Integration of UAS and Backpack-LiDAR to Estimate Aboveground Biomass of Picea crassifolia Forest in Eastern Qinghai, China

Abstract

Precise aboveground biomass (AGB) estimation of forests is crucial for sustainable carbon management and ecological monitoring. Traditional methods, such as destructive sampling, field measurements of Diameter at Breast Height with height (DBH and H), and optical remote sensing imagery, often fall short in capturing detailed spatial heterogeneity in AGB estimation and are labor-intensive. Recent advancements in remote sensing technologies, predominantly Light Detection and Ranging (LiDAR), offer potential improvements in accurate AGB estimation and ecological monitoring. Nonetheless, there is limited research on the combined use of UAS (Uncrewed Aerial System) and Backpack-LiDAR technologies for detailed forest biomass. Thus, our study aimed to estimate AGB at the plot level for Picea crassifolia forests in eastern Qinghai, China, by integrating UAS-LiDAR and Backpack-LiDAR data. The Comparative Shortest Path (CSP) algorithm was employed to segment the point clouds from the Backpack-LiDAR, detect seed points and calculate the DBH of individual trees. After that, using these initial seed point files, we segmented the individual trees from the UAS-LiDAR data by employing the Point Cloud Segmentation (PCS) method and measured individual tree heights, which enabled the calculation of the observed/measured AGB across three specific areas. Furthermore, advanced regression models, such as Random Forest (RF), Multiple Linear Regression (MLR), and Support Vector Regression (SVR), are used to estimate AGB using integrated data from both sources (UAS and Backpack-LiDAR). Our results show that: (1) Backpack-LiDAR extracted DBH compared to field extracted DBH shows about (R² = 0.88, RMSE = 0.04 m) whereas UAS-LiDAR extracted height achieved the accuracy (R² = 0.91, RMSE = 1.68 m), which verifies the reliability of the abstracted DBH and height obtained from the LiDAR data. (2) Individual Tree Segmentation (ITS) using a seed file of X and Y coordinates from Backpack to UAS-LiDAR, attaining a total accuracy F-score of 0.96. (3) Using the allometric equation, we obtained AGB ranges from 9.95–409 (Mg/ha). (4) The RF model demonstrated superior accuracy with a coefficient of determination (R²) of 89%, a relative Root Mean Square Error (rRMSE) of 29.34%, and a Root Mean Square Error (RMSE) of 33.92 Mg/ha compared to the MLR and SVR models in AGB prediction. (5) The combination of Backpack-LiDAR and UAS-LiDAR enhanced the ITS accuracy for the AGB estimation of forests. This work highlights the potential of integrating LiDAR technologies to advance ecological monitoring, which can be very important for climate change mitigation and sustainable environmental management in forest monitoring practices.

1. Introduction

Forest biomass significantly impacts various ecosystem processes, such as carbon sequestration, water cycling, and energy fluxes, which affect climate [1]. Monitoring AGB plays a significant role in forest ecology by determining carbon storage and general ecosystem efficiency [2]. For better forest management, precise assessment of forest AGB estimation is crucial for advancing ecological studies and climate change mitigation strategies [3]. Up to now, two types of techniques have been employed to estimate AGB: ground-based methods and remote sensing. Ground-based methods are categorized into destructive and non-destructive techniques. Destructive approaches entail cutting down trees in the field, followed by procedures such as drying to remove moisture and weighing them to obtain precise data, as delineated in [4]. Non-destructive ground-based techniques for assessing forest AGB have been studied for several decades [5,6]. Non-destructive methods depend on equations that correlate biomass with quantifiable biophysical characteristics, including tree height, DBH, wood density, and crown dimensions, which are commonly used as predictors in allometric equations to estimate biomass [4]. Although these methods yield the most accurate measurements of forest biomass, they are intrinsically time-consuming and labor-intensive [7]. In addition, ground-based measurements have limited access to far-flung areas in the forest, which can greatly limit their accuracy [8]. However, RS-based methods provide substantial advantages for the rapid and automated assessment of biomass structural and biochemical characteristics. These approaches operate across many spatial and temporal scales with high resolution [9]. RS techniques comprise a range of technologies, including digital imaging modalities, such as hyperspectral, multispectral, and optical (red–green–blue, RGB) sensors, radar systems, photogrammetry, and laser scanning (LiDAR). Moreover, advances in sensor integration have enabled the integration of several sensor types on diverse platforms, thereby enabling efficient biomass evaluation [10].

LiDAR technology has considerably broadened the biomass estimation rate and piqued the attention of experts due to its exceptional accuracy in three-dimensional (3D) scanning properties by time-of-flight approaches at the centimeter level. LiDAR operates on the principle of laser light reflection emitted by the sensor to different surfaces (e.g., vegetation, ground, and buildings). These laser pulses reflect after hitting the target surfaces and measure the time it takes for the laser pulses to return (time-of-flight) and the intensity of the reflected light. The reflected light is then combined to generate dense 3D point clouds [11,12]. Among the LiDAR methods, UASs and Backpack-LiDAR are receiving widespread attention for estimating AGB [13]. They can measure distances with laser light reflection, detailed representation of the canopy, and the underlying topography [14]. LiDAR can extract comprehensive 3D models of the forest structure for accurate AGB assessment to overcome traditional forest survey challenges with less fieldwork and more efficiency [15].

Individual Tree Segmentation (ITS) or Individual Tree Detection (ITD) techniques based on LiDAR data analyze 3D LiDAR point clouds to identify and isolate individual trees and are of great significance in AGB estimation once the trees are accurately segmented with high precision, which allows the direct extraction of tree parameters such as tree height and crown diameter [16,17,18]. Subsequently, forest AGB is calculated using an allometric equation, which provides a valid estimate when calibrated for particular species and regions, and is supported by accurate field observations [19,20,21]. The LiDAR data-driven ITS employs both point cloud data directly and a Canopy Height Model (CHM) generated using point cloud data [17,22,23]. CHM-based segmentation is impractical for detecting understory trees using only the first echo of laser point clouds [17,23,24,25,26]. Therefore, recent studies have used point-based segmentation. For example, Reitberger et al. [27] used a random sample convergence approach to differentiate individual tree trunks with a 70% accuracy rate. Based on the characteristic wider crown spacing of coniferous trees compared to that of the understory, Li et al. [28] employed the top laser point cloud from tree crowns as seed points, utilizing a top-to-down geographical growth method, resulting in an average precision of 97%. While this method is only efficient for coniferous trees, its accuracy is reduced when used in deciduous forests. The irregular shape, with crossed and overlapped branches and leaves, makes it challenging to determine seed points from the crown top using point cloud data. Using intensity data from LiDAR point clouds obtained during a deciduous season, Lu et al. [25] extracted a topological link between a tree trunk and the point cloud used to segment the trunk, achieving a total accuracy of 90%. Nevertheless, this approach has limitations when used during the plant growth period; the intensity value of the leaves is extremely high, which complicates the process of extracting the trunk.

AGB estimation using LiDAR data is often performed by constructing a model that links the significant variables acquired from LiDAR sample data with forest inventory attributes [29]. Studies have demonstrated a significant correlation between forest biomass, height, and density variables retrieved from LiDAR point cloud data [12,30,31]. However, the ability to describe the overall canopy layer and vertical heterogeneity using only LiDAR metrics of forest height and density information is insufficient [12,32]. Therefore, utilizing the canopy height distribution and branch and leaf profiles defines typical variables of the canopy profile [16]. The initial laser point cloud echo provides an overview of the structure of the upper canopy, and the final echo distribution reveals the most profound penetration of the laser signal through the vegetation layer. It can be utilized to distinguish among different forest types with different degradation competencies while improving biomass valuation accuracy [33].

UASs are a viable alternative to satellites and aircraft for forest inventory due to their flexibility, autonomous solutions, and cost-effectiveness [34,35]. UASs mounted with LiDAR sensors provide high-resolution data for estimating forest canopy height, forest volume, and forest biomass for a portion of the time [36]. UASs with detailed 3D forest structures swiftly cover extensive regions above the ground to a certain height [37,38]. UASs-LS often pose an occlusion effect, which represents a substantial limitation in LiDAR-based remote sensing systems encompassing UAS-LiDAR and aerial platforms, such as ALS [39]. Although ALS operates at elevated altitudes and encompasses extensive regions, its coarse resolution frequently inadequately captures intricate features, especially in understory vegetation [40]. Conversely, UAS-LiDAR, functioning at lower altitudes, provides high-resolution data, yet is more vulnerable to occlusions from dense, multi-layered canopies. Elements such as scanner configurations (e.g., pulse repetition frequency and scan angle) and flight planning (e.g., altitude and overlap) are essential for alleviating these impacts [41]. Enhancing the flight path overlap and adjusting the scanner parameters help mitigate occlusion effects [42]. It has been proven that UAS-LiDAR can measure individual tree height at the forest stand level with considerable accuracy [43,44]. Nevertheless, it is challenging to measure individual tree DBH using UAS-LiDAR. This is because it can pose limitations attributed to the weakening of the laser beam when it interacts with a dense overstory forest. Hence, the estimation accuracy decreases and is primarily dependent on the diameter of the tree stem [45].

In contrast, the Backpack-LiDAR system can provide individual tree information in a “bottom to top” sequence, which is viable for accurately estimating DBH by individual tree segmentation and is used as a complement in forest inventory applications [45]. Therefore, integrating both UASs and Backpack-LiDAR systems potentially addresses the issue of inaccurate assessment (under- or over-estimation) of broadleaved forest biomass at the individual tree level due to low ITS accuracy. UAS-Backpack-LiDAR produces high-quality data to improve accuracy by capturing vertical and lateral (understory and overstory) forest complexities, specifically lower vegetation such as shrub layers, often overlooked in traditional aerial and ground-based surveys [46,47].

Machine learning (ML) has proven competent in accurately modeling intricate spatial patterns by employing a wide range of comprehensive input data [48]. The utilization of ML techniques for estimating AGB in forests has shown considerable growth in recent years [49,50]. ML techniques have progressively found their way into agroforestry systems [51,52,53]. Previous studies have shown that ML techniques, when integrated with methods for selecting features, can attain significant accuracy in predicting vegetation physiological parameters. Nevertheless, the feature extraction process is both complicated and relies on particular domain knowledge and expertise [54,55]. Furthermore, the types and quantity of extracted feature variables have a strong connection to feature selection algorithms, which can significantly constrain the capabilities and validity of ML prediction models based on these variables [56,57].

Numerous studies have reported work regarding AGB estimation in the Qinghai−Tibetan region using remote sensing technologies [18,58,59,60]. A prompt and precise assessment of AGB offers a scientific benchmark for the management and sustainable utilization of grassland resources [18]. Zhao et al. [18] reported work on AGB estimation based on MODIS and SRTM data in Qinghai. However, the work defines the accuracy assessment well but does not provide information regarding species cover by neglecting different vegetation types. Therefore, neglecting vegetation types greatly limits the accuracy. Chen et al. [61] reported work on mountain grassland by integrating the UAS-LiDAR with Multispectral Data for AGB estimation; the said work used SuperView-1, a Chinese commercial satellite that has a 0.5 m panchromatic with 2 m multispectral bands to estimate AGB. The reported work has drawbacks in defining the DBH of individual trees at the forest stand level, which is also an important factor in countering precise AGB estimations. Wang et al. [62] also reported work on integrating UAS-LiDAR with GF-2 (Gaofen Satelite) optical data to estimate the AGB of spruce plantations in Qinghai, China. However, different satellite images have different observation dates and are not well aligned with the data acquisition dates of UASs due to cloud contamination and various atmospheric factors, which causes a slight error in AGB estimations. Therefore, the combined use of UAS and Backpack-LiDAR provides an exceptional opportunity for thorough canopy interpretation, which offers a multidimensional understanding from the canopy base to the apex, simplifying an extensive, precise approach to AGB estimation at the plot level [63].

Regarding biomass assessment, the integration of these two technologies has not yet been fully explored. Picea crassifolia forests (Qinghai spruce) are widely spread in the high-altitude regions of the Qinghai−Tibetan Plateau in the eastern part of Qinghai Province, China. The terrain complexity, high elevation, and unusual weather conditions make it challenging to acquire data and estimate AGB accurately.

