic and real point cloud datasets. Dataset ID: MF = Melville Forest, HP = Hensley Park, 1995.1 = area no. 1 in the 1995 planatation year

Panel of Figure 5 — Figure 5: Learning curves of the deep learning models trained on synthetic point cloud data downsampled using the farthest point sampling method: (a) PointNet, (b) PointNet++, (c) DGCNN, and (d) PointConv.

Model	Random Sampling		Farthest Point Sampling
	Validation Error (m 3 )	MAPE (%)	Validation Error (m 3 )	MAPE (%)
PointNet	1.034 $\pm$ 0.019	8.999 $\pm$ 0.189	0.959 $\pm$ 0.029	8.111 $\pm$ 0.204
PointNet++	0.151 $\pm$ 0.016	1.438 $\pm$ 0.178	0.190 $\pm$ 0.010	1.797 $\pm$ 0.149
DGCNN	0.195 $\pm$ 0.010	1.627 $\pm$ 0.060	0.202 $\pm$ 0.015	1.691 $\pm$ 0.146
PointConv	0.202 $\pm$ 0.019	1.705 $\pm$ 0.138	0.286 $\pm$ 0.040	2.307 $\pm$ 0.254

Table 5: Mean

\pm

STD of validation error and MAPE for PointNet, PointNet++, DGCNN, and PointConv trained on synthetic datasets, based on 5-fold cross-validation

Model	MF1995.1		MF1999.2		MF2007.2		HP2010.2		Knewleave20
	AGB (t/ha)	C (t/ha)	AGB (t/ha)	C (t/ha)	AGB (t/ha)	C (t/ha)	AGB (t/ha)	C (t/ha)	AGB (t/ha)	C (t/ha)
PointNet	215 $\pm$ 20	107 $\pm$ 10	279 $\pm$ 30	139 $\pm$ 15	195 $\pm$ 13	97 $\pm$ 7	277 $\pm$ 34	139 $\pm$ 17	206 $\pm$ 27	103 $\pm$ 14
PointNet++	200 $\pm$ 2	100 $\pm$ 1	237 $\pm$ 2	119 $\pm$ 1	181 $\pm$ 1	91 $\pm$ 1	225 $\pm$ 4	113 $\pm$ 2	205 $\pm$ 7	103 $\pm$ 3
DGCNN	154 $\pm$ 12	77 $\pm$ 6	200 $\pm$ 12	100 $\pm$ 6	149 $\pm$ 8	75 $\pm$ 4	205 $\pm$ 15	103 $\pm$ 7	170 $\pm$ 11	85 $\pm$ 5
PointConv	186 $\pm$ 26	93 $\pm$ 13	201 $\pm$ 17	100 $\pm$ 9	150 $\pm$ 18	75 $\pm$ 9	188 $\pm$ 28	94 $\pm$ 14	222 $\pm$ 89	111 $\pm$ 44

Table 6: AGB and carbon (t/ha) from real point clouds downsampled with fathest point sampling for Jigsaw Farm (MF1995.1, MF1999.2, MF2007.2, HP2010.2) and Knewleaves Farm (Knewleave20)

Model	MF1995.1		MF1999.2		MF2007.2		HP2010.2		Knewleave20
	AGB (t/ha)	C (t/ha)	AGB (t/ha)	C (t/ha)	AGB (t/ha)	C (t/ha)	AGB (t/ha)	C (t/ha)	AGB (t/ha)	C (t/ha)
PointNet	130 $\pm$ 15	65 $\pm$ 7	193 $\pm$ 33	97 $\pm$ 16	125 $\pm$ 19	63 $\pm$ 10	179 $\pm$ 37	89 $\pm$ 19	114 $\pm$ 5	57 $\pm$ 3
PointNet++	135 $\pm$ 2	67 $\pm$ 1	188 $\pm$ 3	94 $\pm$ 2	119 $\pm$ 1	60 $\pm$ 1	170 $\pm$ 4	85 $\pm$ 2	113 $\pm$ 2	57 $\pm$ 1
DGCNN	125 $\pm$ 5	63 $\pm$ 3	170 $\pm$ 9	85 $\pm$ 4	112 $\pm$ 5	56 $\pm$ 3	158 $\pm$ 8	79 $\pm$ 4	104 $\pm$ 4	52 $\pm$ 2
PointConv	130 $\pm$ 7	65 $\pm$ 3	181 $\pm$ 8	91 $\pm$ 4	117 $\pm$ 5	58 $\pm$ 2	164 $\pm$ 11	82 $\pm$ 6	115 $\pm$ 4	58 $\pm$ 2

Table 7: AGB and carbon (t/ha) from real point clouds downsampled with random sampling for Jigsaw Farm (MF1995.1, MF1999.2, MF2007.2, HP2010.2) and Knewleaves Farm (Knewleave20)

Figure 6: Distribution of wood volume across synthetic plots estimated by the four models trained on synthetic data downsampled using random sampling and farthest point sampling methods.

Figure 7: The distribution of wood volume across real point cloud dataset in the MF1995.1, as estimated by the four models trained on synthetic data downsampled using random sampling and farthest point sampling methods.

