A hybrid method for forest aboveground biomass estimation: fusion of individual tree- and area-based approaches over northeast China

Abstract

Currently, large-scale above-ground biomass (AGB) estimation mainly depends on the area-based approach (ABA). However, with advancements in individual tree detection technology and the availability of multi-platform remote sensing data, the individual tree-based approach (ITA) presents new opportunities for accurate, nondestructive AGB estimation. Nevertheless, research on integrating ITA with ABA for scaling AGB estimates from individual trees to the regional level remains limited. This study introduced an innovative hybrid framework incorporating ITA and ABA for regional-scale estimation of forest AGB through the combination of handheld, unmanned, and satellite LiDAR, together with Landsat 8 imagery. The results demonstrated that the segmentation of individual trees using a fusion of unmanned aerial vehicle laser scanning (ULS) and handheld laser scanning (HLS) was effective (r = 0.84, p = 0.76, F-score = 0.79) for the secondary forests in northeast China (NEC), particularly achieving high accuracy of diameter at breast height (DBH) with R² = 0.93 and RMSE of only 1.84 cm. Both individual tree and plot-level AGB estimates achieved satisfactory accuracy, with biases of −3.9 kg and −16,56 Mg/ha, respectively. The error-in-variable (EIV) model was established for stand-level AGB estimation using the plot AGB estimates from ITA (R² = 0.69; RMSE = 19.59 Mg/ha; rRMSE = 15.1%) and employed to estimate footprint-level AGB based on the canopy height data extracted from ICESat-2/ATL08, significantly expanding the limited sample size (by 150 times) for regional AGB estimation. Two periods (2019–2020, 2021–2022) of continuous AGB of NEC were mapped using the integrated method of random forests (RF) and empirical Bayesian Kriging (EBK). The accuracy of the RF-EBK model is markedly enhanced in comparison to that of the RF model for AGB estimation (R² increased by about 37% − 49%, RMSE and rRMSE declined by 25%). This study provides technical support for upscaling AGB estimation from individual tree to extensive forest and lays a solid foundation for precise forestry, sustainable forest management, and the attainment of carbon neutrality.

Full text

A hybrid method for forest aboveground biomass estimation: fusion of individual
tree- and area-based approaches over northeast China
Zhen Zhen
a,b
, Xiang Li
a,b
, Ye Ma
c
, Yinghui Zhao
a,b
and Xiaochun Wang
a,d
a
Key Laboratory of Sustainable Forest Ecosystem Management-Ministry of Education, Northeast Forestry University, Harbin, P.R. China;
b
School of Forestry, Northeast Forestry University, Harbin, P.R. China;
c
Heilongjiang Province Key Laboratory of Geographical Environment
Monitoring and Spatial Information Service in Cold Regions, College of Geographical Sciences, Harbin Normal University, Harbin, China;
d
School of Ecology, Northeast Forestry University, Harbin, P.R. China
ABSTRACT
Currently, large-scale above-ground biomass (AGB) estimation mainly depends on the area-based
approach (ABA). However, with advancements in individual tree detection technology and the
availability of multi-platform remote sensing data, the individual tree-based approach (ITA) pre-
sents new opportunities for accurate, nondestructive AGB estimation. Nevertheless, research on
integrating ITA with ABA for scaling AGB estimates from individual trees to the regional level
remains limited. This study introduced an innovative hybrid framework incorporating ITA and ABA
for regional-scale estimation of forest AGB through the combination of handheld, unmanned, and
satellite LiDAR, together with Landsat 8 imagery. The results demonstrated that the segmentation
of individual trees using a fusion of unmanned aerial vehicle laser scanning (ULS) and handheld
laser scanning (HLS) was eective (r = 0.84, p = 0.76, F-score = 0.79) for the secondary forests in
northeast China (NEC), particularly achieving high accuracy of diameter at breast height (DBH) with
R
2
= 0.93 and RMSE of only 1.84 cm. Both individual tree and plot-level AGB estimates achieved
satisfactory accuracy, with biases of −3.9 kg and −16,56 Mg/ha, respectively. The error-in-variable
(EIV) model was established for stand-level AGB estimation using the plot AGB estimates from ITA
(R
2
= 0.69; RMSE = 19.59 Mg/ha; rRMSE = 15.1%) and employed to estimate footprint-level AGB
based on the canopy height data extracted from ICESat-2/ATL08, signicantly expanding the
limited sample size (by 150 times) for regional AGB estimation. Two periods (2019–2020,
2021–2022) of continuous AGB of NEC were mapped using the integrated method of random
forests (RF) and empirical Bayesian Kriging (EBK). The accuracy of the RF-EBK model is markedly
enhanced in comparison to that of the RF model for AGB estimation (R
2
increased by about 37% −
49%, RMSE and rRMSE declined by 25%). This study provides technical support for upscaling AGB
estimation from individual tree to extensive forest and lays a solid foundation for precise forestry,
sustainable forest management, and the attainment of carbon neutrality.