Thus, this research proposes a method to estimate the AGB by employing two platforms: Backpack-LiDAR and UAS-LiDAR point clouds with field-measured data. We aim to integrate Backpack and UAS-LiDAR for the precise extraction of forest structural metrics, such as tree height and DBH, to estimate AGB. We preprocessed the data and integrated both platforms for tree segmentation. As a seeding file, tree trunk location from Backpack-LiDAR point clouds was extracted to assist tree segmentation from UAS-LiDAR point clouds. After extracting the (DBH) and (H), the AGB of Picea crassifolia forests in eastern Qinghai was calculated using allometric equations. Furthermore, the normalized UAS-LiDAR data were used to extract optimal metrics as dependent variables, along with AGB calculated as independent variables in three machine learning regression model approaches, MLR, RF, and SVR, to predict AGB. The performances of these three models were evaluated and compared.

2. Materials and Methods

2.1. Study Area

The Alpine grassland on the Qinghai−Tibetan plateau is an important source of animal husbandry production base and acts as an ecological security barrier in China. It is extremely sensitive to global climate change and human activities [59]. The Huangnan Prefecture, geographically located between (34°0′–36°0′N and 100°33′–102°37′E), situated in Qinghai Province, China, with a total area of approximately 17,921 km², as depicted in Figure 1c, was selected as the research area. The Picea crassifolia forest plots were selected randomly within the region from three different townships, Lancia Township, Qukuhu Township, and Maixiu Township, as shown in Figure 1c. The area experiences a continental plateau climate with average annual temperatures ranging from −0.9 °C to 8.5 °C and yearly precipitation between 329 and 505 mm [64]. Vegetation varies, with dense coverage in southeastern Qinghai and low vegetation in the northwest, mainly in grasslands and deserts [65]. Urbanization has increased significantly since 1990, from the Lanzhou–Xining city area to the central-southern Qinghai–Tibet Plateau [66].

2.2. Dataset Acquisition

2.2.1. UAS-LiDAR Data

The UAS-LiDAR data were collected using the LiDAR system ZENMUSE-L1 manufactured by DJI, Shenzhen, China. mounted on a DJI-M300 UAS-LS with a 60% side-to-side overlap rate between two successive flight paths [67]. Zenmuse-L1 features a horizontal field of view of 70.4° and a vertical field of view of 4.5°, capable of producing 240,000 points per second at a maximum height of 450 m aboveground level (AGL) [68]. The detailed sensor specifications for DJI ZENMUSE-L1 are available in [67]. The colored point clouds with RGB values from optical imagery in the ASPRS LAS version 1.4 [69] were acquired by post-processing RTK, IMU, optical imagery, and laser scanning datasets employing the DJI Terra software version 4.4.0 application provided by the manufacturer [70]. Table 1 summarizes the three plots of the UAS-LiDAR. In October 2023, UAS-LiDAR point cloud data were acquired from three quadrats within the designated study area. The sample plots were located at an altitude of 3766 m, while the UAS was used at altitudes ranging from 80 to 100 m from the terrain surface, maintaining a consistent flight ground speed of 5 m/s and covering a square area of 100 × 100 m. The average point cloud density was recorded at 100 points per square meter with a ranging accuracy of 15/10 mm. The LiDAR sensor captured comprehensive waveform information, with each laser return point containing data on the echo, intensity, coordinates, and other pertinent attributes. The LiDAR point cloud was georeferenced using a WGS84 UTM 47N projection system. Overall, intensity normalization can enhance the precision of forest attributes; however, this enhancement was nominal. Thus, we refrained from implementing intensity normalization [68].

2.2.2. Backpack-LiDAR Data

The Green Valley Backpack-LiDAR System was used with a LiDAR scanner (FEIMA, SLAM100) manufactured by Shenzhen Feima Robotics Co., Ltd., based in Shenzhen, China, a Position Orientation System (POS), and a handheld touchpad. Backpack-LiDAR employs industry-level SLAM algorithms to capture high-precision 3D point cloud data without requiring light. Similarly, three sample plots of 30 × 30 m were surveyed using the Backpack-LiDAR system in the abovementioned region. The data were collected using an ‘S’ shape trajectory, and the artificial marker was positioned at predetermined coordinates at the midpoint of the sampling plots, as shown in Figure 2. The laser scanning distance was 120 m, laser pulse repetition rate 320 kHz, scanning frequency 320,000 pts·s⁻¹, laser wavelength 903 nm, laser level class 1 Eye-Safe Laser, mean point density 7135 m⁻² with 270° × 360° angle of field view with absolute accuracy 5 cm, relative accuracy 2 cm, horizontal field of view angle 200°, and vertical field of view angle 200° along with UniStrong G970II Geodetic GNSS RTK manufactured by Beijing UniStrong Science and Technology Co., Ltd, Beijing, China, GPS were used [71]. The WGS84 coordinate system and UTM 47N were employed for the data projections.

2.2.3. Field Data

We followed the data collection methodology of Wang et al. and Lu et al. [17,72] and established five 10 × 10 m subplots at each 30 × 30 m plot’s four corners and one at its center. Field data were gathered for each plot, with two standard trees chosen per subplot, in October 2023 (Figure 3). The geographic coordinates of the two standard trees within each subplot were recorded. In each 30 × 30 m plot, 15 tree coordinates were recorded using a UniStrong G970II GPS device. Moreover, measuring taps with a laser altimeter were used to calculate the DBH and H of each tree.

2.3. Methodology

Firstly, dataset point clouds (UAS and Backpack-LiDAR) were initially projected, normalized, and aligned to eradicate the noise ratio, as reported by [17]. Secondly, the Comparative Shortest Path (CSP) algorithm [73] was employed to segment the point clouds from the Backpack-LiDAR, detect seed points, and calculate the DBH of individual trees. After that, using these initial seed point files, the UAS-LiDAR data segmented the individual trees using the PCS method and measured tree heights from the point clouds, which enabled the calculation of observed/measured AGB across three specific areas using an allometric equation. Consequently, we extracted the UAS-LiDAR metric (height, intensity, and density-related) and observed the most important variable used to build three parametric and non-parametric models: MLR, RF, and SVR. Finally, MLR, RF, and SVR model results were compared and evaluated to predict AGB, as shown in Figure 4.

2.3.1. Aligning Backpack and UAS-LiDAR System

Variation in scale and mismatch discrepancies between point clouds are two significant limitations of planimetric and vertical dimensions derived from Backpack-LiDAR [74]. Our proposed methodology addresses these issues through a two-stage process: terrain normalization and alignment of point clouds. Initially, advanced progressive Triangulated Irregular Network (TIN) densification filtering techniques were used to construct the Digital Terrain Model (DTM) [75]. Subsequently, the remaining ground points were interpolated by employing the Inverse Distance Weighting Algorithm (IDW), which allows adjusting the elevation variations of the point cloud heights by subtracting the DTM-derived ground surface height to normalize the point clouds. Therefore, this normalization ensures that the ground points brought from various sources of LiDAR are on a common horizontal plane; it effectively mitigates vertical discrepancies. Afterward, we follow the alignment suggested by Polewski et al. [76] to manually select at least three pairs of corresponding points within both datasets for the alignment of point clouds from the Backpack-LiDAR and UAS-LiDAR datasets. This intervention is laborious and reasonably accurate because one can guarantee that the three datasets are co-registered as depicted in Table 2. The second process of normalization addresses any residual differences that can be present in the z-values to preserve the integrity of the vertical dimension post-matching, as shown in Figure 5.

2.3.2. Tree Segmentation of Backpack and UAS-LiDAR System

Green Valley International LiDAR 360 software version 5.2 was used for individual tree segmentation. The data-driven Backpack and UAS-LiDAR point clouds underwent a segmentation process using the LiDAR 360 software, as shown in Figure 6a. Initially, we employed a Density-based Spatial Clustering Application with a Noise algorithm (DBSCAN) [77] to the individual tree trunks of Backpack point cloud data. Moreover, with the CSP algorithm [73], standing tree point clusters were located, followed by isolated parts of the point cloud up to the distance of 1.3 m tree trunk height and volume of tree DBH, using a cylindrical-fitting method, as depicted in Figure 6b. The DBH fitting results are obtained by manual adjustment, ensuring the measurement accuracy of each tree, as demonstrated in Figure 6c. The CSP algorithm identifies points within a specified DBH radius or closest to a point to initiate the clustering of seed points for further segmentation, utilizing the 3D coordinates of these seed points. The precision of this segmentation process is critical to the overall accuracy of tree modeling and the subsequent estimation of individual trees [78]. Each segmented tree’s point cloud underwent individual verification, manually adjusted for incorrect classifications. After removing irrelevant data points, seed point files containing the X and Y coordinates and DBH were finalized, facilitating the isolation of individual trees within the UAS-LiDAR point cloud data.

Afterwards, using Point Cloud Segmentation (PCS) [28], the UAS-LiDAR data were employed to define the tree tops. Individual trees were segmented distinctly from the point cloud by exploiting their relative spacing. In contrast to the PCS method implementation, where the canopy peak is used as the seed point of tree canopy extraction, the X and Y coordinates of a tree trunk were used as a seed point acquired by Backpack-LiDAR. The aligned and normalized data from Backpack-LiDAR were used as reference data to assess data quality in terms of tree segmentation accuracy, as shown in Figure 6d.

2.3.3. Segmentation Accuracy Assessment

There are three types of tree segments and levels of accuracy, as depicted in Equations (1)–(3). If a tree is correctly segmented, it is called a true positive (TP). False Negative (FN) refers to undetected trees missed by the segmentation process. Whereas, false positive (FP) denotes incorrectly detected trees from the point clouds. The recall is the tree detection rate, and the F-score incorporates commission and omission errors, which provides a comprehensive assessment of accuracy. Recall (r), precision (p), and the F-score are measured on a scale from 0 to 1, where scores between 0 and 1 indicate varying levels of accuracy. A higher score indicates precise detection and segmentation accuracy, while a lower score indicates considerable error in tree detection and identification [28]. Both recall and precision need to be elevated to attain a high F-score. For example, if every tree is segmented flawlessly, the recall and precision reach one, leading to an F-score of one [79]. Three statistical parameters were applied to evaluate the effectiveness of the ITS algorithm:

(1)$r={\displaystyle \frac{TP}{TP+FN}}$

(2)$p={\displaystyle \frac{TP}{TP+FP}}$

(3)$F=2\times {\displaystyle \frac{r\times p}{r+p}}$

where (r) recall is the tree detection rate, (p) represents the precision of the precisely identified tree, and (F) F-score is a composite measure of accuracy. The above equations (e.g., (1–3)) described by Goutte, C. and Gaussier, E was used for ITS segmentation accuracy [79].

2.3.4. Biomass Calculation of Picea Crassifolia

The allometry equation of specific species (Picea crassifolia) Equations (4)–(7) were used in this study [80], and tree biomass components such as W_L (leaves), W_T (trunk), and W_B (branch) were calculated from DBH and height. Within Huangnan Prefecture, three forest regions were surveyed using UAS and Backpack-LiDAR, from which three 30 × 30 m plots were extracted using LiDAR 360 software and were further subdivided into 10 × 10 m, a total of 27 subplots, as mentioned earlier in Figure 3.

The AGB is the sum of each tree component (trunk, branch, and leaves) and was then totaled to obtain the AGB of each subplot [81]. We detected the individual trees from the point cloud data [82]; the field quadrat survey data were used as the response variable for the calculation of the actual biomass of the quadrat; its value is considered reliable and was used to build and test forest AGB estimation models. The forest AGB of all subplots is shown in Table 3. The following equation (e.g., (4–7)) described by Zhou Guoyi et al. [80] was used to calculate the forest AGB.

(4)$W_L=0.0120{\left(D^{2}H\right)}^{0.8654}$

(5)$W_T=0.0447{\left(D^{2}H\right)}^{0.8564}$

(6)$W_B=0.0184{\left(D^{2}H\right)}^{0.8539}$

(7)$AGB=\sum _n(W_T+W_B+W_L)$

Allometric growth equation of Picea crassifolia species, where W_T is the trunk biomass, W_B is the branch biomass and W_L, is the leaf biomass, D² is the diameter at breast height in cm, and H is the height of each tree in m.

