Estimating Forest Aboveground Biomass at the Stand Scale Using Voxel-Based 3D Canopy Structures from Airborne LiDAR

Abstract

Accurate estimation of forest aboveground biomass (AGB) is pivotal for assessing forest carbon sequestration and informing global change studies. Conventional LiDAR-based AGB estimation approaches primarily rely on height and density metrics, which inadequately characterize the complex three-dimensional (3D) structure of forest canopies. This study developed and evaluated a novel method utilizing voxel-based 3D canopy structural metrics derived from airborne LiDAR (ALS) to improve AGB estimation accuracy across diverse forest types. First, voxel-based metrics (Voxel Canopy Height Model (VCHM), canopy volume, and canopy surface area) were extracted from voxelized point clouds. Their distribution patterns across five forest types (Pinus massoniana, Cunninghamia lanceolata, coniferous, broadleaf, and mixed conifer–broadleaf forests) and their correlations with AGB were systematically examined. The results revealed distinct 3D canopy architectures among forest types, with all three voxel metrics showing highly significant positive correlations with AGB; VCHM demonstrated the strongest association. We then constructed two Random Forest models: a baseline model using traditional metrics only, and an enhanced model integrating both traditional and voxel-based metrics. The 10-fold cross-validation indicated that the model incorporating voxel metrics achieved markedly higher accuracy (R² in 0.490–0.684) than the traditional model (R² in 0.480–0.607), representing a relative improvement of 2.1% to 32.7%. The most substantial gain occurred in structurally complex broadleaf forests. The enhanced model was subsequently applied to generate a wall-to-wall AGB map of the study region, yielding a total estimated AGB stock of 8.36 × 10⁶ t, which exhibited a patchy spatial distribution. Pinus massoniana forests accounted for the largest proportion (57.8%) of the total stock. This study demonstrates that voxel-based 3D canopy metrics can more effectively capture forest structural heterogeneity and substantially improve the accuracy of AGB estimation models, particularly for complex forest stands. The findings provide a significant advancement toward precise, stand-scale forest biomass monitoring founded on detailed 3D structural information.

1. Introduction

Forests are the principal component of terrestrial ecosystems [1,2]. Forest aboveground biomass (AGB) is a core parameter for characterizing forest productivity, assessing carbon sink function, and understanding the global carbon cycle [3,4,5]. Accurate and efficient estimation of forest AGB at regional to global scales is of critical scientific and practical importance for addressing climate change, implementing sustainable forest management, and fulfilling relevant international conventions [6,7,8].

In recent years, active remote sensing technologies, particularly ALS, have become pivotal in addressing the limitations of traditional optical remote sensing in AGB estimation, due to their ability to directly and accurately capture forest vertical structural information [9,10,11]. However, most current LiDAR-based AGB estimation methods largely rely on statistical traditional metrics such as height and density metrics [12,13]. These metrics are well-suited for capturing the vertical profile or statistical properties of the canopy, but they fall short of describing its intricate internal three-dimensional architecture. Such dimensionality constraints may impair model performance in terms of both accuracy and generalizability, particularly when applied to structurally complex forests [14,15,16].

In recent years, voxel-based LiDAR metrics have gradually emerged as an important class of structural characterization methods [17]. This approach first partitions the LiDAR point cloud into a 3D grid of uniform volumetric units. Compared to traditional statistical metrics based on height or intensity distributions, voxel-based metrics possess the potential to represent canopy information layer by layer in the vertical dimension, enabling finer characterization of the spatial distribution patterns of LiDAR returns [18,19,20,21]. This advantage is particularly pronounced in stands with strong vertical structural heterogeneity. However, although the voxelization method offers a new perspective for 3D forest characterization, its application in the field of airborne LiDAR remains relatively limited [22]. Current studies have predominantly focused on extracting statistical indicators—such as height percentiles and point cloud density—using voxels, while paying insufficient attention to 3D canopy attributes with clear physical meanings [20,23]. For instance, 3D structural parameters directly derived from voxel grids, such as canopy height distribution, total volume occupied by solid canopy material, and outer surface area of the canopy, have not been fully explored or applied in voxel-based analyses. Therefore, this study introduces a voxel-based 3D reconstruction and analysis method. By voxelizing the LiDAR point cloud, the aim is to directly quantify the aforementioned canopy structural parameters in 3D space, thereby deepening the understanding of the real spatial configuration of forest canopies and further improving the estimation accuracy of aboveground forest biomass. The metrics extracted by this method directly reflect the physical configuration of the canopy in 3D space. Theoretically, they can more mechanistically characterize tree spatial competition and light resource allocation processes, thereby establishing a closer association with AGB accumulation [24,25,26,27].

