The precise quantification of forest biomass is a cornerstone for understanding carbon sequestration dynamics, biodiversity assessments, and sustainable forest resource management. Traditional ground-based forest inventory methods, while accurate, are notoriously labor-intensive, time-consuming, and often impractical for covering large or inaccessible areas. The advent of Unmanned Aerial Vehicle (UAV) remote sensing technology has ushered in a paradigm shift for forestry applications. UAV drones equipped with advanced sensors, particularly Light Detection and Ranging (LiDAR), offer an unparalleled ability to rapidly acquire high-density, three-dimensional point cloud data over forest canopies. This capability enables the detailed characterization of forest structure, paving the way for accurate, non-destructive estimation of critical parameters like tree height, diameter at breast height (DBH), crown width, and ultimately, above-ground biomass (AGB).
Among the various remote sensing platforms, UAV drones provide a unique balance of flexibility, high spatial resolution, and cost-effectiveness. UAV LiDAR systems can penetrate the canopy to varying degrees, capturing information from the top of the trees down to the understory and ground, which is crucial for constructing detailed Digital Terrain Models (DTM) and Digital Surface Models (DSM). The difference between these, the Canopy Height Model (CHM), is a fundamental data layer for individual tree detection and segmentation. However, accurately delineating individual tree crowns (ITCs) from the CHM, especially in dense or complex stands, remains a significant challenge that directly impacts the accuracy of subsequent biomass estimations.
This research focuses on developing and validating a robust methodological framework for individual tree AGB estimation by synergistically integrating two key algorithms applied to UAV LiDAR data: Canopy Contour Extraction (CCE) and Multi-scale Crown Width (MCW) estimation. The CCE algorithm is designed for precise individual tree segmentation from the CHM, while the MCW algorithm enhances the accuracy of crown dimension retrieval. We hypothesize that the combined use of CCE for segmentation and MCW for feature extraction will yield superior structural metrics, leading to more accurate allometric biomass models compared to using either method in isolation or relying on conventional approaches.
Methodological Framework: Integrating CCE and MCW with UAV LiDAR Data
The proposed workflow encompasses data acquisition, pre-processing, individual tree segmentation, feature extraction, and biomass model development and validation, all centered on data collected by UAV drones.
1. Study Area and UAV LiDAR Data Acquisition
The study was conducted in typical plantation forests, including Larch (Larix spp.) and Chinese Pine (Pinus tabuliformis) stands. A custom-built UAV drone platform was employed, equipped with a high-performance LiDAR scanner (e.g., Velodyne Puck Lite VLP-16). Flight missions were planned to ensure sufficient point cloud density (exceeding 200 points/m²) with significant overlap. Precise georeferencing was achieved using a Global Navigation Satellite System (GNSS) base station. Concurrently, a rigorous ground truth campaign was carried out, measuring DBH, tree height, and crown width for a large sample of trees. A subset of these trees were destructively sampled to develop site-specific allometric equations for AGB, serving as the validation benchmark.
2. LiDAR Data Pre-processing and Canopy Height Model Generation
The raw LiDAR point clouds were processed using standard software (e.g., TerraSolid, LAStools). The sequence involved noise filtering, classification of ground points using an iterative triangulation algorithm, and generation of a high-resolution DTM. A DSM was then created from the highest non-ground points within each raster cell. The crucial CHM was derived by subtracting the DTM from the DSM:
$$CHM(x, y) = DSM(x, y) – DTM(x, y)$$
where \(CHM(x, y)\) represents the height of the canopy above ground at coordinate \((x, y)\). This CHM serves as the primary input for the subsequent tree segmentation process.
3. Individual Tree Segmentation using the Canopy Contour Extraction (CCE) Algorithm
The CCE algorithm is designed to overcome limitations of simple local maxima detection, which often fails in complex canopies. The process involves several key steps applied to the CHM derived from UAV LiDAR data:
a) Local Maxima Detection with Adaptive Filtering: An adaptive circular window is used to identify potential tree apex locations (local maxima). The window size varies based on the surrounding CHM values to account for varying crown sizes.
b) Canopy Contour Delineation: From each identified apex, a region-growing or watershed segmentation algorithm is initiated. However, CCE incorporates an edge-detection enhancement. It uses a Sobel or Canny operator on the CHM to highlight crown boundaries more precisely before segmentation, helping to separate touching crowns. The growth or watershed flooding is constrained by these detected edges and a height threshold relative to the apex.
c) Contour Refinement: The initial segmented polygons are refined using morphological operations (opening, closing) to smooth irregularities and fill small gaps, resulting in the final individual tree crown (ITC) polygons.
The performance of CCE can be evaluated using precision, recall, and F1-score metrics when compared to manually delineated reference trees:
$$Precision = \frac{TP}{TP + FP}, \quad Recall = \frac{TP}{TP + FN}, \quad F1 = 2 \cdot \frac{Precision \cdot Recall}{Precision + Recall}$$
where \(TP\) (True Positives) are correctly segmented trees, \(FP\) (False Positives) are non-tree objects segmented as trees, and \(FN\) (False Negatives) are missed trees.