2.3.5. UAS-LiDAR Metrics

UAS-LiDAR metrics, as outlined in previous research [33,83,84,85,86,87], encompass height, density, and intensity measures obtained from all echoes, including leaf points extracted from the last echo, which were used for AGB prediction in different machine learning models. Height-related metrics include percentiles (e.g., $H_1$, $H_{25}, H_{50}$, $H_{75}$, $H_{95}$, $H_{99}$), mean height ($H_{mean}$), coefficient of variation ($H_{cv}$), and variance ($H_{variance}$). Density metrics were used to explain the proportion of canopy returns exceeding specific percentiles. Intensity metrics are parallel to height metrics, but use intensity values instead. Fifty-two metrics (Table 4) were extracted from UAS-LiDAR using LiDAR 360 software v5.2, from which 15 optimal metrics were selected using Waikato Environment for Knowledge Analysis (WEKA), a machine learning and data analysis software, RF variable importance ranking, and Pearson correlation were prepared for AGB prediction in the Picea crassifolia forest to use within the predictive models (MLR, SVR, RF).

2.3.6. Model Development for AGB Estimation

Modeling techniques such as comparison of (MLR, RF, and SVR), parametric, and non-parametric methods impact the AGB estimation results in accuracy, as reported by [17,88]. Subsequently, MLR, RF, and SVR were implemented to compute the AGB in the forest subplots; the independent variables were LiDAR-derived metrics, and the dependent variables were the measured forest AGB.

MLR is a simple parametric method that considers predictor dependencies and correlations; this technique has been widely described by [33,89] as one of the most used methods in the AGB estimation process. However, most earlier studies have adopted logarithmic transformations to enhance the ability to fit the model’s fitness [90]. The current study also considered logarithmic transformations to enable the presented case to fit appropriately; note that our LiDAR metrics’ values are not negative. Hence, the optimal LiDAR metrics were identified by considering the CfsSubsetEval evaluator and scoring attributes (i.e., metrics in this study) based on their predictive ability and redundancy before implementing the MLR model using the WEKA software suite version 3.9.6. The search was conducted using a forward strategy, beginning with an empty set of attributes and systematically adding the most suitable attribute at each iteration. The WEKA-evaluated metrics were positively correlated in comparison with the Pearson correlation, whereas the optimal metrics evaluated by WEKA, RF, and Pearson correlation exhibited lower internal correlation among them. After selecting the optimal metrics, MLR was utilized to derive the optimal AGB model. Subsequently, the model’s performance was evaluated using the Leave-One-Out Cross-Validation (LOOCV) method.

The RF algorithm, distinguished by its non-parametric decision-tree-based classification framework, excels in mitigating overfitting and ensuring robustness against outliers and noise [91]. Optimizing the RF model involves adjusting two key parameters: ntree, which is the total number of tree nodes in the model, and mtry, which is the number of variables considered at each tree node [92,93]. Subsequently, based on the error distribution and interpretation rate, our proposed study suggests that the optimal parameter value be set to 1000 ntrees and one-third of the predictive variables for mtry. Conversely, metric significance was assessed based on an increase in the mean squared error ensuing its elimination. Similarly, the RF model accuracy was evaluated using LOOCV, similar to the MLR model evaluation.

SVR is a hyperplane model based on structural risk reduction. The AGB prediction model performance was also assessed by employing the SVR method. SVR model optimization requires tuning of epsilon (ε), the regularization parameter (C), and the kernel width parameter (γ), three crucial hyperparameters to improve model accuracy and efficiency [94]. The vector (w) is problematic in discriminating within lower dimensional space mapped to high-dimensional space by nonlinear transformation through kernel function, improving to linearly separable. Using a regularization parameter (C) and kernel parameter gamma (γ) decreases the residuals between the data and hyperplane to enhance the model accuracy [95,96]. The kernel function type, (C), and (γ) parameters effectively reduce the error to obtain the optimal SVR model [97]. There are three methods of kernel function (line, polynomial, and RBF). The radial basis function (RBF) performs better than line and polynomials [98]. The RBF kernel significantly enhances the prediction of forest AGB, predominantly esteemed for its efficiency in simplifying parameters [99]. Thus, the RBF was used in the SVR regression model for AGB prediction. Subsequently, the LOOCV approach was employed to validate the model’s accuracy. Hence, for the accurate estimation of AGB, the proposed study underlines the importance of selecting suitable modeling parameters and techniques, demonstrating the strengths and applications of the MLR, RF, and SVR methods in environmental research. Afterward, using the coefficient of determination R², RMSE, and rRMSE, the model accuracy was further assessed using the following equation (e.g., (8–10)), as described by Lu et al. [17] and detailed below:

(8)$R^2=1-{\displaystyle \frac{\sum _{i=1}^m{({\widehat{y}}^{\left(i\right)}-y^{\left(i\right)})}^2}{\sum _{i=1}^m{({\overline{y}}^{\left(i\right)}-y^{\left(i\right)})}^2}}$

(9)$RMSE=\sqrt{{\displaystyle \frac{1}{m}}\sum _{i=1}^m{ \left(y^{\left(i\right)}-{\widehat{y}}^{\left(i\right)}\right)}^2}$

(10)$rRMSE \left(\%\right)={\displaystyle \frac{\sqrt{\frac{1}{m}\sum _{i=1}^m{ \left(y^{\left(i\right)}-{\widehat{y}}^{\left(i\right)}\right)}^2}}{\overline{y}}}\times 100\%$

where (m) represents the number of estimated values, $\left(y^{\left(i\right)}\right)$ denotes the ith measured value, (${\widehat{y}}^{\left(i\right)}$) signifies the ith estimated value, and ${(\overline{y}}^{\left(i\right)})$ is the average of the measured values.

3. Results

3.1. Individual Tree Segmentation

Data analysis for Picea crassifolia forests based on Backpack-LiDAR using the ITS technique showed considerable accuracy (Table 5). The r value for ITS ranged from 0.93–0.97, with an average of 0.95. Correspondingly, the p-value is between 0.85–0.90, with an average of 0.88. Moreover, the F-value varied from 0.89 to 0.94, averaging 0.91. Similarly, the UAS-LiDAR segmentation method (Table 6) achieved higher accuracy using seed points obtained through the Backpack-LiDAR method. The value for r reached 0.96–92, with an average of 0.95, while the average p-value stood at 0.98; the F-value ranged from 0.97–0.95, with an average of 0.96.

3.2. Assessment of DBH and H from LiDAR

Figure 7 compares the field-measured DBHs and height of standard trees in 27 subplots with the mean value of the DBH and height extracted from UAS and Backpack-LiDAR in 27 subplots. Backpack-LiDAR extracted DBH compared to field extracted DBH shows about (R² = 0.88, RMSE = 0.04 m) as depicted in Figure 7a. Figure 7b demonstrates the comparison between the mean tree height value extracted from UAS-LiDAR of 27 subplots based on seed points, and the field-measured heights of standard trees confirmed the accuracy (R² = 0.91, RMSE = 1.68 m), which verifies the reliability of the abstracted DBH and height obtained from the LiDAR data.

3.3. MLR-Based Biomass Assessment

Table 7 demonstrates the forest AGB estimation results by the MLR model, the LiDAR-derived metrics selected by WEKA, RF variable importance ranking, and Pearson correlation compared to select optimal predictive metrics, as shown in Figure 8. The MLR model yielded a moderate estimation accuracy with an average R² of 0.77, RMSE of 48.18 (Mg/ha), and rRMSE of 41.88% at the three forest plots, in addition to the selection of height-related metrics and density-related metrics, density_metrics_4, elev_percentile_95th, and elev_percentile_99th stood out as most important metrics for AGB prediction, which were obtained by RF variable importance ranking. Figure 9a presents the cross-validation results of the predicted AGB using the MLR models and the field-measured AGB in the three plots of Picea crassifolia forests.

3.4. RF-Based Biomass Assessment

The importance ranks of the LiDAR-derived metrics determined by the RF models indicate that density_metrics_4, elev_percentile_95th, and elev_percentile_99th are the essential LiDAR metrics in the AGB estimation model in Picea crassifolia forests in the three plots. The RF variable importance ranking (Figure 8a) demonstrates that the density_metrics_4 is the most crucial LiDAR metric, followed by elev_percentile_95th and elev_percentile_99th. Figure 8b shows the Pearson correlation, where the abovementioned variables positively correlate with AGB as the most important metric. Figure 9b also validates the RF plot, field-measured AGB validation results, and predicted AGB by the RF model, where the RF-fitted model outperformed the results of both the MLR and SVR models. The RF model produced the best estimation accuracy over the SVR and MLR models with R² of 0.89, RMSE 33.92 (Mg/ha), and rRMSE 29.34% values at the three forest plots.

3.5. SVR-Based Biomass Assessment

Figure 9c shows the SVR model results for AGB prediction, where the model was trained on 80% of the dataset, with the remaining 20% reserved for testing. The LiDAR-derived metrics selected by WEKA were compared with RF, and the Pearson correlation was used against the target variable in a predictive model. To evaluate the SVR model’s performance, a LOOCV approach was used as a cross-validation strategy on the training data to ensure its ability. We selected a radial basis function (RBF) kernel-based SVR model due to its ability to handle nonlinear patterns observed in the preliminary data analysis. The SVR model utilized the “svm ()” function from the relevant R studio build 524 package, specifying epsilon-regression to focus on predicting continuous AGB values. Before training, the input features were normalized to ensure that the scale of the data did not bias the model’s performance. The metrics evaluation criteria on the test data for SVR models specified an R² of 0.82, RMSE of 45.09 (Mg/ha), and rRMSE of 39.00%. AGB estimation comparative analysis with different machine learning algorithms suggests that the RF model outperforms both MLR and SVR models in accuracy.

4. Discussion

4.1. Accuracy Assessment for Tree Segmentation

The CSP algorithm [73] and ITS were employed to process Backpack-LiDAR data for tree segmentation, as illustrated in Figure 6. The segmentation accuracy values were between 0.92 and 0.94, with an average value of 0.91, obtained by Equations (1)–(3). The ITS approach allowed precise determination of DBH for each tree, achieving high accuracy (R² = 0.88, RMSE = 0.04 m) (Figure 7a) for a representative tree in each subplot. However, the present results, compared with those of Lu et al. [17] (R² = 0.94, RMSE = 1.02 cm) show slightly lower accuracy due to the uneven terrain condition of our study area, which greatly limits the accuracy. Parsetyo et al. [100] propose the efficacy of low-cost Backpack-LiDAR technology for estimating carbon stocks. Their findings regarding diameter at breast height (DBH) indicate a (R² = 0.99822) and (RMSE = 0.53 cm). In comparison to the current study, their results demonstrated marginally superior accuracy, attributable to varying factors, such as altitude, climate, and species type, which significantly constrain precision. To process the UAS-LiDAR data shown in Figure 7b, the PCS method was employed for tree segmentation. The segmentation accuracy was between 0.95 and 0.97, with an average value of 0.96. The PCS method allows the precise determination of each tree in each subplot, obtaining a high accuracy of (R² = 0.91 RMSE = 1.68 m) compared to the description by Chen et al. [37], with a flight height of 80 m; the tree height estimation achieved considerable accuracy of (R² = 0.75, RMSE = 2.65). Lu et al. [17] also reported in their research that the tree height compared to the field height also shows considerable accuracy (R² = 0.83, RMSE = 1.48 m). In the comparison of existing studies, our tree height estimation accuracy achieved the highest accuracy of (R² = 0.91 RMSE = 1.68 m); this occurred due to the Picea crassifolia forest, which is generally recognized as a coniferous forest type, allowing for straightforward identification of tree tops through the PCS method utilizing UAS-LiDAR. Consequently, it enhances the precision of tree height estimation. Without using tree location data from Backpack-LiDAR as initial seed points, the ITS method’s accuracy significantly decreased by 0.18 (dropping from 0.96 to 0.78), as verified by comparing Table 6 and Table 8, leading to frequent over- or under-segmentation of many trees. This decrease can be attributed to using the PCS algorithm [28] to extract individual tree tops/crowns and employing a top-to-bottom region growing technique to ascertain tree spacing, which is notably effective in coniferous forests. However, the dense, intertwined canopies, proximity of trees, and abundant shrubbery in our study area complicated the detection of tree crowns and trunks at breast height. To counteract these challenges and improve ITS accuracy for individual tree detection, we integrated tree location data (seed points) sourced from Backpack-LiDAR, which enabled the precise extraction of individual tree metrics using the tree segmentation tool within the LiDAR 360 software. According to the data presented in (Table 3), incorporating Backpack-LiDAR-derived tree locations led to the accurate segmentation of over 80% of the trees across the three study plots. In AGB estimation, values typically range from 50 to over 300 Mg/ha, depending on forest type, age, and environmental conditions [101]. We established three sample plots of Picea crassifolia in eastern Qinghai, Huangnan Prefecture, representing local forest conditions. Our RF model performed well, with an RMSE of 33.92 Mg/ha and an estimated AGB range of 50–400 Mg/ha, consistent with coniferous forests like Picea crassifolia. These forests exhibit moderate to high biomass due to their dense canopy, with variability influenced by factors such as forest maturity, altitude, and human disturbance. The RF model outperformed the MLR and SVR models. Compared to previous research, Zhang et al. [60] estimated AGB at 130–157 Mg/ha in Picea crassifolia forests of the nearby Qilian Mountains using in situ measurements, UAS data, and MODIS imagery. Chen et al. [37], in similar ecological zones, reported AGB values of 50–100 Mg/ha at lower elevations. In some cases, specific estimates for Picea crassifolia forests showed growth rates of around 55.99 t/ha per decade [102]. These findings are consistent with the lower range of our estimates. Differences in AGB estimates may result from previous studies calculating potential AGB under ideal conditions, while we estimated actual AGB, reflecting environmental stresses and disturbances. Additionally, variations in the input parameters, sample size, sample plot number, and spatial distribution significantly impact model accuracy. The number of ground samples remains a key factor influencing the AGB estimation precision [103]. The terrestrial/airborne LiDAR point cloud (ALS) data combination significantly boosts the ITS method’s accuracy and operational efficiency. Apart from that, we could not obtain the individual tree height from the UAS-LiDAR data until the individual trees were segmented correctly, and thus, both individual tree and plot-level field-estimated AGB could not be calculated.