ARTICLE HISTORY
Received 11 December 2024
Accepted 18 April 2025
KEYWORDS
AGB; LiDAR; ICESat-2; error-in
-variance model; Bayesian
Kriging
1. Introduction
Forests are primarily vital to land-based ecosystems and
the worldwide carbon cycle because they are crucial for
capturing and storing carbon. As a key indicator of forest
carbon sinks and productivity, forest biomass is also
a core parameter for assessing forest carbon balance
(Simon et al. 2021). Forest biomass encompasses both
aboveground and belowground biomass, including
wood, woody debris, litter, and soil organic matter
(Satoo and Madgwick 2012), with AGB making up the
largest portion (Vicharnakorn et al. 2014). AGB can be
obtained through direct measurements or estimation
methods (Satoo and Madgwick 2012). Accurately and
eciently estimating large-scale AGB has been a focus
of sustainable forest management and planning,
providing strong support for assessing forest carbon
sequestration capacity and quantifying its role in the
global carbon cycle (Chen et al. 2019; Zhang et al. 2019).
Although traditional destructive sampling is the most
precise technique for assessing forest biomass, it is
resource-intensive, expensive, and impractical for exten-
sive areas, which limits its application in large-scale
research (Lu et al. 2016). An alternative approach to
estimating forest biomass involves using remote sensing
technologies. Remote sensing data, characterized by
timely acquisition, low cost, and non-invasiveness, have
become the primary data source for large-scale AGB
estimation (Forkuor et al. 2020; Liu et al. 2023; Liu et al.
2024a; Zhao et al. 2021). LiDAR, capable of penetrating
the canopy to capture the three-dimensional structure
CONTACT Yinghui Zhao yinghuizhao@nefu.edu.cn; Xiaochun Wang wangx@nefu.edu.cn
Supplemental data for this article can be accessed online at https://doi.org/10.1080/15481603.2025.2497629
GISCIENCE & REMOTE SENSING
2025, VOL. 62, NO. 1, 2497629
https://doi.org/10.1080/15481603.2025.2497629
© 2025 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group.
This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial License (http://creativecommons.org/licenses/by-nc/4.0/), which
permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited. The terms on which this article has been
published allow the posting of the Accepted Manuscript in a repository by the author(s) or with their consent.

of forest canopies, has seen widespread use in recent
years for forest AGB estimation (Cao et al. 2016).
According to platforms, LiDAR data can be classied
into spaceborne LiDAR such as the Ice, Cloud, and
Land Elevation Satellite (ICESat-2) and the Global
Ecosystem Dynamics Investigation (GEDI); airborne
laser scanning (ALS); unmanned aerial vehicle laser
scanning (ULS); mobile laser scanning (MLS); terres-
trial laser scanning (TLS); and handheld laser scanning
(HLS), among others. Spaceborne LiDAR systems are
capable of delivering a vast array of discrete footprint
data, making them a dependable resource for exten-
sive forest research. The ICESat-2 stands out as the
sole photon-counting LiDAR system with a high repe-
tition rate, and oers global geolocated photon data
(ATL03) with an impressive accuracy (horizontal: 3.6
m, vertical: 0.1 m) (Magruder, Neumann, and Kurtz
2021; Markus et al. 2017). This level of accuracy sig-
nicantly outperforms the GEDI system, which has
been found to have substantial systematic geoloca-
tion errors that may consequently impact the accu-
rate measurement of forest canopy parameters (Tang
et al. 2023).