Therefore, this study aims to systematically explore the potential and value of 3D canopy structural metrics extracted from ALS voxelization technology for forest AGB estimation. The research will focus on the following core questions: (1) How are these 3D metrics correlated with forest AGB, and is their estimation performance universal across different forest types? (2) Can the integration of voxel-based 3D metrics into modern machine learning estimation models improve the accuracy of AGB estimation for multiple forest types, and if so, to what extent? (3) What is the total regional forest AGB stock and its distribution patterns among different stand types derived from the inversion of the optimized model? Therefore, this study seeks to develop and validate a novel, accurate method for estimating forest biomass by leveraging 3D structural information. It endeavors to promote a paradigm shift in LiDAR-based forest remote sensing from 2D statistics to 3D analysis. The developed “plot–3D metrics–regional inversion” technical framework is expected to provide key technical support for the current transition from forest inventory to integrated forest–grassland monitoring, addressing the operational need to translate sample-based survey data into actionable information at the regional management scale.

2. Materials and Methods

2.1. Materials

2.1.1. Study Area

Dongyang City in central Zhejiang Province was selected as the study area for this work (Figure 1a). It is geographically located between 28°58′08″ N and 29°29′55″ N, 120°04′17″ E and 120°44′03″ E. Its maximum east–west extent is 71.6 km, and its maximum north–south extent is 56.1 km, with a total area of 1746.81 km². The terrain generally exhibits a higher elevation in the east and a lower elevation in the west, with altitudes ranging from 100 m to 1000 m, predominantly consisting of hilly and low mountainous landscapes. The area experiences a subtropical monsoon climate, with an average monthly precipitation of approximately 123–154 mm. The forest coverage rate is approximately 59.3%, and the vegetation is primarily dominated by Cunninghamia lanceolata forest, with the distribution of the main tree species illustrated in Figure 1b. The species composition and stand boundaries were derived from the 2024 Dongyang City Forest–Grassland–Wetland Survey dataset.

2.1.2. Field Inventory Data

This study was based on the 2024 forest–grassland–wetland inventory data of Dongyang City, Zhejiang Province. Supported by this dataset and with comprehensive consideration of key factors such as forest type, stand volume, and spatial distribution, a systematic grid of 25 m × 25 m square sample plots was established throughout the study area. The spacing between sample plots is 4*6 km. The four corner points (southwest, northwest, northeast, southeast) of each plot were precisely positioned using RTK equipment. The deployment of all 464 plots was completed in July 2024. Ultimately, five representative stand types within the region were selected as the research subjects: Masson pine forest, Chinese fir forest, coniferous forest, broadleaf forest, and mixed coniferous–broadleaf forest (

Table 1. Distribution of aboveground biomass across forest types.

Type	N	Minimum	Maximum	Range	SD	Average
All Data Combined	349	0.82	261.447	260.627	50.667	109.309
Pinus massoniana	67	8.172	173.03	164.858	41.985	80.702
Cunninghamia lanceolata	60	13.339	251.966	238.627	48.783	128.629
Broadleaf	102	0.82	261.447	260.627	57.413	117.578
Coniferous	43	41.084	171.854	130.77	33.173	104.818
Mixed Conifer–Broadleaf	77	7.955	221.482	213.526	47.607	110.7

).

Plot surveys followed standardized national forest inventory protocols [28]. Within each sample plot, full callipering was conducted for all trees with a diameter at breast height (DBH) ≥ 5 cm. The species, DBH, and tree height of each individual tree were measured and recorded. Individual tree biomass was calculated by applying species-specific tree height models, crown length models, and biomass equations. The stand-level biomass for each plot was then derived by summing the biomass of all its constituent trees. The foundational biomass models used in these calculations are presented in

Table 2. Crown length models and metrics for individual tree species.

Type	Model	c0	c1	c2
Pine	$L=c_0{H}^{c_1}$	0.908155	0.741309
Fir	$L=c_0{D}^{c_1}{H}^{c_2}$	0.486967	0.170689	0.897133
Hardwood Broadleaf	$L=c_0{H}^{c_1}{e}^{c_{2}H}$	0.631614	1.180106	−0.05109
Softwood Broadleaf	$L=c_0{H}^{c_1}{e}^{c_{2}H}$	0.441339	1.377011	−0.06033

and

Table 3. Bivariate biomass models.