4.2. Evaluation of MLR SVR and RF Model

In this research, we focused on calculating the AGB in forest areas using allometry Equations (4)–(7), which depend on two parameters: the height of trees H and their DBH. Typically, these measurements are taken directly from trees in designated plots, a method known for its precision and notably labor-intensiveness. Recent advancements have demonstrated that LiDAR technology, both from drones (Airborne/UAS) and ground-based systems [104,105] (terrestrial/backpack), can accurately capture these measurements with considerable accuracy [104,106]. Our findings (Figure 7) echo these advancements, showing that tree height and DBH can be efficiently derived from UAS and Backpack-LiDAR data. Given these insights, we calculated the AGB using LiDAR-derived measurements.

While advantageous in terms of accuracy, this approach requires substantial data processing and computation time. Consequently, we limited this technique to 27 small plots of 10 × 10 m to calculate the AGB using Equations (4)–(7). We further analyzed 52 metrics (Table 4) from the UAS-LiDAR point cloud data from which 15 optimal metrics were extracted (Table 9) and used within the AGB prediction models to forecast the AGB across the study area, specifically for Picea crassifolia forests. Among all the employed modeling techniques (MLR, SVR, and RF), the RF model showed superior performance in terms of prediction accuracy and error rates compared to the MLR and SVR models, as shown in (Table 10) and (Figure 10). MLR showed an R² of 0.77, with an rRMSE of 41.68%; RF showed an R² of 0.89, with an rRMSE of 29.34%; and SVR showed an R² of 0.82, with an rRMSE of 39.00% of the total predicted AGB value.

4.3. Variable Importance for AGB Estimation

Variables such as height and density are strongly correlated with the structural features of forests and their biomass, and these factors are mostly assigned higher weights in the AGB model [107]. Although they are less predictive on their own, lower-ranked variables offer supplementary information [108]. Figure 8a shows the higher important variable features like density_metrics.4, elev_percentile_95th, and elev_percentile_99th in the model. In contrast, such variables indicate the structural features of forest biomass, like canopy height and density, as they have a high predictive relationship with AGB [109]. The Elev IQ (Interquartile range) of elevation measures variability in vegetation height, which is an important predictor of forest structure and biomass. However, it appears as the top 6th most predictive variable in the model; hilly mountainous regions with steep slopes might be the reason. While canopy cover has a minor impact on the model outcomes, it typically complements height-based metrics by offering an estimate of horizontal vegetation distribution, which is vital for biomass modeling [110]. At the same time, lower significance variables, including density_metrics.6 and density_metrics.9, may capture less relevant structural features or redundant data in comparison with higher-ranked metrics. Such features could overlap considerably with stronger predictors, weakening their independent contribution to the AGB estimation. In addition, elev_kurtosis and int_kurtosis (Figure 8a) reflect the statistical distribution of elevation or intensity but are less directly linked to biomass than height percentiles or density metrics [111].

According to the importance ranking of variables by the RF model and Pearson correlation in R studio (Figure 8a and Table 9), the top fifteen ranked variables in Picea crassifolia forest included variables relating to canopy horizontal distribution, height-related, and intensity-related distribution. The RF model, reselected density_metric 4, previously identified by the WEKA machine learning software in the MLR model, is significant for AGB estimation in Picea crassifolia forests. These metrics, along with density_metric 3 and elevation percentiles (95, 99, and 70), were found to have a strong positive correlation (Figure 8b) with AGB by Pearson correlation. We set a Pearson correlation (Figure 8b) threshold of 0.8 to manage multicollinearity in our prediction models. We observed that pairs such as elev_percentile_99th and elev_percentile_95th, as well as density_metrics.4 and density_metrics.3, exceeded this threshold. Therefore, we retained the variable with higher feature importance or theoretical relevance to the AGB estimation. The following approach ensured that each predictor variable in our regression model contributed unique information, reducing redundancy and enhancing model stability, indicating their high predictive value, which underscores the relationship between specific LiDAR-derived metrics and AGB.

Dividing the study area into more petite plots introduces methodological challenges such as the overlapping of tree crowns across plot boundaries, which could potentially impact the accuracy of our measurements [112]. Despite these challenges, our research shows that the LiDAR technique is the most reliable and effective way to accurately evaluate individual tree attributes for AGB prediction among remote sensing technologies.

4.4. Performance Comparison of Each Model

The proposed research represents the model’s effects using the LiDAR technique to estimate forest AGB. Figure 9 and Table 10 summarize the comparative analyses of the three biomass prediction models. Therefore, using field-measured forest AGB values for training and testing purposes, these models ensure that the models are developed using actual data collected in the real world. The model performances of RF, MLR, and SVR were compared, and the MLR model demonstrated the lowest accuracy, as shown in Figure 9a. The RF model stood out by attaining maximum accuracy, with a coefficient of determination R² of 0.89, rRMSE of 29.34%, and RMSE of 33.92 Mg/ha, as depicted in Figure 9b. Thus, according to the following research, the RF model slightly surpassed the MLR and SVR model performances, as illustrated in Figure 9a,c. Machine learning models, such as RF and SVR, establish superior performance owing to their capacity to manage nonlinear relationships and interactions among variables [113]. Linear Models such as MLR may underperform as they depend on linearity, which could fail to adequately capture the intricacies of forest biomass estimation [114].

The dimensions of the sample plot, along with its spatial arrangement, greatly altered the efficacy of the machine learning models in the present study. More extensive plots potentially encompassed more variability in tree structure and biomass, thereby strengthening the model training, as indicated by Hernández et al. [115]. In contrast, smaller or unevenly dispersed plots may have introduced biases, diminishing the model’s ability to generalize across the research area. Prior studies such as Su et al., Schäfer et al., and Lu et al. [116,117,118] further substantiate that larger and more evenly dispersed plots improve the correlation between LiDAR-derived metrics and AGB, hence improving predictive accuracy. The results show that LiDAR technology strengthens the feasibility of using LiDAR data for forest biomass estimation, signifying its ability to capture the vertical structure of canopies accurately. Figure 11 demonstrates the schematic presentation of segmentation accuracy, whereas (Figure 11b) shows individual trees segmented using seed points, which significantly improved segmentation accuracy as compared with (Figure 11c) segmentation without using seed points. Subsequently, previous studies using LiDAR methods reported that the AGB prediction models for biomass estimation show an R² ranging from 0.43 to 0.95. Our results are well aligned with those of prior studies [16,17,83,88], demonstrating the consistency of our approach in biomass estimation. However, the overall performance of the RF model specified by the RMSE and rRMSE showed considerable accuracy with a slight variance in the data. This was primarily because the study area was in a hilly mountainous region, making it challenging to collect larger sample plots and resulting in fewer trees being assessed. Thus, increasing the number of sample plots is subjected to increase the accuracy of the predicted models. Increasing the number of sample plots will be the focus of our future work.

5. Conclusions

Our findings highlight the effectiveness of combining UAS and Backpack-LiDAR technologies to address the challenges of tree segmentation accuracy, particularly within coniferous tree forests during the growing season. The conclusions were derived from the experimental results. Notably, Backpack-LiDAR extracted DBH compared to field extracted DBH shows about (R² = 0.88, RMSE = 0.04 m) whereas UAS-LiDAR extracted height achieved the accuracy (R² = 0.91, RMSE = 1.68 m), which verifies the reliability of the abstracted DBH and height obtained from the LiDAR data. The application of Backpack-LiDAR-derived tree trunk positions as seed points for ITS using UAS-LiDAR data markedly enhanced segmentation accuracy, achieving a total accuracy F-score of 0.96. Using the allometric equation, we obtained AGB ranges from 9.95–409 (Mg/ha). This advancement signifies a step forward in applying lightweight and versatile LiDAR systems in forest management and conservation efforts. The combined use of UASs with Backpack-LiDAR data offers an efficient solution to address the issue of under- or over-segmentation in coniferous forests by utilizing (ITS) approach. The RF model demonstrated superior performance, achieving the highest accuracy (R² of 89%, with an rRMSE of 29.34% and an RMSE of 33.29 Mg/ha).

Conversely, the MLR model exhibited a low accuracy (R² of 77% with a variance of rRMSE of 41.68% and an RMSE of 48.18 Mg/ha). Analysis revealed that the RF model also slightly surpassed the SVR model in terms of accuracy (R² of 82% with a variance showing rRMSE of 39.00% and an RMSE of 45.09 Mg/ha); hence, RF and SVR both outperformed the MLR model. Furthermore, by applying WEKA machine learning software and Pearson correlation with the RF importance variable ranker, our study identified optimal LiDAR metrics for AGB estimation, highlighting the importance of height-related and density-related metrics. Specifically, density_metrics_4, elev_percentile_95th, and elev_percentile_99th emerged as critical indicators, with density_metrics_4 being the most influential in accurately estimating AGB across different forest plots. These insights provide a robust framework for employing advanced LiDAR technologies and machine learning models for accurate forest biomass structure evaluation. Such advancements hold significant implications for forest management practices, enabling more informed decision-making processes in the conservation and sustainable management of Picea crassifolia forests. Our study suggests that the data quality, canopy type, and geographical location of the study area can affect model performance. Moreover, future research with additional machine learning techniques such as Extreme Boosting (XGBoost) with increasing plot size and number is recommended to improve AGB prediction accuracy.

Author Contributions

J.S.A.: Data curation, Formal analysis, Methodology, Software, Validation, Visualization, Writing—original draft. L.C., B.L. and C.W.: UAS-LiDAR and Backpack-LiDAR Data collection. F.Z.: Supervision and Funding, Y.A.B. and S.A.J.: Review and editing, Y.N.: Supervision, Conceptualization and Funding. All authors have read and agreed to the published version of the manuscript.

Data Availability Statement

The original contributions presented in this study are included in the article. Further inquiries can be directed to the corresponding author.

Conflicts of Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1. Maps of the selected Picea crassifolia forest study area and distribution of sample plots. (a) Location of the Qinghai−Tibetan Plateau in China. (b) The Digital Elevation Model (DEM) stands for the height from sea level (c) Normalized Difference Vegetation Index (NDVI) shows the vegetation distribution ratio map of three Picea crassifolia forest sites using Google Earth Engine (GEE) sentinel satellite images with Google Earth images showing the study area (d) LiDAR point cloud cross-sectional view of alignment; the black dots represent Backpack-LiDAR data, while chromatic dots show UAS-LiDAR data alignment of both platforms over each other.

Figure 2. One plot of 30 × 30 m for the Backpack-LiDAR data acquisition method; the yellow lines are the distance from one end to the other end of the plot, and the red lines represent the data acquisition track.

Figure 3. The schematic diagram for one plot data segmentation subdivided each 30 × 30 m plot into 10 × 10 m.