In general, researchers applied two methods for esti-
mating AGB using LiDAR: the individual tree-based
approach (ITA) (e.g. Fu et al. 2018) and the area-based
approach (ABA) (e.g. Su et al. 2016). ABA estimates forest
AGB across an entire area by extracting stand character-
istics from sample plots and applying statistical models.
Since plots are fundamental for tting, obtaining su-
cient sample plots presents a signicant challenge. The
ITA method oers several advantages: rstly, it increases
the sample size and provides detailed information at the
individual tree, thereby enhancing the understanding of
biomass uncertainty across dierent spatial scales (Xu
et al. 2018). Second, ITA demonstrates high precision by
capturing individual tree characteristics, allowing for
more rened estimates that are suitable for studies
focused on specic species or growth stages. Its adapt-
ability enables exible applications across various forest
types and structures, meeting the research needs of
diverse ecosystems.
There are, however, still a lot of errors at both the
individual tree and stand levels, regardless of which of
the ITA and ABA are used to estimate AGB. These include
errors in the estimated parameters of individual trees,
conversion factors, allometric equations, plot location
errors, and more. As a result, the variables used to create
AGB prediction models for the whole study area often
have errors, which violates the assumptions of the pre-
diction models.
Another type of model called Errors-in-Variables
(EIV) can be used with both linear and nonlinear
frameworks. It allows for variables with measurement
errors and simplies the estimation of parameters for
each model. Recent studies have attempted to
address the issue of errors in predictor variables for
AGB estimation models by integrating multi-source
data through EIV models (Dong, Zhang, and Li
2015). The EIV model eectively addresses data
aws, guaranteeing a globally optimal solution for
the system of equations through comprehensive ana-
lysis. Therefore, it may provide a reliable tool for AGB
estimation using remotely sensed data.
In summary, conicts arise in the application of
ABA and ITA within the current AGB estimation utiliz-
ing remote sensing techniques. While ABA is suitable
for large-scale AGB estimation, the uncertainty arising
from limited training samples may hinder the model’s
ability to accurately represent the spatial heteroge-
neity and variability of biomass in a specic region,
thereby impacting the precision of AGB estimates.
The ITA is suitable for AGB estimation across various
spatial scales with increasing sample sizes; however,
it is aected by various errors that propagate from
ITA-based AGB estimation procedures. To address the
conicts between ITA and ABA, this study proposes
a nondestructive, integrated framework for estimat-
ing forest AGB in northeastern China, utilizing an AGB
upscaling estimation model based on various LiDAR
datasets and Landsat 8 imagery. The framework
included: (1) estimating individual tree parameters
from the fused ULS and HLS (i.e. U-HLS) data to
achieve individual tree- and plot-level AGB estima-
tions; (2) developing an EIV model for standard AGB
estimation using ULS data; (3) performing sample
augmentation of reference AGB using the EIV model
and canopy height data extracted from ICESat-2/
ATL08; and (4) estimating the accurate ne-
resolution (30 m) forest AGB using random forest
and empirical Bayesian Kriging (RF-EBK) method and
exploring the recent AGB change (2019–2022) in NEC.
This nondestructive, integrated framework provides
technical support for accurate large-scale AGB esti-
mation and regional AGB change analysis, thus
improving the temporal interval of the large-scale
(national, regional) forest resource inventory (e.g.
from 5 to 2 or 3 years).
By incorporating the EIV model to account for mea-
surement errors and uncertainties in model parameters,
systematic biases between AGB and canopy height
derived from ULS can be quantied, thereby enhancing
the robustness of AGB estimation at the standard scale.
Furthermore, the integration of the canopy height
extracted from ICESat-2/ATL08 data addresses the bot-
tleneck of insucient samples for large-scale AGB
2Z. ZHEN ET AL.

estimation while ensuring sample representativeness
and diversity.