Type	Component Type	Model	a	b	c
Pine	Stem	$W_2=a{H}^b{D}^c$	0.06	0.79342	1.800545
	Crown	$W_3=a{D}^b{L}^c$	0.137708	1.487266	0.405207
	Root	$W_4=a{H}^b{D}^c$	0.0417	−0.078	2.261759
Fir	Stem	$W_2=a{H}^b{D}^c$	0.064741	0.89585	1.487953
	Crown	$W_3=a{D}^b{L}^c$	0.0971	1.781359	0.0346
	Root	$W_4=a{H}^b{D}^c$	0.0617	−0.10374	2.115252
Hardwood Broadleaf	Stem	$W_2=a{H}^b{D}^c$	0.055984	0.809933	1.813982
	Crown	$W_3=a{D}^b{L}^c$	0.098	1.64808	0.461
	Root	$W_4=a{H}^b{D}^c$	0.0549	0.106846	2.095267
Softwood Broadleaf	Stem	$W_2=a{H}^b{D}^c$	0.0444	0.719686	1.7095
	Crown	$W_3=a{D}^b{L}^c$	0.0856	1.22657	0.397
	Root	$W_4=a{H}^b{D}^c$	0.0459	0.106662	2.024681

.

2.1.3. ALS Data Acquisition

From June to July 2024, ALS data were acquired over the Dongyang study area using a Cessna 208 fixed-wing manned aircraft equipped with a Leica TerrainMapper-1 laser scanning system. The mission utilized a GNSS/IMU unit for position and attitude measurement. The total coverage of the LiDAR survey was approximately 1205 km², with flight line spacing designed to ensure a side overlap of no less than 20%. The average point density exceeded 20 pts/m², the flight altitude was 2000 m above mean sea level, and the total flight line length was 1280 km. The survey area was divided into 1205 mapping frames (each 1000 m × 1000 m). Key flight and sensor parameters are summarized in

Table 4. Key parameters of the airborne LiDAR survey.

Parameter Type	Design Parameter
Flight speed (km/h)	240
Relative flight height (m)	2200
Pulse repetition frequency (kHz)	1250
Field of view (degrees)	40
Scanning frequency (Hz)	150
Swath width (m)	713–1546
Minimum point density per flight line (pts/m2)	7.2
Average point density (pts/m2)	11
Side overlap	20%

.

2.2. Methods

2.2.1. Extraction of ALS Metrics

The ALS point clouds were preprocessed using LiDAR360 software. The raw point cloud data first underwent strip adjustment and noise filtering to correct for systematic discrepancies between flight lines and remove outlier points. We then classified ground returns using the iterative progressive triangular densification (IPTD) algorithm. From these ground points, a 1 m resolution digital elevation model (DEM) was generated through irregular triangular network (TIN) interpolation. Then, the point cloud data were normalized to remove the influence of terrain elevation. Finally, points above a height threshold of 2 m were extracted as canopy points, based on which height percentile metrics and density variables were calculated (

Table 5. LiDAR basic metrics.

Metric Category	Metric Name	Description
Height-related metrics	H_5, H_10, H_20, H_25, H_30, H_40, H_50, H_60, H_70, H_80, H_90, H_95, H_99	Percentile heights (5th, 10th, 20th, 25th, 30th, 40th, 50th, 60th, 70th, 80th, 90th, 95th, 99th) of the canopy height distribution of first returns
	H_mean	Mean height above ground of all first returns
	H_max	Maximum height above ground of all first returns
	H_median	Median height above ground of all first returns
	H_iq	Interquartile range of heights of all first returns
	H_sq	Root mean square of heights of all first returns
	H_kurtosis	Kurtosis of the height distribution of all first returns
	H_cv	Coefficient of variation in heights of all first returns
	H_variance	Variance of heights of all first returns
Density-related metrics	D1, D3, D5, D7, D9	Canopy return density: Proportion of points above the 10th, 30th, 50th, 70th, and 90th height percentiles to total number of first returns

).

2.2.2. Extraction of Voxel-Based Metrics

To further characterize the 3D structural features of the forest canopy, this study introduced a voxelization method to spatially discretize the normalized LiDAR point cloud and extract canopy structural metrics at the voxel scale. Voxelization is the process of partitioning 3D point cloud space into regular cubic units, which can effectively describe both vertical and horizontal heterogeneity within the canopy. First, the boundaries of the 3D grid were determined based on the spatial extent of the point cloud. Considering both point density and canopy structural scale, the voxel size was set to 1 m × 1 m × 1 m. This resolution ensures perfect alignment with the 25 m × 25 m plot grids, avoiding edge artifacts, and follows the common practice in voxel-based forest studies to maintain comparability with the existing literature [22].