Figure 4. The workflow overview for estimating forest aboveground biomass using LiDAR data.

Figure 5. Data alignment of both platforms (Backpack and UAS-LiDAR): (a) Marking numbers for the same tree point cloud in the UAS and Backpack-LiDAR point clouds, (b) Normalizing 30 × 30 m LiDAR point clouds, (c) Cross section line on point clouds chromatic color represents Backpack while RGB is UAS point clouds, (d) Overlay maps of original Backpack-LiDAR (black) and UAS-LiDAR (chromatic) point, (e) Overlay maps of normalized Backpack-LiDAR and UAS-LiDAR point clouds. A yellow line in figure (c,d) represents cross-sectional region on point clouds from “A towards ”B at both ends.

Figure 6. LiDAR 360 software interface for Backpack-LiDAR data segmentation; the result shows a trunk slice at 1.3 m, (a) Backpack-LiDAR point cloud data for a single tree. (b) Point cloud data fitted to DBH at 1.3 m. (c) Incorrectly classified trees manually corrected. (d) Segmentation results of Backpack-LiDAR point clouds.

Figure 7. Comparisons between field-measured DBH, H, and extracted DBH from Backpack-LiDAR and H from UAS-LiDAR point cloud data. (a) DBH comparison, and (b) Height comparison.

Figure 8. LiDAR variables essential for AGB prediction (a) The importance rank of the variable based on random forest. (b) Pearson correlation between LiDAR variables.

Figure 9. Field-estimated Forest AGB (Mg/ha) versus predicted forest AGB (Mg/ha). (a) MLR model. (b) RF model (c) SVR model. The solid line represents the fitting model and the gray areas show the 95% confidence intervals of the fitting models.

Figure 10. The performance comparison of MLR, RF, and SVR, using R2, RMSE, and rRMSE, where R2 shows a random forest with the highest score of 89%.

Figure 11. Graphical representation of the results for tree segmentation. Where (a) illustrates graphical results of Backpack-LiDAR point cloud data segmentation, (b) shows the results of the UAS-LiDAR data based on the segmentation results using seed points, and (c) demonstrates the segmentation results of the UAS-LiDAR data without seed points. Note that each color in the above figure characterizes a different tree species, whereas the polygon areas overlapped on (b,c) refer to the distinct trees and/or tree crowns resulting from the graphical illustration.

Table 1

Summary of collected Zenmuse L1 UAS-LiDAR data sets for three sites.

ID	Survey Date	Altitude (m)	Frequency (kHz)	Area (m2)	Pts/m2	Total Pts (Million)
Plot 1	14 October 2023	80	240	27655	710	3.6
Plot 2	15 October 2023	80	240	46334	710	3.7
Plot 3	17 October 2023	80	240	51032	710	4.7

Table 2

Registration accuracy in the three plots.

Plot ID	Minimum Error (m)	Maximum Error (m)	Root Mean Square (m)
Plot 1 Lancai Township	0.662935	0.85483	0.762279
Plot 2 Qukuhu Towship	0.353413	0.760136	0.5704247
Plot 3 Maixiu Township	0.144129	0.276742	0.220785

Table 3

The parameters for field-estimated forest characteristics in the three plots.

Variables	Range	Mean	SD
Height (meters)	4.64–29.46	14.39	5.69
DBH (meters)	0.072–0.494	0.213	0.112
AGB (Mg/ha)	9.95–409.67	149.43	149.32

Table 4

Summary of UAS-LiDAR extracted metrics.

LiDAR Metrics	Metrics	Description
Height-related metrics	Percentile height ($H_1$, $H_5,H_{10},H_{20},H_{25},H_{50}$, $H_{75}$, $H_{95}$, $H_{99}$)	The percentiles of the height distributions (1st, 5th, 10th, 20th, 25th, 30th, 40th, 50th, 60th, 70th, 75th, 80th, 90th, 95th, 99th) of all points above 2 m
	MADmedian ($H_{MAD}$)	Median absolute deviation from the median
	Median of Height ($H_{median}$)	The median of the heights above 2 m of all points
	Mean of height ($H_{mean})$	The mean height above 2 m of all points
	The coefficient of variance of height ($H_{cv})$	The coefficient of variation of heights of all points above 2 m
	Kurtosis of heights ($H_{Kurtosis}$)	The kurtosis of the heights of all points above 2 m
	Interquartile distance of height ($H_{IQ})$	The Interquartile distance of height of all points above 2 m
	Variance of height ($H_{Varinace})$	The variance of the heights of all points above 2 m
	Absolute average deviation ${(H}_{AAD}$)	The absolute average deviation of the heights of all points above 2 m
	Standard deviation (${(H}_{std})$	The standard deviation of heights of all points above 2 m
	Maximum Height ${(H}_{max})$	The maximum height of all points above 2 m
	Skewness of heights ($H_{ske}$)	The skewness of the heights of all points above 2 m
Density-related metrics	Canopy return density ($D_1$, $D_2,D_3,D_4,D_5,D_6$, $D_7$, $D_8$, $D_9$)	The proportion of points above the quantiles (10th, 20th, 30th, 40th, 50th, 60th, 70th, 80th, 90th) to total number of points
Intensity-related metrics	Canopy cover above 2 m (CC)	Percentages of first returns above 2 m
	Intensity percentile ($I_1$, $I_5,I_{10},I_{20},I_{25},I_{30}$, $I_{40}$, $I_{50}$, $I_{60},I_{70},I_{75},I_{80},I_{90},I_{95},I_{99}$)	The percentiles of the cumulative intensities’ distributions (1st, 5th,10th, 20th, 30th, 40th, 50th, 60th, 70th, 80th, 90th, 95th, 99th) of all points above 2 m
Laser penetration rate	LP	Percentages of last returns above ground

Table 5

Accuracy assessment for the method based on Backpack-LiDAR.

Plot ID	Number of Trees	Tree Segmented	TP	FP	FN	r	p	F
Plot 1 Lancai Township	63	68	60	8	3	0.95	0.88	0.92
Plot 2 Qukuhu Township	98	105	95	10	3	0.97	0.90	0.94
Plot 3 Maixiu Township	44	48	41	7	3	0.93	0.85	0.89
Mean Value	68.33	73.67	65.	8.33	3.00	0.95	0.88	0.91

Table 6

Accuracy assessment for the ITS method based on UAS-LiDAR.

Plot ID	Number of Trees	Tree Segmented	TP	FP	FN	r	p	F
Plot 1 Lancai Township	63	59	58	1	5	0.92	0.98	0.95
Plot 2 Qukuhu Township	98	95	94	1	4	0.96	0.99	0.97
Plot 3 Maixiu Township	44	42	41	1	3	0.93	0.98	0.95
Mean Value	68.33	65.33	64.33	1.00	4.00	0.94	0.98	0.96

Table 7

The summary of linear predictive models and accuracy assessment results at the three different plots.

Model	Predictive Model	R2	RMSE (Mg/ha)	rRMSE (%)
MLR(AGB)	(−71.51 × elev_percentile_95th) + (1210.32 × density_metrics_4.) + (92.19 × elev_AIH_99th) + (−707.66 × density_metrics_3.) + (−21.48 × elev_max) + (−332.37 × density_metrics_2) + (−44.98 × elev_IQ) + (−0.34 × elev_kurtosis) + (−170.00 × density_metrics_1) + (2198.31 × density_metrics_9) + (45.110 × CanopyCover) + (12.77 × int_kurtosis) + (−1827.93 × density_metrics_6) + (23.30 × elev_percentile_70th).	0.77	48.18	41.68

Table 8

Accuracy assessment of UAS-LiDAR tree segmentation without seed points derived from Backpack-LiDAR point cloud data at the three plots.

Plot ID	Number of Trees	Tree Segmented	TP	FP	FN	r	p	F
Plot 1 Lancai Township	63	42	41	1	22	0.65	0.98	0.78
Plot 2 Qukuhu Township	98	105	88	17	10	0.90	0.84	0.87
Plot 3 Maixiu Township	44	48	32	16	12	0.73	0.67	0.70
Mean Value	68.33	65.00	53.6	11.33	14.67	0.76	0.83	0.78

Table 9

The summary of UAS-LiDAR optimal metrics for use in regression models to predict AGB.

LiDAR Metrics	Metrics	Description
Height-related metrics	Percentile height (H_70, H_95,H_99, AIH_99)	The percentiles of the height distributions of all above 2
	Interquartile distance of Height (H_IQ)	The Interquartile distance of height of all points above 2 m
	Kurtosis of heights (H_kurtosis)	The kurtosis of the heights of all points above 2 m
	Maximum heights (H_max)	Tree Height Maximum
Density-related metrics	Canopy return density (D_1, D_2, D_3, D_4, D_6, D_ (9))	The proportion of points above the quantiles to the total number of points
Intensity-related metric	Canopy cover above 2 m (CC)	Percentages of first returns above 2 m
	Intensity Kurtosis (Int_kurtosis)	The kurtosis of the heights of all points above 2 m

Table 10

Performance comparison of MLR, RF, and SVR regression model estimation accuracy statistics for prediction of AGB.

Method	R2 %	RMSE (Mg/ha)	rRMSE/(%)
MLR	77	48.18	41.68
SVR	82	45.09	39.00
RF	89	33.92	29.34

References

1. Urbazaev, M.; Thiel, C.; Cremer, F.; Dubayah, R.; Migliavacca, M.; Reichstein, M.; Schmullius, C. Estimation of forest aboveground biomass and uncertainties by integration of field measurements, airborne LiDAR, and SAR and optical satellite data in Mexico. Carbon Balance Manag.; 2018; 13, 5. [DOI: https://dx.doi.org/10.1186/s13021-018-0093-5] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/29468474]

2. Matasci, G.; Hermosilla, T.; Wulder, M.A.; White, J.C.; Coops, N.C.; Hobart, G.W.; Zald, H.S. Large-area mapping of Canadian boreal forest cover, height, biomass and other structural attributes using Landsat composites and lidar plots. Remote Sens. Environ.; 2018; 209, pp. 90-106. [DOI: https://dx.doi.org/10.1016/j.rse.2017.12.020]

3. Wallace, L.; Lucieer, A.; Malenovský, Z.; Turner, D.; Vopěnka, P. Assessment of forest structure using two UAV techniques: A comparison of airborne laser scanning and structure from motion (SfM) point clouds. Forests; 2016; 7, 62. [DOI: https://dx.doi.org/10.3390/f7030062]

4. Ma, T.; Zhang, C.; Ji, L.; Zuo, Z.; Beckline, M.; Hu, Y.; Li, X.; Xiao, X. Development of forest aboveground biomass estimation, its problems and future solutions: A review. Ecol. Indic.; 2024; 159, 111653. [DOI: https://dx.doi.org/10.1016/j.ecolind.2024.111653]

5. Paul, A.M.; Alfred, J.U. Quantifying Carbon Sequestration in Woody Trees of Makurdi Zoological Garden, Benue State, Nigeria: A Non-Destructive Approach. UMYU Sci.; 2024; 3, pp. 223-234. [DOI: https://dx.doi.org/10.56919/usci.2432.024]

6. Rodda, S.R.; Nidamanuri, R.R.; Mayamanikandan, T.; Rajashekar, G.; Jha, C.S.; Dadhwal, V.K. Non-destructive allometric modeling for tree volume estimation in tropical dry deciduous forests of India using terrestrial laser scanner. J. Indian Soc. Remote Sens.; 2024; 52, pp. 825-839. [DOI: https://dx.doi.org/10.1007/s12524-022-01664-y]

7. Adhikari, S.; Ma, Q.; Poudel, K.; Renninger, H.J. Aboveground woody biomass estimation of young bioenergy plantations of Populus and its hybrids using mobile (backpack) LiDAR remote sensing. Trees For. People; 2024; 18, 100665. [DOI: https://dx.doi.org/10.1016/j.tfp.2024.100665]

8. Fayolle, A.; Ngomanda, A.; Mbasi, M.; Barbier, N.; Bocko, Y.; Boyemba, F.; Couteron, P.; Fonton, N.; Kamdem, N.; Katembo, J. A regional allometry for the Congo basin forests based on the largest ever destructive sampling. For. Ecol. Manag.; 2018; 430, pp. 228-240. [DOI: https://dx.doi.org/10.1016/j.foreco.2018.07.030]

9. Zhu, X.; Chen, X.; Ma, L.; Liu, W. UAV and Satellite Synergies for Mapping Grassland Aboveground Biomass in Hulunbuir Meadow Steppe. Plants; 2024; 13, 1006. [DOI: https://dx.doi.org/10.3390/plants13071006] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/38611535]

10. Kganyago, M.; Adjorlolo, C.; Mhangara, P.; Tsoeleng, L. Optical remote sensing of crop biophysical and biochemical parameters: An overview of advances in sensor technologies and machine learning algorithms for precision agriculture. Comput. Electron. Agric.; 2024; 218, 108730. [DOI: https://dx.doi.org/10.1016/j.compag.2024.108730]

11. Gura, D.; Berkova, E.; Panyutischeva, A.; Zaidi, M.M.; Turk, G. Overview of the DJI ZENMUSE L1 Laser Scanner for Spatial Planning Purposes. Proceedings of the International Scientific and Practical Conference Digital and Information Technologies in Economics and Management; Dushanbe, Tajikistan, 21–23 November 2022; pp. 32-41.