2. Materials
2.1. Study area
The study area is primarily located in NEC, encompassing
Heilongjiang Province, Jilin Province, Liaoning Province,
and the eastern part of the Inner Mongolia Autonomous
Region (including Hulunbeier City, Xing’an League,
Tongliao City, Chifeng City, and Xilingole League), the
study area spans from 111°14′E to 135°05′E in longitude
and from 38°72′N to 53°56′N in latitude. The NEC is rich
in tree species, with forest cover accounting for 37% of
the national total and a total forest area of approxi-
mately 680,000 km
2
. The forest stock volume is about
3.2 billion m
3
, constituting 30% of the national total
forest stock volume (Pan et al. 2011). Five study sites
were selected in typical natural forests within NEC (see
Figure 1), located in the Wulikuma Forest Farm in Genhe
City (GH), Inner Mongolia Autonomous Region; the
Heping Forest Farm in Muling City (ML) and the
Maoershan Experimental Forest Farm in Shangzhi City
(SZ), Heilongjiang Province; the Huangsongpu Forest
Farm in Yanji City (YJ), Jilin Province; and the Wandian
Forest Farm in Fushun City (FS), Liaoning Province. The
main forest types in these areas include evergreen nee-
dle-leaved forest (ENLF), deciduous needle-leaved forest
(DNLF), deciduous broad-leaved forest (DBLF), and
mixed forests (MF).
2.2. Data preparation
2.2.1. Reference data
The eld data used in this study were collected from 109
plots from 2021 to 2023, primarily focusing on the domi-
nant tree species within the study area. Among these
plots, 66 have an area of 0.09 hectares (30 m × 30 m)
and 43 have an area of 0.04 hectares (20 m × 20 m).
Data was gathered for a grand total of 8,208 trees,
encompassing measurements of tree height (H, m) and
diameter at breast height (DBH, cm). In addition, for
trees with a DBH exceeding 5 cm within the plots, the
species, DBH, height, location, and growing condition of
each tree were recorded. The reference AGB of indivi-
dual trees was typically calculated using species-specic
additive biomass equations (see Equation 1) (Dong,
Zhang, and Li 2015). The species-specic coecients
used in this study were shown in Supplementary Table
S1. The reference AGB was then obtained by summing
up the reference AGB of all individual trees within the
plot. The descriptive statistics for all collected trees are
presented in Table 1.
Figure 1. The distribution of (a) Northeast China. (b) The locations of the five study sites. GH-Genhe City; SZ-Shangzhi City; ML- Muling
City; YJ- Yanji City; FS- Fushun City.
Table 1. Individual tree parameters of all the 109 plots.
Factors Count Minimum Maximum Average Midpoint Standard deviation
H (m) 8208 4.00 33.30 15.73 16.20 4.94
DBH (cm) 8208 5.10 58.30 18.34 17.60 7.45
AGB (kg) 8208 4.25 1605.72 128.34 87.31 145.18
GISCIENCE & REMOTE SENSING 3

In Equation (1), W
a
, W
s
, W
b
, and W
f
represent the
above-ground, stem, branch, and leaf biomass (kg),
respectively. D denotes the DBH (cm), and H represents
the H (m). The term ln is the natural logarithm, ε is the
error term, and ai,bi,ci,a�
i,b�
i and c�
i are regression coe-
cients, where i = [s,b,f] corresponds to the regression
coecients for the stem, branch, and leaf, respectively.
2.2.2. LiDAR data and preprocessing
2.2.2.1. Near-ground LiDAR data. Two types of near
ground LiDAR data were collected in this study for indi-
vidual tree biomass estimation during the growing sea-
sons (June to September) of 2021–2023: ULS and HLS
data. The ULS data were acquired using a Feima D200
drone equipped with a D-LiDAR200 system, covering
a total of 7 km
2
within the study area. Concurrently, the
HLS data were collected using the SLAM100 system. The
parameters of LiDAR datasets are summarized in Table 2.
The preprocessing of the ULS and HLS data involved:
(1) noise removal; (2) ground point classication using
improved progressive triangular irregular network (TIN)
densication (Zhao et al. 2016); (3) point cloud normal-
ization; (4) fusion of ULS and HLS data; (5) canopy
height extraction. The 98
th
percentile canopy height
(RH98) was extracted from the U-HLS data for subse-
quent analysis.