(a)

Voxel Canopy Height Model (VCHM)

VCHM refers to the average height of the highest non-empty voxel within each voxel column. Specifically, for each horizontal grid cell (i.e., a voxel column), the height of the highest voxel containing point clouds in the vertical direction is extracted, representing the canopy surface height for that column. The average voxel canopy height for the study area is then calculated as the mean of these highest voxel heights across all columns. The calculation formula is as follows:

$$VCHM={\displaystyle \frac{1}{N^2}}\sum _{i,j=1}^{N}max\left\{Z_{i,j}\right\}$$

where the study area is divided into $N\times N$ regular grid cells (voxel columns) in the horizontal plane, with each cell denoted by the two-dimensional index $\left(i,j\right)$; $Z_{i,j}$ is the set of normalized heights of all LiDAR point clouds within the voxel $\left(i,j\right)$; $max\left\{Z_{i,j}\right\}$ represents the maximum height of the point clouds in that column, i.e., the canopy surface height at that location; ${\sum }_{i,j=1}^N\left(\right)$ denotes the summation over all grid cells; and dividing by $N^2$ yields the average canopy surface height across the entire study area.

(b)

Canopy Volume

Canopy volume refers to the sum of the volumes of all voxels containing canopy point clouds, representing the total 3D space occupied by the canopy. A height threshold (2 m in this study) was first applied to distinguish canopy points from non-canopy points, and only voxels containing canopy points were counted towards the canopy volume. The calculation formula is as follows:

$$Volume=\sum _{j=1}^m{V}_{voxel}\ast I_j$$

where $V_{voxel}$ is the volume of a single voxel (1 m³ in this study), $I_j$ is an indicator function that takes the value of 1 if the j-th voxel contains canopy points and 0 otherwise, and m is the total number of voxels.

(c)

Canopy Surface Area

Canopy surface area is defined as the sum of all exposed voxel faces at the canopy’s exterior, and thus serves as a metric reflecting its structural complexity. It is calculated by extracting the surface voxels (i.e., voxels that contain canopy points and have at least one adjacent voxel that is empty) and summing the area of their exposed external faces. Surface area is calculated by summing the exposed faces of all surface voxels. For each such voxel, we count only those sides that are adjacent to empty space or non-canopy voxels, with each exposed face contributing 1 m² to the total. This measure effectively quantifies the canopy’s outer roughness and gap structure, properties that directly influence key ecological functions like light capture and gas exchange. In this study, the canopy surface area was calculated from the voxelized point cloud data using the adjacent voxel counting method.

$$AREA=(N\times 6a)-\sum _{i=1}^6(N_i\times a\times i)$$

where the voxel edge length is $l$ and the area of a single voxel face is $a=l^2$. Let the total number of voxels containing canopy points in the study area be $N$. For any canopy voxel, the number of faces it shares with adjacent canopy voxels in the 3D grid (i.e., the number of adjacent canopy voxels) is denoted as k $\left(k=\mathrm{0,1},\dots ,6\right)$. Let $N_k$ represent the number of voxels that have exactly k adjacent canopy voxels, satisfying ${\sum }_{k=0}^6{N}_k=N$.

2.2.3. Random Forest Model

The Random Forest algorithm, developed within the ensemble learning framework by Breiman and Cutler, is essentially an extension and ensemble form of decision tree models [29,30]. The method employs bootstrap sampling with replacement, drawing multiple subsets from the original dataset. Each subset is used to train an individual decision tree. Predictions are then made by aggregating the outputs of all trees, for example, through averaging or majority voting. This ensemble approach significantly improves the model’s overall predictive accuracy and stability.

In this study, the Random Forest model was implemented using the randomForest package in R. Three key parameters were set during model construction: ntree (number of decision trees), nodesize (minimum number of samples in a terminal node, with a default of 5), and mtry (number of variables randomly sampled as candidates at each split, defaulting to one-third of the total number of predictors). In this study, ntree was set to 2000, while the other parameters retained their default package values.

2.2.4. Model Accuracy Assessment Using 10-Fold Cross-Validation

To evaluate the generalization performance of the model and mitigate the risk of overfitting, a Random Forest algorithm was implemented for predictive modeling, coupled with a rigorous 10-fold cross-validation procedure. Specifically, the complete dataset was randomly divided into 10 equally sized subsets. In each iteration, nine subsets were used to train the model, while the remaining subset served as an independent validation set. This process was repeated 10 times so that each subset was used exactly once for validation. The final model performance metrics were calculated as the average of the results obtained from all 10 validation folds. This approach ensures that the model evaluation is not biased by a particular data partition, thereby yielding a robust estimate of its predictive performance on unseen data [31,32,33].