12. Puletti, N.; Grotti, M.; Ferrara, C.; Chianucci, F. Lidar-based estimates of aboveground biomass through ground, aerial, and satellite observation: A case study in a Mediterranean forest. J. Appl. Remote Sens.; 2020; 14, 044501. [DOI: https://dx.doi.org/10.1117/1.JRS.14.044501]

13. Ahmed, O.S.; Franklin, S.E.; Wulder, M.A.; White, J.C. Characterizing stand-level forest canopy cover and height using Landsat time series, samples of airborne LiDAR, and the Random Forest algorithm. ISPRS J. Photogramm. Remote Sens.; 2015; 101, pp. 89-101. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2014.11.007]

14. Mielcarek, M.; Stereńczak, K.; Khosravipour, A. Testing and evaluating different LiDAR-derived canopy height model generation methods for tree height estimation. Int. J. Appl. Earth Obs. Geoinf.; 2018; 71, pp. 132-143. [DOI: https://dx.doi.org/10.1016/j.jag.2018.05.002]

15. Guo, Q.; Su, Y.; Hu, T.; Guan, H.; Jin, S.; Zhang, J.; Zhao, X.; Xu, K.; Wei, D.; Kelly, M. Lidar boosts 3D ecological observations and modelings: A review and perspective. IEEE Geosci. Remote Sens. Mag.; 2020; 9, pp. 232-257. [DOI: https://dx.doi.org/10.1109/MGRS.2020.3032713]

16. Li, M.; Im, J.; Quackenbush, L.J.; Liu, T. Forest biomass and carbon stock quantification using airborne LiDAR data: A case study over Huntington Wildlife Forest in the Adirondack Park. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.; 2014; 7, pp. 3143-3156. [DOI: https://dx.doi.org/10.1109/JSTARS.2014.2304642]

17. Lu, J.; Wang, H.; Qin, S.; Cao, L.; Pu, R.; Li, G.; Sun, J. Estimation of aboveground biomass of Robinia pseudoacacia forest in the Yellow River Delta based on UAV and Backpack LiDAR point clouds. Int. J. Appl. Earth Obs. Geoinf.; 2020; 86, 102014. [DOI: https://dx.doi.org/10.1016/j.jag.2019.102014]

18. Zhao, L.; Zhou, W.; Peng, Y.; Hu, Y.; Ma, T.; Xie, Y.; Wang, L.; Liu, J.; Liu, Z. A new AG-AGB estimation model based on MODIS and SRTM data in Qinghai Province, China. Ecol. Indic.; 2021; 133, 108378. [DOI: https://dx.doi.org/10.1016/j.ecolind.2021.108378]

19. Daba, D.E.; Soromessa, T. The accuracy of species-specific allometric equations for estimating aboveground biomass in tropical moist montane forests: Case study of Albizia grandibracteata and Trichilia dregeana. Carbon Balance Manag.; 2019; 14, 18. [DOI: https://dx.doi.org/10.1186/s13021-019-0134-8] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31858282]

20. Fareed, N.; Numata, I. Evaluating the impact of field-measured tree height errors correction on aboveground biomass modeling using airborne laser scanning and GEDI datasets in Brazilian Amazonia. Trees For. People; 2025; 19, 100751. [DOI: https://dx.doi.org/10.1016/j.tfp.2024.100751]

21. He, H.; Zhang, C.; Zhao, X.; Fousseni, F.; Wang, J.; Dai, H.; Yang, S.; Zuo, Q. Allometric biomass equations for 12 tree species in coniferous and broadleaved mixed forests, Northeastern China. PLoS ONE; 2018; 13, e0186226. [DOI: https://dx.doi.org/10.1371/journal.pone.0186226] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/29351291]

22. Liu, L.; Lim, S.; Shen, X.; Yebra, M. A hybrid method for segmenting individual trees from airborne lidar data. Comput. Electron. Agric.; 2019; 163, 104871. [DOI: https://dx.doi.org/10.1016/j.compag.2019.104871]

23. Zhen, Z.; Quackenbush, L.J.; Zhang, L. Trends in Automatic Individual Tree Crown Detection and Delineation—Evolution of LiDAR Data. Remote Sens.; 2016; 8, 333. [DOI: https://dx.doi.org/10.3390/rs8040333]

24. Dai, W.; Yang, B.; Dong, Z.; Shaker, A. A new method for 3D individual tree extraction using multispectral airborne LiDAR point clouds. ISPRS J. Photogramm. Remote Sens.; 2018; 144, pp. 400-411. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2018.08.010]

25. Lu, X.; Guo, Q.; Li, W.; Flanagan, J. A bottom-up approach to segment individual deciduous trees using leaf-off lidar point cloud data. ISPRS J. Photogramm. Remote Sens.; 2014; 94, pp. 1-12. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2014.03.014]

26. Yang, J.; Kang, Z.; Cheng, S.; Yang, Z.; Akwensi, P.H. An individual tree segmentation method based on watershed algorithm and three-dimensional spatial distribution analysis from airborne LiDAR point clouds. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.; 2020; 13, pp. 1055-1067. [DOI: https://dx.doi.org/10.1109/JSTARS.2020.2979369]

27. Reitberger, J.; Schnörr, C.; Krzystek, P.; Stilla, U. 3D segmentation of single trees exploiting full waveform LIDAR data. ISPRS J. Photogramm. Remote Sens.; 2009; 64, pp. 561-574. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2009.04.002]

28. Li, W.; Guo, Q.; Jakubowski, M.K.; Kelly, M. A new method for segmenting individual trees from the lidar point cloud. Photogramm. Eng. Remote Sens.; 2012; 78, pp. 75-84. [DOI: https://dx.doi.org/10.14358/PERS.78.1.75]

29. Donager, J.J.; Sánchez Meador, A.J.; Blackburn, R.C. Adjudicating perspectives on forest structure: How do airborne, terrestrial, and mobile lidar-derived estimates compare?. Remote Sens.; 2021; 13, 2297. [DOI: https://dx.doi.org/10.3390/rs13122297]

30. Soraya, E.; Umarhadi, D.A.; Wardhana, W.; Senawi, F.A.; Ardiansyah, F. Aerial imagery and UAV-LiDAR data fusion for quantifying aboveground carbon stock of teak plantation. Ecol. Eng. Environ. Technol.; 2025; 26, pp. 178-192. [DOI: https://dx.doi.org/10.12912/27197050/197151] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/24466494]

31. Thinley, J.; Pickering, C.; Ndehedehe, C. Modelling above ground biomass for a mixed-tree urban arboretum forest based on a LiDAR-derived canopy height model and field-sampled data. Geomatica; 2025; 77, 100047. [DOI: https://dx.doi.org/10.1016/j.geomat.2025.100047]

32. Zhang, Z.; Cao, L.; She, G. Estimating forest structural parameters using canopy metrics derived from airborne LiDAR data in subtropical forests. Remote Sens.; 2017; 9, 940. [DOI: https://dx.doi.org/10.3390/rs9090940]

33. Cao, L.; Coops, N.C.; Innes, J.L.; Sheppard, S.R.; Fu, L.; Ruan, H.; She, G. Estimation of forest biomass dynamics in subtropical forests using multi-temporal airborne LiDAR data. Remote Sens. Environ.; 2016; 178, pp. 158-171. [DOI: https://dx.doi.org/10.1016/j.rse.2016.03.012]

34. Dandois, J.P.; Olano, M.; Ellis, E.C. Optimal altitude, overlap, and weather conditions for computer vision UAV estimates of forest structure. Remote Sens.; 2015; 7, pp. 13895-13920. [DOI: https://dx.doi.org/10.3390/rs71013895]

35. Sankey, T.; Donager, J.; McVay, J.; Sankey, J.B. UAV lidar and hyperspectral fusion for forest monitoring in the southwestern USA. Remote Sens. Environ.; 2017; 195, pp. 30-43. [DOI: https://dx.doi.org/10.1016/j.rse.2017.04.007]

36. Anderson, K.; Gaston, K.J. Lightweight unmanned aerial vehicles will revolutionize spatial ecology. Front. Ecol. Environ.; 2013; 11, pp. 138-146. [DOI: https://dx.doi.org/10.1890/120150]

37. Chen, S.; Nian, Y.; He, Z.; Che, M. Measuring the Tree Height of Picea crassifolia in Alpine Mountain Forests in Northwest China Based on UAV-LiDAR. Forests; 2022; 13, 1163. [DOI: https://dx.doi.org/10.3390/f13081163]

38. Gonzalez, P.; Asner, G.P.; Battles, J.J.; Lefsky, M.A.; Waring, K.M.; Palace, M. Forest carbon densities and uncertainties from Lidar, QuickBird, and field measurements in California. Remote Sens. Environ.; 2010; 114, pp. 1561-1575. [DOI: https://dx.doi.org/10.1016/j.rse.2010.02.011]

39. Brede, B.; Bartholomeus, H.M.; Barbier, N.; Pimont, F.; Vincent, G.; Herold, M. Peering through the thicket: Effects of UAV LiDAR scanner settings and flight planning on canopy volume discovery. Int. J. Appl. Earth Obs. Geoinf.; 2022; 114, 103056. [DOI: https://dx.doi.org/10.1016/j.jag.2022.103056]

40. Hyyppä, J.; Yu, X.; Hakala, T.; Kaartinen, H.; Kukko, A.; Hyyti, H.; Muhojoki, J.; Hyyppä, E. Under-canopy UAV laser scanning providing canopy height and stem volume accurately. Forests; 2021; 12, 856. [DOI: https://dx.doi.org/10.3390/f12070856]

41. Zhang, B.; Fischer, F.J.; Prober, S.M.; Yeoh, P.B.; Gosper, C.R.; Zdunic, K.; Jucker, T. Robust retrieval of forest canopy structural attributes using multi-platform airborne LiDAR. Remote Sens. Ecol. Conserv.; 2024; 10, pp. 725-742. [DOI: https://dx.doi.org/10.1002/rse2.398]

42. Roşca, S.; Suomalainen, J.; Bartholomeus, H.; Herold, M. Comparing terrestrial laser scanning and unmanned aerial vehicle structure from motion to assess top of canopy structure in tropical forests. Interface Focus; 2018; 8, 20170038. [DOI: https://dx.doi.org/10.1098/rsfs.2017.0038] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/29503719]

43. Brede, B.; Lau, A.; Bartholomeus, H.M.; Kooistra, L. Comparing RIEGL RiCOPTER UAV LiDAR derived canopy height and DBH with terrestrial LiDAR. Sensors; 2017; 17, 2371. [DOI: https://dx.doi.org/10.3390/s17102371] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/29039755]

44. Liu, K.; Shen, X.; Cao, L.; Wang, G.; Cao, F. Estimating forest structural attributes using UAV-LiDAR data in Ginkgo plantations. ISPRS J. Photogramm. Remote Sens.; 2018; 146, pp. 465-482. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2018.11.001]

45. Wieser, M.; Mandlburger, G.; Hollaus, M.; Otepka, J.; Glira, P.; Pfeifer, N. A case study of UAS borne laser scanning for measurement of tree stem diameter. Remote Sens.; 2017; 9, 1154. [DOI: https://dx.doi.org/10.3390/rs9111154]