4. Crown Feature Extraction using the Multi-scale Crown Width (MCW) Algorithm
Accurate crown width is a vital predictor for biomass. The MCW algorithm improves upon single-direction measurement by analyzing the crown shape at multiple scales. For each ITC segmented by CCE:
a) Crown Rasterization: The ITC polygon is converted to a binary raster mask.
b) Multi-scale Analysis: The algorithm scans across the crown mask in multiple directions (e.g., 0°, 45°, 90°, 135°). For each direction, it calculates the crown width profile. It then applies a multi-scale mean filter to the profile to smooth out fine-scale irregularities caused by branches or gaps while preserving the overall crown boundary.
c) Crown Width Determination: The maximum value from the smoothed profiles across all analyzed directions is taken as the representative crown width (\(CW_{MCW}\)). This method is more robust than simply taking the maximum diameter of a fitted ellipse. Other features like crown projection area (\(CPA\)) and crown volume (\(CV\)) are also calculated from the 3D points within each ITC.
The tree height (\(H\)) for each ITC is directly extracted as the 95th or 98th percentile height of LiDAR points within the crown to minimize noise from outliers.
5. Biomass Model Development and Estimation
Allometric models were developed by establishing relationships between the ground-measured AGB (from destructive sampling) and LiDAR-derived metrics. Commonly used models include the power-law form:
$$AGB = a \cdot (D^{2}H)^{b}$$
or linear forms after log-transformation. In this study, we used LiDAR-derived equivalents: Crown Diameter (\(CD_{MCW}\)) and Tree Height (\(H_{LiDAR}\)). Multiple forms were tested:
$$AGB = \beta_0 + \beta_1 \cdot (CD_{MCW}^2 \cdot H_{LiDAR}) \quad \text{(Linear)}$$
$$\ln(AGB) = \alpha_0 + \alpha_1 \cdot \ln(CD_{MCW}^2 \cdot H_{LiDAR}) \quad \text{(Log-Log)}$$
The models were fitted using ordinary least squares regression. The performance of models using features from the integrated CCE-MCW method was compared against models using features from a standard watershed segmentation (WS) and local maxima (LM) method, as well as against the benchmark K nearest neighbors (KNN) and random forest (RF) imputation approaches applied to the UAV LiDAR data.
Model accuracy was assessed using the coefficient of determination (\(R^2\)), Root Mean Square Error (RMSE), and relative RMSE (rRMSE):
$$RMSE = \sqrt{\frac{1}{n}\sum_{i=1}^{n}(y_i – \hat{y}_i)^2}, \quad rRMSE = \frac{RMSE}{\bar{y}} \times 100\%$$
where \(y_i\) is the observed AGB, \(\hat{y}_i\) is the predicted AGB, \(\bar{y}\) is the mean observed AGB, and \(n\) is the number of samples.
Experimental Results and Analysis
1. Accuracy of Individual Tree Detection and Segmentation
The integration of CCE and MCW algorithms with data from UAV drones significantly improved individual tree detection rates across different forest types. The CCE algorithm, with its edge-aware segmentation, demonstrated superior performance in dense stands where crowns were overlapping.
| Forest Type | Method | Precision | Recall | F1-Score |
|---|---|---|---|---|
| Larch Plantation | CCE-MCW | 0.92 | 0.88 | 0.90 |
| WS-LM | 0.85 | 0.82 | 0.83 | |
| Pine Plantation | CCE-MCW | 0.89 | 0.85 | 0.87 |
| WS-LM | 0.81 | 0.78 | 0.79 |
The table above clearly shows that the CCE-based method achieved higher F1-scores, indicating a better balance between correctly finding trees and minimizing false detections, which is critical for reliable plot-level biomass aggregation.
2. Accuracy of LiDAR-Derived Tree Metrics
The MCW algorithm provided more accurate estimates of crown width compared to simple methods like fitting a minimum bounding circle or ellipse to the segmented crown polygon.
| Metric | Forest Type | Method | R² | RMSE | rRMSE (%) |
|---|---|---|---|---|---|
| Crown Width (m) | Larch | MCW | 0.86 | 0.31 | 8.5 |
| Bounding Circle | 0.72 | 0.48 | 13.1 | ||
| Pine | MCW | 0.83 | 0.35 | 9.8 | |
| Bounding Circle | 0.68 | 0.52 | 14.6 | ||
| Tree Height (m) | Larch | LiDAR (98th %ile) | 0.95 | 0.65 | 4.2 |
| Pine | LiDAR (98th %ile) | 0.93 | 0.78 | 5.1 |
The high correlation between LiDAR-derived tree heights and field measurements underscores the capability of UAV LiDAR to accurately capture vertical structure.