We evaluated performance with the coefficient of determination (R²), which indicates the goodness of fit, and the root mean square error (RMSE), a measure of prediction error. Here, a model is considered better fitted when R² is higher, while a lower RMSE points to more accurate predictions. Model accuracy was compared in this way to select the optimal biomass estimation model for each dominant tree species.

$$R^2=1-{\displaystyle \frac{{\sum }_{i=1}^n{(y_i-\widehat{y_i})}^2}{{\sum }_{i=1}^n(y_i-\overline{y_i}{)}^2}}$$

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}\sum _{i=1}^n{(y_i-\widehat{y_i})}^2}$$

where $y_i$ is the actual observed value, $\widehat{y_i}$ is the model-predicted value, $\overline{y_i}$ is the mean of the actual observed values, and n is the number of validation samples.

3. Results

3.1. Distributions of Voxel-Based Metrics

Based on voxelized ALS data, this study extracted three structural parameters: the VCHM, canopy volume, and canopy surface area. Descriptive statistics of these metrics are presented across different forest types (Figure 2 and

Table 6. Variance of voxel-based metrics across different forest types.

Forest Type	VCHM	Canopy Volume	Surface Area
All tree species	14.42	1.24 × 106	9.40 × 106
Pinus massoniana	14.18	1.03 × 106	8.13 × 106
Cunninghamia lanceolata	11.01	1.66 × 106	1.13 × 107
Broadleaf	17.11	1.04 × 106	8.63 × 106
Mixed conifer–broadleaf	14.97	1.23 × 106	9.48 × 106
Pinus massoniana	14.18	1.03 × 106	8.13 × 106

). Coniferous forests (e.g., Cunninghamia lanceolata) exhibited a relatively uniform canopy structure in terms of VCHM, with a median of 11.01 m, an interquartile range of 6.52 m, and a variance of 28.34. Broadleaved forests showed a higher median VCHM (17.07 m), a larger interquartile range (11.53 m), and a variance of 68.24, reflecting greater variability in canopy height. Mixed forests displayed intermediate VCHM metrics (median 14.97 m, interquartile range 10.24 m, variance 55.36), with a wider distribution that may suggest more complex structural characteristics. Notably, Pinus massoniana forests demonstrated a considerable capacity for spatial occupancy, with a median canopy volume of 1.03 × 10⁵ m³ (variance 1.56 × 10¹⁰ and a median canopy surface area of 8.18 × 10⁴ m² (variance 4.35 × 10⁸). In comparison, coniferous forests (exemplified by Cunninghamia lanceolata) had a larger median canopy volume (1.66 × 10⁵ m³) but a relatively smaller median canopy surface area (1.13 × 10⁵ m²), indicating a distinct spatial configuration strategy.

3.2. Correlation Analysis Between Voxel-Based Metrics and AGB

All voxel-extracted metrics, VCHM, canopy volume, and canopy surface area, showed highly significant positive correlations with AGB across all forest types, demonstrating strong potential as predictors for AGB estimation (

Table 7. Relationship between voxel-based 3D canopy structures and AGB.

Forest Type	VCHM		Canopy Volume		Surface Area
	r	p	r	p	r	p
All tree species	0.672	3.6 × 10−47	0.462	7.9 × 10−20	0.525	3.9 × 10−26
Pinus massoniana	0.836	1.5 × 10−18	0.598	9.0 × 10−8	0.701	3.9 × 10−11
Cunninghamia lanceolata	0.822	7.7 × 10−16	0.615	1.7 × 10−7	0.657	1.2 × 10−8
Broadleaf	0.530	1.0 × 10−8	0.361	2.0 × 10−47	0.417	1.3 × 10−47
Coniferous	0.643	3.4 × 10−6	0.364	1.6e × 10−3	0.438	3.3 × 10−4
Mixed conifer–broadleaf	0.651	1.4 × 10−10	0.443	5.5 × 10−5	0.501	3.5 × 10−6

) and comprehensively reflecting the close relationship between canopy structural characteristics and AGB. For the pooled sample of all tree species, the correlation coefficients for VCHM, canopy surface area, and canopy volume reached 0.672, 0.525, and 0.462, respectively, confirming the general relevance of voxel-based 3D metrics for biomass estimation. Analyzed by forest type, all three metrics maintained significant correlations within each type, with particularly strong associations in coniferous forests. In Pinus massoniana and Cunninghamia lanceolata forests, the correlation coefficients for VCHM were as high as 0.836 and 0.822, for canopy surface area 0.701 and 0.657, and for canopy volume 0.598 and 0.615, respectively. In broadleaved and mixed forests, all three metrics also reached significant levels, with VCHM showing the strongest correlation, followed by canopy surface area; canopy volume, although somewhat lower, remained significantly correlated.