46. Tran, G.; Nguyen, D.; Milenkovic, M.; Pfeifer, N. Potential of full waveform airborne laser scanning data for urban area classification–transfer of classification approaches between missions. Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci.; 2015; 40, pp. 1317-1323. [DOI: https://dx.doi.org/10.5194/isprsarchives-XL-7-W3-1317-2015]

47. Treitz, P.; Lim, K.; Woods, M.; Pitt, D.; Nesbitt, D.; Etheridge, D. LiDAR sampling density for forest resource inventories in Ontario, Canada. Remote Sens.; 2012; 4, pp. 830-848. [DOI: https://dx.doi.org/10.3390/rs4040830]

48. Wu, C.; Shen, H.; Shen, A.; Deng, J.; Gan, M.; Zhu, J.; Xu, H.; Wang, K. Comparison of machine-learning methods for above-ground biomass estimation based on Landsat imagery. J. Appl. Remote Sens.; 2016; 10, 035010. [DOI: https://dx.doi.org/10.1117/1.JRS.10.035010]

49. Maxwell, A.E.; Warner, T.A.; Fang, F. Implementation of machine-learning classification in remote sensing: An applied review. Int. J. Remote Sens.; 2018; 39, pp. 2784-2817. [DOI: https://dx.doi.org/10.1080/01431161.2018.1433343]

50. Zhang, Y.; Ma, J.; Liang, S.; Li, X.; Li, M. An evaluation of eight machine learning regression algorithms for forest aboveground biomass estimation from multiple satellite data products. Remote Sens.; 2020; 12, 4015. [DOI: https://dx.doi.org/10.3390/rs12244015]

51. Filippi, A.M.; Güneralp, İ.; Randall, J. Hyperspectral remote sensing of aboveground biomass on a river meander bend using multivariate adaptive regression splines and stochastic gradient boosting. Remote Sens. Lett.; 2014; 5, pp. 432-441. [DOI: https://dx.doi.org/10.1080/2150704X.2014.915070]

52. Sainuddin, F.V.; Malek, G.; Rajwadi, A.; Nagar, P.S.; Asok, S.V.; Reddy, C.S. Estimating above-ground biomass of the regional forest landscape of northern Western Ghats using machine learning algorithms and multi-sensor remote sensing data. J. Indian Soc. Remote Sens.; 2024; 52, pp. 885-902. [DOI: https://dx.doi.org/10.1007/s12524-024-01836-y]

53. Suchenwirth, L.; Stümer, W.; Schmidt, T.; Förster, M.; Kleinschmit, B. Large-scale mapping of carbon stocks in riparian forests with self-organizing maps and the k-nearest-neighbor algorithm. Forests; 2014; 5, pp. 1635-1652. [DOI: https://dx.doi.org/10.3390/f5071635]

54. Le, Q.; Boydell, O.; Mac Namee, B.; Scanlon, M. Deep learning at the shallow end: Malware classification for non-domain experts. Digit. Investig.; 2018; 26, pp. S118-S126. [DOI: https://dx.doi.org/10.1016/j.diin.2018.04.024]

55. Wang, L.; Yan, J.; Mu, L.; Huang, L. Knowledge discovery from remote sensing images: A review. Wiley Interdiscip. Rev. Data Min. Knowl. Discov.; 2020; 10, e1371. [DOI: https://dx.doi.org/10.1002/widm.1371]

56. Yang, X.; Ye, Y.; Li, X.; Lau, R.Y.; Zhang, X.; Huang, X. Hyperspectral image classification with deep learning models. IEEE Trans. Geosci. Remote Sens.; 2018; 56, pp. 5408-5423. [DOI: https://dx.doi.org/10.1109/TGRS.2018.2815613]

57. Yuan, Q.; Shen, H.; Li, T.; Li, Z.; Li, S.; Jiang, Y.; Xu, H.; Tan, W.; Yang, Q.; Wang, J. Deep learning in environmental remote sensing: Achievements and challenges. Remote Sens. Environ.; 2020; 241, 111716. [DOI: https://dx.doi.org/10.1016/j.rse.2020.111716]

58. Chu, D. Aboveground biomass estimates of grassland in the North Tibet using MODIS remote sensing approaches. Appl. Ecol. Environ. Res.; 2020; 18, pp. 7655-7672. [DOI: https://dx.doi.org/10.15666/aeer/1806_76557672]

59. Liu, S.; Cheng, F.; Dong, S.; Zhao, H.; Hou, X.; Wu, X. Spatiotemporal dynamics of grassland aboveground biomass on the Qinghai-Tibet Plateau based on validated MODIS NDVI. Sci. Rep.; 2017; 7, 4182. [DOI: https://dx.doi.org/10.1038/s41598-017-04038-4] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28646198]

60. Zhang, H.; Tang, Z.; Wang, B.; Kan, H.; Sun, Y.; Qin, Y.; Meng, B.; Li, M.; Chen, J.; Lv, Y. A 250m annual alpine grassland AGB dataset over the Qinghai-Tibetan Plateau (2000–2019) based on in-situ measurements, UAV images, and MODIS Data. Earth Syst. Sci. Data Discuss.; 2022; 2022, pp. 821-846.

61. Chen, A.; Wang, X.; Zhang, M.; Guo, J.; Xing, X.; Yang, D.; Zhang, H.; Hou, Z.; Jia, Z.; Yang, X. Fusion of LiDAR and Multispectral Data for Aboveground Biomass Estimation in Mountain Grassland. Remote Sens.; 2023; 15, 405. [DOI: https://dx.doi.org/10.3390/rs15020405]

62. Wang, Z.; Yi, L.; Xu, W.; Zheng, X.; Xiong, S.; Bao, A. Integration of UAV and GF-2 Optical Data for Estimating Aboveground Biomass in Spruce Plantations in Qinghai, China. Sustainability; 2023; 15, 9700. [DOI: https://dx.doi.org/10.3390/su15129700]

63. Roussel, J.-R.; Caspersen, J.; Béland, M.; Thomas, S.; Achim, A. Removing bias from LiDAR-based estimates of canopy height: Accounting for the effects of pulse density and footprint size. Remote Sens. Environ.; 2017; 198, pp. 1-16. [DOI: https://dx.doi.org/10.1016/j.rse.2017.05.032]

64. Luo, Z.; Wu, W.; Yu, X.; Song, Q.; Yang, J.; Wu, J.; Zhang, H. Variation of net primary production and its correlation with climate change and anthropogenic activities over the Tibetan Plateau. Remote Sens.; 2018; 10, 1352. [DOI: https://dx.doi.org/10.3390/rs10091352]

65. Gao, Q.; Guo, Y.; Xu, H.; Ganjurjav, H.; Li, Y.; Wan, Y.; Qin, X.; Ma, X.; Liu, S. Climate change and its impacts on vegetation distribution and net primary productivity of the alpine ecosystem in the Qinghai-Tibetan Plateau. Sci. Total Environ.; 2016; 554, pp. 34-41. [DOI: https://dx.doi.org/10.1016/j.scitotenv.2016.02.131]

66. Fu, S.; Zhang, X.; Kuang, W.; Guo, C. Characteristics of Changes in Urban Land Use and Efficiency Evaluation in the Qinghai–Tibet Plateau from 1990 to 2020. Land; 2022; 11, 757. [DOI: https://dx.doi.org/10.3390/land11050757]

67. Štroner, M.; Urban, R.; Línková, L. A new method for UAV lidar precision testing used for the evaluation of an affordable DJI ZENMUSE L1 scanner. Remote Sens.; 2021; 13, 4811. [DOI: https://dx.doi.org/10.3390/rs13234811]

68. Zhou, C.; Ye, H.; Sun, D.; Yue, J.; Yang, G.; Hu, J. An automated, high-performance approach for detecting and characterizing broccoli based on UAV remote-sensing and transformers: A case study from Haining, China. Int. J. Appl. Earth Obs. Geoinf.; 2022; 114, 103055. [DOI: https://dx.doi.org/10.1016/j.jag.2022.103055]

69. St. Peter, J.; Drake, J.; Medley, P.; Ibeanusi, V. Forest structural estimates derived using a practical, open-source lidar-processing workflow. Remote Sens.; 2021; 13, 4763. [DOI: https://dx.doi.org/10.3390/rs13234763]

70. Diara, F.; Roggero, M. Quality Assessment of DJI Zenmuse L1 and P1 LiDAR and Photogrammetric Systems: Metric and Statistics Analysis with the Integration of Trimble SX10 Data. Geomatics; 2022; 2, pp. 254-281. [DOI: https://dx.doi.org/10.3390/geomatics2030015]

71. Guan, T.; Shen, Y.; Wang, Y.; Zhang, P.; Wang, R.; Yan, F. Advancing Forest Plot Surveys: A Comparative Study of Visual vs. LiDAR SLAM Technologies. Forests; 2024; 15, 2083. [DOI: https://dx.doi.org/10.3390/f15122083]

72. Wang, H.; Pu, R.; Zhu, Q.; Ren, L.; Zhang, Z. Mapping health levels of Robinia pseudoacacia forests in the Yellow River delta, China, using IKONOS and Landsat 8 OLI imagery. Int. J. Remote Sens.; 2015; 36, pp. 1114-1135. [DOI: https://dx.doi.org/10.1080/01431161.2015.1007253]

73. Tao, S.; Wu, F.; Guo, Q.; Wang, Y.; Li, W.; Xue, B.; Hu, X.; Li, P.; Tian, D.; Li, C. et al. Segmenting tree crowns from terrestrial and mobile LiDAR data by exploring ecological theories. ISPRS J. Photogramm. Remote Sens.; 2015; 110, pp. 66-76. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2015.10.007]

74. Chen, P.; Shi, W.; Bao, S.; Wang, M.; Fan, W.; Xiang, H. Low-drift odometry, mapping and ground segmentation using a backpack LiDAR system. IEEE Robot. Autom. Lett.; 2021; 6, pp. 7285-7292. [DOI: https://dx.doi.org/10.1109/LRA.2021.3097060]

75. Zhao, X.; Guo, Q.; Su, Y.; Xue, B. Improved progressive TIN densification filtering algorithm for airborne LiDAR data in forested areas. ISPRS J. Photogramm. Remote Sens.; 2016; 117, pp. 79-91. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2016.03.016]

76. Polewski, P.; Yao, W.; Cao, L.; Gao, S. Marker-free coregistration of UAV and backpack LiDAR point clouds in forested areas. ISPRS J. Photogramm. Remote Sens.; 2019; 147, pp. 307-318. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2018.11.020]

77. Wu, J.; Cawse-Nicholson, K.; van Aardt, J. 3D Tree reconstruction from simulated small footprint waveform lidar. Photogramm. Eng. Remote Sens.; 2013; 79, pp. 1147-1157. [DOI: https://dx.doi.org/10.14358/PERS.79.12.1147]

78. Ruhan, A.; Du, W.; Ying, H.; Wei, B.; Shan, Y.; Dai, H. Estimation of Aboveground Biomass of Individual Trees by Backpack LiDAR Based on Parameter-Optimized Quantitative Structural Models (AdQSM). Forests; 2023; 14, 475. [DOI: https://dx.doi.org/10.3390/f14030475]

79. Goutte, C.; Gaussier, E. A probabilistic interpretation of precision, recall and F-score, with implication for evaluation. Proceedings of the European Conference on Information Retrieval; Santiago de Compostela, Spain, 21–23 March 2005; pp. 345-359.

80. Guoyi, Z.; Yin, G.; Xuli, T. Carbon Storage of China’s Forest Ecosystems—Biomass Equations (中国森林生态系统碳储量—生物量方程); National Publication Foundation: Beijing, China, 2018; (In Chinese)

81. Hao, H.; Li, W.; Zhao, X.; Chang, Q.; Zhao, P. Estimating the Aboveground Carbon Density of Coniferous Forests by Combining Airborne LiDAR and Allometry Models at Plot Level. Front. Plant Sci.; 2019; 10, 917. [DOI: https://dx.doi.org/10.3389/fpls.2019.00917] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31354780]

82. Hyyppä, E.; Yu, X.; Kaartinen, H.; Hakala, T.; Kukko, A.; Vastaranta, M.; Hyyppä, J. Comparison of Backpack, Handheld, Under-Canopy UAV, and Above-Canopy UAV Laser Scanning for Field Reference Data Collection in Boreal Forests. Remote Sens.; 2020; 12, 3327. [DOI: https://dx.doi.org/10.3390/rs12203327]

83. Donev, P.; Wang, H.; Qin, S.; Meng, P.; Lu, J. Estimating the forest above-ground biomass based on extracted LiDAR metrics and predicted diameter at breast height. J. Geod. Geoinf. Sci.; 2021; 4, 13.