3. Individual Tree Above-Ground Biomass Estimation Accuracy
The allometric models developed using the structural metrics from the integrated CCE-MCW framework yielded the highest prediction accuracy. The following table compares the performance of different modeling approaches for the Larch plantation.
| Model Input (Method) | Model Form | R² | RMSE (kg) | rRMSE (%) |
|---|---|---|---|---|
| \(CD_{MCW}^2 \cdot H\) (CCE-MCW) | Log-Log | 0.89 | 24.1 | 18.3 |
| \(CD_{WS}^2 \cdot H\) (WS-LM) | Log-Log | 0.81 | 31.5 | 23.9 |
| KNN Imputation | Non-parametric | 0.85 | 27.8 | 21.1 |
| RF Imputation | Non-parametric | 0.87 | 25.9 | 19.7 |
The model based on CCE-MCW features (Log-Log) achieved an R² of 0.89 and an rRMSE of 18.3%, outperforming both the model based on standard segmentation features and the sophisticated non-parametric imputation techniques. This demonstrates that accurate physical segmentation and feature extraction can sometimes surpass purely statistical prediction methods when the underlying allometric relationship is strong and well-captured by the LiDAR metrics.
The predictive equations for the two main species were:
Larch: $$\ln(AGB) = -2.15 + 0.87 \cdot \ln(CD_{MCW}^2 \cdot H_{LiDAR})$$
Chinese Pine: $$\ln(AGB) = -1.92 + 0.83 \cdot \ln(CD_{MCW}^2 \cdot H_{LiDAR})$$
4. Plot-Level Biomass Aggregation and Validation
The individual tree AGB estimates were summed to obtain plot-level AGB. The plot-level estimates using the CCE-MCW allometric approach showed excellent agreement with the ground-based plot inventory estimates (derived from field-measured DBH and species-specific allometric equations).
| Plot ID | Forest Type | Ground AGB (Mg/ha) | CCE-MCW Est. (Mg/ha) | Relative Error (%) |
|---|---|---|---|---|
| L1 | Larch | 145.2 | 138.7 | -4.5 |
| L2 | Larch | 162.8 | 157.1 | -3.5 |
| P1 | Pine | 98.5 | 103.2 | +4.8 |
| P2 | Pine | 115.7 | 110.4 | -4.6 |
The mean absolute relative error at the plot level was less than 5%, which is highly satisfactory for operational forest carbon stock assessment using UAV drones.
Discussion
The integration of the Canopy Contour Extraction (CCE) and Multi-scale Crown Width (MCW) algorithms presents a significant advancement for processing UAV LiDAR data for forestry applications. The success of this method hinges on several factors inherent to UAV drone technology: the high point density allows for the generation of very fine-resolution CHMs, which is a prerequisite for effective crown boundary detection in CCE. The flexibility of UAV drones to fly at lower altitudes and under clouds ensures data acquisition under optimal conditions.
The CCE algorithm’s main contribution is its improved handling of crown overlap. Traditional watershed segmentation often leads to under-segmentation (multiple trees merged into one segment) in dense areas. By explicitly incorporating canopy edge information into the segmentation process, CCE provides more realistic crown delineations, which directly translates to more accurate crown width measurements and, consequently, better biomass predictions. The MCW algorithm complements this by providing a robust, scale-invariant measure of crown width that is less sensitive to crown asymmetry and small gaps within the crown, common issues in coniferous species.
While non-parametric methods like Random Forest (RF) showed competitive accuracy, the allometric approach based on physically meaningful metrics (crown area, height) from CCE-MCW offers greater interpretability and potential for transferability. The derived allometric constants (\(\alpha_0, \alpha_1\)) have biological meaning related to species-specific wood density and crown architecture. This model can be more easily adapted or recalibrated for new sites with some field samples, whereas machine learning models like RF may require extensive retraining data.
Challenges remain. The performance of CCE can degrade in extremely dense, multi-layered tropical forests where the CHM represents a nearly continuous surface. Future work could integrate full-waveform or multi-spectral LiDAR data from UAV drones to better distinguish species, which would allow for the application of species-specific allometric models, further reducing estimation error. Furthermore, automating the calibration of algorithm parameters (e.g., the scale of the mean filter in MCW, the edge detection threshold in CCE) based on forest type or stand density would enhance the operational robustness of this framework.
Conclusion
This study demonstrates a highly effective and accurate framework for estimating individual tree and plot-level above-ground biomass by leveraging the unique capabilities of UAV LiDAR technology. The synergistic integration of the Canopy Contour Extraction (CCE) algorithm for improved tree segmentation and the Multi-scale Crown Width (MCW) algorithm for robust feature extraction addresses key limitations in conventional methods. The resulting LiDAR-derived structural metrics served as excellent predictors in species-specific allometric models, achieving estimation accuracies (rRMSE ~18%) that meet or exceed the requirements for many carbon monitoring and reporting protocols.
The methodology underscores the transformative potential of UAV drones as a platform for rapid, precise, and non-destructive forest inventory. By providing a direct link between high-resolution 3D structural data and biomass allometry, this approach facilitates scalable forest carbon assessment. It offers forest managers and ecologists a powerful tool to monitor forest growth, health, and carbon sequestration potential with unprecedented spatial detail and temporal frequency, supporting evidence-based decision-making for sustainable forest management and climate change mitigation.