3.3. Results of Random Forest Model

To assess the value of voxel-based 3D canopy metrics for estimating forest AGB, we developed and compared two Random Forest models. Model 1 used only traditional LiDAR metrics, while Model 2 combined these with the novel voxel-based metrics. Their performance, evaluated via 10-fold cross-validation (Figure 3), showed that Model 2 consistently and significantly outperformed Model 1 across all forest types. For the combined dataset, Model 2 achieved an R² of 0.593—an improvement of 0.089 (17.7%) over Model 1’s R² of 0.504. This demonstrates that integrating voxel-based 3D structural information effectively enhances AGB estimation accuracy. Analysis by individual forest type further confirmed the universally beneficial effect of voxel metrics. In Chinese fir forests, Model 2 achieved the highest R² of 0.689 among all types, which is an improvement of 0.091 (15.2%) over Model 1. For mixed coniferous–broadleaved forests, Model 2 attained an R² of 0.536, showing the most substantial relative increase of 0.132 (32.7%) compared to Model 1. This underscores the particularly significant gain provided by voxel metrics for biomass estimation in structurally complex mixed forests. It is noteworthy that, despite variations in sample size and structural characteristics among the different forest types, Model 2 consistently maintained a clear accuracy advantage over Model 1 in all cases.

To evaluate the value of voxel-based 3D canopy metrics for estimating forest AGB, we developed and compared two Random Forest models. Model 1 used only traditional LiDAR metrics, whereas Model 2 combined traditional metrics with the novel voxel-based metrics. Model performance was evaluated via 10-fold cross-validation. The results showed that Model 2 consistently and significantly outperformed Model 1 across all forest types (Figure 3). The 95% confidence intervals around the regression lines in Figure 3 visually represent the estimation uncertainty; the generally narrower confidence intervals for Model 2 indicate improved estimation precision. For the combined dataset, Model 2 achieved an R² of 0.593, an increase of 0.089 (17.7%) over the R² of 0.504 obtained by Model 1. This demonstrates that incorporating voxel-based 3D structural information effectively enhances the accuracy of AGB estimation. Further analysis by forest type confirmed the consistent benefit of voxel metrics. In Chinese fir forests, Model 2 attained the highest R² among all forest types (0.689), which represents an improvement of 0.091 (15.2%) over Model 1. For mixed coniferous–broadleaved forests, Model 2 yielded an R² of 0.536, corresponding to the largest relative increase of 0.132 (32.7%) compared with Model 1. Accordingly, the confidence intervals for Model 2 in mixed forests were markedly narrower than those for Model 1, underscoring the reduction in estimation uncertainty achieved by voxel metrics in structurally complex stands. It is noteworthy that despite variations in sample size and structural characteristics among forest types, Model 2 consistently maintained a clear advantage over Model 1 in both accuracy and precision, as systematically evidenced by the higher R² values and generally tighter confidence bands shown in Figure 3.

The variable importance analysis results from the Random Forest model (Figure 4) reveal, from different perspectives, the critical role of the proposed voxel-based 3D canopy structural parameters in improving the estimation accuracy of forest AGB. Specifically, in the comprehensive all-species model, the VCHM exhibited the highest feature importance (%IncMSE = 16%), significantly outperforming other traditional LiDAR metrics. The canopy surface area (14%) and canopy volume (10%) also ranked within the top five, highlighting the dominant role of 3D structural features in comprehensively explaining spatial variations in biomass. In the species-specific group analyses, different forest types demonstrated distinct feature importance patterns aligned with their structural characteristics: In the Cunninghamia lanceolate model, topographic factors (e.g., elev_percentile_1st, 10%) ranked highest, but canopy surface area and canopy volume (both 9%) followed closely, with VCHM (8%) also ranking within the top four, indicating that although topography played a dominant role, 3D structural parameters remained important synergistic predictors. In the Pinus massoniana model, VCHM (18%) emerged as the most important feature, corroborating the significant influence of vertical canopy structure on biomass in this species, followed by canopy surface area. In broadleaf forests, VCHM (13%) and canopy surface area (12%) ranked just below the topographic coefficient of variation (elev_cv_z, 15%), with their high importance reflecting the close relationship between canopy external morphological complexity and biomass in broadleaf forests. In mixed conifer–broadleaf forests, although the absolute values of VCHM (5%) and canopy volume (4%) were relatively low, they remained key predictors in the feature ranking for this group, demonstrating the persistent explanatory power of 3D parameters in mixed forest structures. Overall, these results indicate that the proposed 3D voxel-based structural parameters generally exhibit high feature importance across different forest types, enabling a more direct and physically meaningful representation of the relationship between canopy structure and AGB. This provides essential structural information for enhancing the accuracy and generalizability of AGB estimation models.