Panel of Figure 8 — Figure 8: Spatial distribution of the predicted AGB across plots based on real point cloud data derived from PointNet++ predictions trained on synthetic data downsampled with farthest point sampling: (a) MF1995.1, (b) MF1999.2, (c) MF2007.2, (d) HP2010.2, and (e) Knewleave20.

( Khoshelham et al., 2023 )	79	40	70	35	50	25	55	28	67	34
Method	MF1995.1		MF1999.2		MF2007.2		HP2010.2		Knewleave20
	AGB	C	AGB	C	AGB	C	AGB	C	AGB	C
	(t/ha)	(t/ha)	(t/ha)	(t/ha)	(t/ha)	(t/ha)	(t/ha)	(t/ha)	(t/ha)	(t/ha)
CHM-segmentation	79	40	70	35	50	25	55	28	67	34
TreeLearn	143	72	132	66	129	65	121	61	162	81
( Henrich et al., 2024 )	143	72	132	66	129	65	121	61	162	81

Table 8: Comparison of AGB and carbon (t/ha) estimates obtained by our direct method (PointNet++ with farthest point sampling datasets (FPS)), traditional methods, and field measurements for Jigsaw Farm (MF1995.1, MF1999.2, MF2007.2, HP2010.2) and Knewleave Farm (Knewleave20)

Forest Aboveground Biomass Estimation

See all 11 papers

Self-Supervised Tree-level Biomass Estimation in Urban Environments From Airborne LiDAR and Optical Observations
Jun 24, 2026Jose Bermudez, Zilong Zhong, Dominic Cyr +2
McMaster University, Hamilton, Ontario, Canada · McMaster University, 1280 Main Street West, Hamilton, Ontario L8S 4K1, Canada · Environment and Climate Change Canada, Montreal, Quebec, Canada +2
Transferable Above-Ground Biomass (AGB) Estimation Model from Multi-Sensor Data with Sparse Field Calibration
Aug 12, 2026Pann Thinzar Seint, Bryan Atwood, Subas ChhatkuliForest Aboveground Biomass Estimation Remote Sensing
DAI Labs, K.K.
Rapid Forest Fuel Load Estimation via Virtual Remote Sensing and Metric-Scale Feed-Forward 3D Reconstruction
May 11, 2026Quanyun Wu, Kyle Gao, Wentao Sun +6Tree Species Classification Remote Sensing
University of Waterloo · Waterloo, Canada · University of Calgary +1

Jun 24, 2026J/K move · Enter open · S save

Self-Supervised Tree-level Biomass Estimation in Urban Environments From Airborne LiDAR and Optical Observations

Jose Bermudez, Zilong Zhong, Dominic Cyr, Camile Sothe+1

McMaster University, Hamilton, Ontario, Canada · McMaster University, 1280 Main Street West, Hamilton, Ontario L8S 4K1, Canada · Environment and Climate Change Canada, Montreal, Quebec, Canada+2

Open paper AskPDF arXiv

Urban tree biomass remains less spatially explicitly quantified than biomass in managed forests because many estimates rely on inventories or coarse products that cannot resolve individual crowns or fine-scale heterogeneity. We present a crown-level above-ground biomass (AGB) framework for an 810~~km $^2$ landscape in Ontario, Canada, using leaf-off airborne LiDAR (8--10~~pulses~~m $^{-2}$ ) and near-infrared RGB orthophotography (0.16--0.20~~m) from 2018 and 2023. A dual-stream cross-attention network trained on rule-based pseudo-labels produced semantic marks for buildings, needleleaf trees, and deciduous trees, supporting crown delineation and functional-type assignment. On independently annotated withheld tiles, global/mean precision, recall, and Dice scores were 0.86, 0.83, and 0.84. Crowns were delineated with multiscale watershed segmentation in mapped tree areas, and AGB was estimated from a crown area--height power-law proxy calibrated to species-specific allometry (Lambert et al., 2005) for 21,921 inventory trees. For 18,713 inventory--segment matched pairs from a 90,726-tree held-out test set, AGB prediction achieved

R^2=0.609

using inventory crown geometry and

R^2=0.570

under operational segmentation, identifying crown delineation as the remaining uncertainty source. Aggregated to 30~~m, estimates yielded total AGB stocks of 1.73~~Tg in 2018 and 1.81~~Tg in 2023 (811--850~~Gg~~C), local densities up to ${\sim}140$ Mgha $^{-1}$ along the Niagara Escarpment, and a net carbon gain of 39~~Gg~C over five years. Deep-ensemble uncertainty maps highlighted high-epistemic-uncertainty areas linked to underrepresented land covers and guided assignment of uncertain crowns to a pooled allometric equation. The framework uses standard provincial data, requires no manual annotation, and produces a public bitemporal crown-level AGB database for trees outside forests at management-relevant resolution.

Abstract

Forest Aboveground Biomass Estimation

Explore similar work

Self-Supervised Tree-level Biomass Estimation in Urban Environments From Airborne LiDAR and Optical Observations

Transferable Above-Ground Biomass (AGB) Estimation Model from Multi-Sensor Data with Sparse Field Calibration

Rapid Forest Fuel Load Estimation via Virtual Remote Sensing and Metric-Scale Feed-Forward 3D Reconstruction