2.2.2.2. Spaceborne LiDAR data. The satellite LiDAR
data used in this study was the ICESat-2 ATL08 (ver-
sion 5) data from June to September 2019–2022,
obtained from the National Snow and Ice Data
Center (NSIDC, https://nsidc.org/data/icesat-2/tools).
The ATL08 (forest canopy height product), one of
the ICESat-2 ATLAS products, provides terrain and
forest canopy height along the satellite’s trajectory
(Li et al. 2020). The forest RH98 canopy height data,
representing the relative canopy height at the 98
th
percentile (Neuenschwander et al. 2020), was
extracted from ATL08. The preprocessing of ICESat-
2/ATL08 data included: (1) excluding non-forest areas;
(2) selecting footprints with sucient ground and
canopy photon counts since insucient photons
could result in errors in canopy height and elevation
estimations (Wu and Shi 2022). The footprints with
a canopy-to-ground photon ratio greater than 5%
and a canopy photon count exceeding 50 (Wu and
Shi 2022) were selected for further regional scale AGB
estimation; (3) selecting footprints with stronger
beams collected during nighttime, because the night-
time data are less aected by solar background noise,
oering more accurate canopy height information
(Rodda et al. 2024).
A total of 16,430 forest canopy height measurements
(8,400 points from the rst period, 2019–2020, and 8,030
points from the second period, 2021–2022) were
obtained in the forests of NEC (Figure 2) and used for
biomass estimation at the ICESat-2/ATL08 footprint
level.
2.2.3. Landsat 8 OLI imagery and its feature
extraction
In this study, Landsat 8 OLI data (Data ID: LANDSAT/
LC08/C02/T1_L2) with a nominal spatial resolution of
30 m were obtained from Google Earth Engine (GEE,
https://code.earthengine.google.com/). After cloud
removal, nine spectral indices were derived from
Landsat 8 OLI imagery and ve textures for each
band (B2-B7) within a 3 × 3 pixel kernel (Conners,
Trivedi, and Harlow 1984) (see Supplementary
Table S3).
2.2.4. Auxiliary data
The Shuttle Radar Topography Mission (SRTM) data
(ID: “USGS/SRTMGL1_003”) with a 30 m resolution
were applied to extract topographical factors after
void-lling processing (Reuter, Nelson, and Jarvis
2007), including elevation, slope, aspect, and hill-
shade variables.
In addition, the global forest coverage data pro-
duct (Hansen et al. 2013) with a spatial resolution of
30 m was applied to extract continuous monitoring of
global forest cover changes from 2000 to 2023. Due
to the little change in forest coverage in NEC from
2019 to 2022, the average forest coverage data (in
percentage) was extracted as a feature for subse-
quent analyses.
3. Methods
The design of this study aimed to accomplish the four
previously mentioned objectives (see Figure 3).
Table 2. The essential parameters of the two types of near-
ground LiDAR data.
Parameters ULS HLS
Sensor RIEGL mini VUX-1UAV SLAM100
Point frequency (kpts/s) 100 320
Ranging accuracy (cm) ±1 ±2
Echo intensity (bit) 16 8
Field of view-horizontal (°) 360 270
Field of view-vertical (°) – 360
Measuring range (m) >250 120
Average point density (pts/m
2
) 220–300 3×10
4
– 6×10
4
4Z. ZHEN ET AL.

3.1. Individual tree- and plot-level AGB estimation
and uncertainty analysis
3.1.1. Individual tree segmentation and parameter
estimation
This study applied the point cloud segmentation (PCS)
algorithm (Li et al. 2012) to segment individual trees
based on U-HLS data. Then, tree height (H
ITA
) and
diameter at breast height (DBH
ITA
) were estimated
by identifying local maxima in canopy height (Zhen
et al. 2022) and at 1.3 m aboveground using
a nonlinear least squares circle-tting algorithm
(Gander, Golub, and Strebel 1994), respectively. To
evaluate the accuracy of the PCS algorithm, a 1:1
matching approach was employed for validation,
including the following steps: (1) identifying all
segmented trees within a 3 m buer around each
reference tree and (2) selecting the uniquely corre-
sponding segmented tree with the smallest dierence
in height and DBH compared to the reference tree.