3.4. AGB Estimation Results

The inversion based on the optimal model revealed that the total AGB of forests in Dongyang City is 8.36 million t (

Table 8. AGB estimation results.

Type	Inversion Results (t)	Percentage (%)
All tree species	8.36 × 106	100.00
Pinus massoniana	4.83 × 106	57.73
Cunninghamia lanceolata	1.72 × 105	2.06
Broadleaf	1.18 × 106	14.03
Coniferous	1.38 × 105	1.65
Mixed conifer–broadleaf	2.04 × 106	24.25

). Spatially, areas with high biomass (≥198 t/ha) are predominantly distributed in patchy patterns across the central and northeastern regions, while low-biomass areas (0–72 t/ha) are widely scattered along the western and southern margins. In terms of species composition, Pinus massoniana forests contributed the most significantly, reaching 4.83 million t and accounting for 57.73% of the total (Figure 5). This was followed by mixed coniferous–broadleaved forests (2.04 million t, 24.25%) and broadleaved mixed forests (1.18 million t, 14.03%). Cunninghamia lanceolata forests and coniferous mixed forests had relatively low biomass stocks of 0.1723 million and 0.1384 million t, respectively, together accounting for only 3.71% of the total.

4. Discussion

4.1. Structural Characterization Advantages of Voxel-Based 3D Metrics

This study found that the three metrics extracted based on voxels can clearly distinguish the structural differences in canopy among various forest types (Section 3.1). Among them, VCHM, as a direct measure of vertical space, maintained the strongest correlation with AGB across all forest types (Table 6), especially in structurally more homogeneous coniferous pure forests, with r as high as 0.82 or above. This indicates that canopy height remains the most critical driving factor for biomass accumulation dominated by vertical structure [34,35]. However, in broadleaved forests and coniferous-broadleaved mixed forests with high structural heterogeneity, the correlation of canopy surface area (r = 0.417–0.501) was significantly superior to that of canopy volume (r = 0.361–0.443). This result may suggest that for stands with complex branching, high spatial heterogeneity in canopy structure, the surface area characterizing the external complexity of the canopy better reflects the photosynthetic area and the allocation strategy of supporting structure biomass compared to the volume characterizing internal filling, thereby establishing a closer link with AGB [36,37,38]. The voxelization method captures a more comprehensive set of structural features by quantifying these 3D attributes simultaneously, moving beyond traditional height and density metrics. This provides a new perspective for understanding the structure–function relationship in forests.

In terms of metric design and application context, this study contrasts with and advances previous research. For instance, Kim et al. [39] in their study of intact tropical rainforests, also employed a voxelization method. However, the metrics they extracted (e.g., number of points per layer, frequency ratio, median intensity) focused more on describing the statistical characteristics of point clouds within different vertical strata. In contrast, this study directly proposes three 3D structural parameters with clear physical meanings—VCHM, canopy volume, and canopy surface area—which more intuitively quantify the geometric configuration of the canopy in three-dimensional space (height distribution, occupied space, external complexity). Pearse et al. [22], in a study of a single-species Pinus radiata plantation, systematically compared the advantages of voxel-based metrics over traditional metrics [17]. These studies laid the foundation for the application of the voxel method in specific ecosystems. The prominent advancement of this study lies in the systematic application of these three-dimensional geometric metrics with clear physical meanings to a forest ecosystem encompassing five distinct structural types (including pure Pinus massoniana and Cunninghamia lanceolata forests, as well as structurally more complex broadleaf, coniferous mixed, and mixed conifer–broadleaf forests). The results (Table 6) indicate that especially in the most structurally complex natural mixed conifer–broadleaf forests, the explanatory power of the canopy surface area metric significantly surpassed that of the canopy volume metric. This new finding reveals that the advantage of voxel-based metrics lies not only in their three-dimensional representational capability but, more importantly, in their ability to capture key morphological attributes that are closely related to the structure–function relationships of specific ecosystems. This moves beyond the abstract statistical descriptions of traditional metrics and some early voxel-based analyses, providing a more mechanistic perspective for understanding the structure–function relationships in diverse forests.