84. García, M.; Saatchi, S.; Ustin, S.; Balzter, H. Modelling forest canopy height by integrating airborne LiDAR samples with satellite Radar and multispectral imagery. Int. J. Appl. Earth Obs. Geoinf.; 2018; 66, pp. 159-173. [DOI: https://dx.doi.org/10.1016/j.jag.2017.11.017]

85. Lim, K.; Treitz, P.; Wulder, M.; St-Onge, B.; Flood, M. LiDAR remote sensing of forest structure. Prog. Phys. Geogr. Earth Environ.; 2016; 27, pp. 88-106. [DOI: https://dx.doi.org/10.1191/0309133303pp360ra]

86. Persson, H.J.; Olofsson, K.; Holmgren, J. Two-phase forest inventory using very-high-resolution laser scanning. Remote Sens. Environ.; 2022; 271, 112909. [DOI: https://dx.doi.org/10.1016/j.rse.2022.112909]

87. Srinivasan, S.; Popescu, S.C.; Eriksson, M.; Sheridan, R.D.; Ku, N.-W. Terrestrial laser scanning as an effective tool to retrieve tree level height, crown width, and stem diameter. Remote Sens.; 2015; 7, pp. 1877-1896. [DOI: https://dx.doi.org/10.3390/rs70201877]

88. Sun, Z.; Wang, Y.-f.; Ding, Z.-d.; Liang, R.-t.; Xie, Y.-h.; Li, R.; Li, H.-w.; Pan, L.; Sun, Y.-j. Individual tree segmentation and biomass estimation based on UAV Digital aerial photograph. J. Mt. Sci.; 2023; 20, pp. 724-737. [DOI: https://dx.doi.org/10.1007/s11629-022-7563-7]

89. Fassnacht, F.; Hartig, F.; Latifi, H.; Berger, C.; Hernández, J.; Corvalán, P.; Koch, B. Importance of sample size, data type and prediction method for remote sensing-based estimations of aboveground forest biomass. Remote Sens. Environ.; 2014; 154, pp. 102-114. [DOI: https://dx.doi.org/10.1016/j.rse.2014.07.028]

90. Næsset, E.; Bollandsås, O.M.; Gobakken, T. Comparing regression methods in estimation of biophysical properties of forest stands from two different inventories using laser scanner data. Remote Sens. Environ.; 2005; 94, pp. 541-553. [DOI: https://dx.doi.org/10.1016/j.rse.2004.11.010]

91. Breiman, L. Random forests. Mach. Learn.; 2001; 45, pp. 5-32. [DOI: https://dx.doi.org/10.1023/A:1010933404324]

92. Mutanga, O.; Adam, E.; Cho, M.A. High density biomass estimation for wetland vegetation using WorldView-2 imagery and random forest regression algorithm. Int. J. Appl. Earth Obs. Geoinf.; 2012; 18, pp. 399-406. [DOI: https://dx.doi.org/10.1016/j.jag.2012.03.012]

93. Yu, X.; Hyyppä, J.; Vastaranta, M.; Holopainen, M.; Viitala, R. Predicting individual tree attributes from airborne laser point clouds based on the random forests technique. ISPRS J. Photogramm. Remote Sens.; 2011; 66, pp. 28-37. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2010.08.003]

94. Hsu, C.-W.; Chang, C.-C.; Lin, C.-J. *A Practical Guide to Support Vector Classification*. Technical Report, Department of Computer Science, National Taiwan University. 2003; Available online: https://www.csie.ntu.edu.tw/~cjlin/papers/guide/guide.pdf (accessed on 10 February 2025).

95. Dong, L.; Qingsong, G.; Wenjiang, H.; Linsheng, H.; Guijun, Y. Remote sensing inversion of leaf area index based on support vector machine regression in winter wheat. Trans. Chin. Soc. Agric. Eng.; 2013; 29, pp. 117-123. (In Chinese) [DOI: https://dx.doi.org/10.3969/j.issn.1002-6819.2013.07.015]

96. Wu, J.; Yao, W.; Choi, S.; Park, T.; Myneni, R.B. A Comparative Study of Predicting DBH and Stem Volume of Individual Trees in a Temperate Forest Using Airborne Waveform LiDAR. IEEE Geosci. Remote Sens. Lett.; 2015; 12, pp. 2267-2271. [DOI: https://dx.doi.org/10.1109/LGRS.2015.2466464]

97. Awad, M.; Khanna, R. Efficient Learning Machines: Theories, Concepts, and Applications for Engineers and System Designers; Springer Nature: Berlin/Heidelberg, Germany, 2015.

98. Yukun, G.; Lu, D.; Li, G.; Wang, G.; Chen, Q.; Liu, L.; Li, D. Comparative Analysis of Modeling Algorithms for Forest Aboveground Biomass Estimation in a Subtropical Region. Remote Sens.; 2018; 10, 627. [DOI: https://dx.doi.org/10.3390/rs10040627]

99. Schölkopf, B.; Smola, A.J. Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond; MIT Press: Cambridge, MA, USA, 2002.

100. Prasetyo, W.E.; Handayani, H.H.; Raharjo, A.B.; Saptarini, D. Advancing Carbon Stock Estimation and 3D Tree Modeling: Harnessing the Potential of Low-Cost Backpack LiDAR Technology. IOP Conf. Ser. Earth Environ. Sci.; 2024; 1406, 012013. [DOI: https://dx.doi.org/10.1088/1755-1315/1406/1/012013]

101. Tian, L.; Wu, X.; Tao, Y.; Li, M.; Qian, C.; Liao, L.; Fu, W. Review of Remote Sensing-Based Methods for Forest Aboveground Biomass Estimation: Progress, Challenges, and Prospects. Forests; 2023; 14, 1086. [DOI: https://dx.doi.org/10.3390/f14061086]

102. Zhu, X.; He, Z.; Chen, L.; Du, J.; Yang, J.; Lin, P.; Li, J. Changes in Species Diversity, Aboveground Biomass, and Distribution Characteristics along an Afforestation Successional Gradient in Semiarid Picea crassifolia Plantations of Northwestern China. For. Sci.; 2016; 63, pp. 17-28. [DOI: https://dx.doi.org/10.5849/forsci.16-041]

103. Morais, T.G.; Teixeira, R.F.; Figueiredo, M.; Domingos, T. The use of machine learning methods to estimate aboveground biomass of grasslands: A review. Ecol. Indic.; 2021; 130, 108081. [DOI: https://dx.doi.org/10.1016/j.ecolind.2021.108081]

104. Mokroš, M.; Mikita, T.; Singh, A.; Tomaštík, J.; Chudá, J.; Wężyk, P.; Kuželka, K.; Surový, P.; Klimánek, M.; Zięba-Kulawik, K. Novel low-cost mobile mapping systems for forest inventories as terrestrial laser scanning alternatives. Int. J. Appl. Earth Obs. Geoinf.; 2021; 104, 102512. [DOI: https://dx.doi.org/10.1016/j.jag.2021.102512]

105. Yao, W.; Krzystek, P.; Heurich, M. Tree species classification and estimation of stem volume and DBH based on single tree extraction by exploiting airborne full-waveform LiDAR data. Remote Sens. Environ.; 2012; 123, pp. 368-380. [DOI: https://dx.doi.org/10.1016/j.rse.2012.03.027]

106. Heo, H.K.; Lee, D.K.; Park, J.H.; Thorne, J.H. Estimating the heights and diameters at breast height of trees in an urban park and along a street using mobile LiDAR. Landsc. Ecol. Eng.; 2019; 15, pp. 253-263. [DOI: https://dx.doi.org/10.1007/s11355-019-00379-6]

107. Li, Y.; Li, M.; Li, C.; Liu, Z. Forest aboveground biomass estimation using Landsat 8 and Sentinel-1A data with machine learning algorithms. Sci. Rep.; 2020; 10, 9952. [DOI: https://dx.doi.org/10.1038/s41598-020-67024-3] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/32561836]

108. Muhammad, B.; Rehman, A.U.R.; Mumtaz, F.; Qun, Y.; Zhongkui, J. Estimation of above-ground biomass in dry temperate forests using Sentinel-2 data and random forest: A case study of the Swat area of Pakistan. Front. Environ. Sci.; 2024; 12, 1448648. [DOI: https://dx.doi.org/10.3389/fenvs.2024.1448648]

109. Luo, S.; Jiang, X.; He, Y.; Li, J.; Jiao, W.; Zhang, S.; Xu, F.; Han, Z.; Sun, J.; Yang, J. et al. Multi-dimensional variables and feature parameter selection for aboveground biomass estimation of potato based on UAV multispectral imagery. Front. Plant Sci.; 2022; 13, 948249. [DOI: https://dx.doi.org/10.3389/fpls.2022.948249] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35968116]

110. Adame-Campos, R.L.; Ghilardi, A.; Gao, Y.; Paneque-Gálvez, J.; Mas, J.-F. Variables Selection for Aboveground Biomass Estimations Using Satellite Data: A Comparison between Relative Importance Approach and Stepwise Akaike’s Information Criterion. ISPRS Int. J. Geo-Inf.; 2019; 8, 245. [DOI: https://dx.doi.org/10.3390/ijgi8060245]

111. Su, H.; Shen, W.; Wang, J.; Ali, A.; Li, M. Machine learning and geostatistical approaches for estimating aboveground biomass in Chinese subtropical forests. For. Ecosyst.; 2020; 7, 64. [DOI: https://dx.doi.org/10.1186/s40663-020-00276-7]

112. Frazer, G.; Magnussen, S.; Wulder, M.; Niemann, K. Simulated impact of sample plot size and co-registration error on the accuracy and uncertainty of LiDAR-derived estimates of forest stand biomass. Remote Sens. Environ.; 2011; 115, pp. 636-649. [DOI: https://dx.doi.org/10.1016/j.rse.2010.10.008]

113. Zhang, L.; Zhao, Y.; Chen, C.; Li, X.; Mao, F.; Lv, L.; Yu, J.; Song, M.; Huang, L.; Chen, J. et al. UAV-LiDAR Integration with Sentinel-2 Enhances Precision in AGB Estimation for Bamboo Forests. Remote Sens.; 2024; 16, 705. [DOI: https://dx.doi.org/10.3390/rs16040705]

114. Ali, H.; Mohammadi, J.; Shataee Jouibary, S. Deep and machine learning prediction of forest above-ground biomass using multi-source remote sensing data in coniferous planted forests in Iran. Eur. J. For. Res.; 2024; 143, pp. 1731-1745. [DOI: https://dx.doi.org/10.1007/s10342-024-01721-w]

115. Hernández-Stefanoni, J.L.; Reyes-Palomeque, G.; Castillo-Santiago, M.Á.; George-Chacón, S.P.; Huechacona-Ruiz, A.H.; Tun-Dzul, F.; Rondon-Rivera, D.; Dupuy, J.M. Effects of Sample Plot Size and GPS Location Errors on Aboveground Biomass Estimates from LiDAR in Tropical Dry Forests. Remote Sens.; 2018; 10, 1586. [DOI: https://dx.doi.org/10.3390/rs10101586]

116. Lu, D.; Jiang, X. A brief overview and perspective of using airborne Lidar data for forest biomass estimation. Int. J. Image Data Fusion; 2024; 15, pp. 1-24. [DOI: https://dx.doi.org/10.1080/19479832.2024.2309615]

117. Schäfer, J.; Winiwarter, L.; Weiser, H.; Novotný, J.; Höfle, B.; Schmidtlein, S.; Henniger, H.; Krok, G.; Stereńczak, K.; Fassnacht, F.E. Assessing the potential of synthetic and ex situ airborne laser scanning and ground plot data to train forest biomass models. For. Int. J. For. Res.; 2024; 97, pp. 512-530. [DOI: https://dx.doi.org/10.1093/forestry/cpad061]

118. Su, R.; Du, W.; Shan, Y.; Ying, H.; Rihan, W.; Li, R. Aboveground Carbon Stock Estimation Based on Backpack LiDAR and UAV Multispectral Imagery at the Forest Sample Plot Scale. Remote Sens.; 2024; 16, 3927. [DOI: https://dx.doi.org/10.3390/rs16213927]