4.2. Universal Improvement of AGB Estimation Models by Incorporating Voxel Metrics

Across all forest types examined, the models developed in this study that incorporated voxel-based metrics consistently surpassed the performance of those relying solely on traditional metrics. This enhancement in universality carries significant methodological implications. Notably, the most substantial improvement in model accuracy (a 32.7% increase in R²) was observed in structurally complex coniferous broadleaved mixed forests. When applied to mixed forests characterized by multiple species, diverse age classes, and distinct vertical stratification, the statistical features of traditional metrics are prone to homogenization, leading to diminished estimation capability [40,41]. In contrast, voxel-based metrics can directly characterize this mixture and stratification phenomenon in 3D space, thereby more effectively capturing the biomass signals corresponding to structural complexity [42,43,44]. This demonstrates the unique value of voxelized 3D metrics in addressing the challenging problem of AGB estimation in complex ecosystems. Even in Chinese fir forests with a relatively small sample size, the model achieved the highest R² (0.689), indicating that this method is also effective for relatively pure stands, showcasing its broad application potential. This confirms the robustness of the proposed method in tackling the challenges of estimating biomass for forests with complex structures at the stand level, providing a superior solution for routine monitoring of stand biomass using ALS.

4.3. Research Limitations and Prospects

Based on the validation of voxelized 3D metrics for AGB estimation in this study, several key limitations are acknowledged, which also point to valuable directions for future research. (1) The exploration of voxel-based metrics remains preliminary. This study primarily relied on three macroscopic, geometrically derived indicators: Voxel Canopy Height Model (VCHM), canopy volume, and canopy surface area. Future work could significantly benefit from extracting and integrating more refined morphological descriptors from the voxel grid. Promising candidates include internal point cloud density distributions, measures of voxel connectivity or adjacency, and detailed vertical profile functions. These advanced metrics could reveal subtler structural characteristics and potentially enhance the explanatory power of models for complex canopies. (2) The methodological framework is inherently dependent on the completeness of LiDAR data within each voxel. The calculation of structural elements, such as volume and surface area, presupposes the presence of lidar points. A significant challenge, not explicitly addressed here, is the handling of gaps or missing data within voxel cells. This issue is particularly pertinent in areas with low point density, dense understory vegetation, or steep terrain where lidar penetration is incomplete. In such cases, voxels may be erroneously classified as “empty,” leading to potential underestimations of canopy attributes. The sensitivity of the three core metrics to variations in point density and their robustness under heterogeneous data coverage conditions require further investigation to ensure reliable application across datasets of differing quality. (3) The geographical specificity of the study presents a constraint on generalizability. The analysis was confined to Dongyang City, China, a region characterized by a subtropical monsoon climate and associated forest compositions. Consequently, the derived model parameters and the observed relationships between the 3D metrics and AGB may not be directly transferable to other biomes (e.g., boreal or tropical dry forests) or regions with fundamentally different ecological and structural conditions. Validating and, if necessary, calibrating this framework across a broader range of geographical and ecological zones is essential to confirm its robustness and universal applicability. In summary, overcoming the aforementioned limitations will be a core task for future research in this field, and successful breakthroughs will significantly enhance the applicability and predictive capacity of voxel-based metrics in complex forest ecosystems.

5. Conclusions

Based on ALS point cloud voxelization, this study systematically investigated the application of extracted 3D canopy structural metrics for estimating forest AGB at the stand scale. The main conclusions are as follows:

The three metrics (VCHM, canopy volume, and canopy surface area) effectively quantified and distinguished differences in the 3D spatial configuration of canopies among different forest types. This provides a novel observational dimension for understanding forest structural heterogeneity. Compared to models using only traditional height and density metrics, the Random Forest model that integrated voxel-based 3D structural metrics consistently and significantly improved AGB estimation accuracy. This improvement was validated across all forest types and was most pronounced in structurally complex mixed coniferous–broadleaved forests. The results confirm that voxel-based metrics hold a unique advantage, making them particularly valuable for estimating biomass in complex, heterogeneous ecosystems. Applying the optimized model, the total forest AGB for the study area was estimated to be 8.36 × 10⁶ t. The inversion further revealed a patchy spatial distribution of high-biomass areas within Pinus massoniana stands. This high-resolution mapping of biomass heterogeneity addresses a critical need in the transition from traditional forest inventories to integrated forest–grassland monitoring. By quantifying the baseline forest AGB and its composition for Dongyang City, the results provide spatially explicit data to directly support forest carbon sink management and evidence-based ecological policy formulation.

In summary, this study demonstrates that analyzing LiDAR data through a true 3D voxel framework yields a more detailed and physically realistic depiction of canopy structure. This enhanced representation offers a robust basis for establishing reliable structure–biomass relationships, providing a methodological pathway for more precise monitoring and assessment of forest AGB at regional to global scales.