The accuracy of individual tree segmentation was
assessed using recall (r), precision (p), and F-score.
Based on the estimated H
ITA
and DBH
ITA
for the seg-
mented trees and the species information from the
1:1 matched trees, this study estimated individual tree
aboveground biomass ðBITAÞand plot-level above-
ground biomass ðBst;ITAÞusing the additive biomass
model described in Section 2.2.1.
3.1.2. Uncertainty analysis of the individual tree
parameter estimation
The Monte Carlo simulation process for quantifying the
uncertainty in AGB estimation caused by individual tree
parameter estimation (i.e. tree height and DBH) is as
follows:
Step 1: Normal distributions were constructed to
simulate tree height and DBH, with the mean of 0 and
the standard deviation of the residual error of the esti-
mated parameter. The function of the normal distribu-
tions is as follows:
Where ε2
D and ε2
H represent the residuals of the esti-
mated DBH and tree height, respectively, while ε0
D and ε0
H
denote the randomly generated residuals for DBH and
tree height estimation.
Step 2: Based on the newly generated random resi-
duals from Step 1, new DBH and tree height values are
simulated as follows:
Figure 2. The distribution of ICESat-2/ATL08 footprints within the study area.
GISCIENCE & REMOTE SENSING 5

Where D and H represent the estimated DBH and tree
height, respectively, while D0and H0denote the newly
simulated DBH and tree height.
Step 3: Based on the simulated DBH and tree height,
the new individual tree AGB is calculated and
aggregated to obtain the new plot-level AGB according
to Equation (1). The process is executed 1000 times, and
the mean plot-level AGB in addition to the standard
deviation (uncertainty) is calculated from these
simulations.
Figure 3. Flowchart of the study.
6Z. ZHEN ET AL.

3.2. Establishment of EIV model for stand-level AGB
estimation
In this study, an Errors-in-Variables (EIV) model was
developed to estimate stand AGB using ULS data as
a medium. The EIV model (see Supplementary Equation
S1) can be applied to both linear and nonlinear models,
and allows for the input of estimated features that con-
tain errors and can simultaneously estimate the para-
meters of each model, yielding a globally optimal
solution for multiple sets of equations (Dong, Zhang,
and Li 2015).
Since the variables (i.e. canopy height and AGB)
estimated from remote sensing have errors, the EIV
model was established to estimate stand-level AGB as
follows:
In the equation, D and H represent the average DBH
and tree height of plots measured from eld surveys,
respectively, and are considered error-free-variables.
HULS is the canopy height extracted from ULS data,
treated as an error-in-variable. Bst;ITA is the stand-level
AGB estimated based on ITA, also considered an
error-in-variable. The ai;bi;ci (i = 1,2,3) are the model
parameters to be tted, and εi (i = 1,2) are the model
residuals. The model is solved using a two-stage error
variance model algorithm (Tang and Wang 2002). The
model II was applied to estimate stand-level AGB in
the subsequent analysis.
3.3. Sample augmentation of reference AGB using
the EIV model and ICESat-2/ATL08 data
The canopy height (H
ICE
) of ICESat-2/ATL08 is dened
as the 98
th
percentile canopy height (Neuenschwander
et al. 2019a). Due to the consistency between the
ICESat-2- and ALS-derived canopy heights (Rai et al.
2024), we replaced the canopy height derived from
ULS (H
ULS
) by that obtained from ICESat-2/ATL08 in
the EIV model (Model II) to estimate the stand-level
AGB (B
st,ICE
) for the ICESat-2/ATL08 footprints. The
stand-level reference AGB (B
st,ICE
) can be applied to
develop RF model for spatially continuous AGB estima-
tion (Huang et al. 2019).
This approach signicantly augmented the number
of AGB samples (i.e. 109 to 16,430) and enabled the
development of RF model for the large-scale AGB
estimation with abundant training samples (see
Figure 4).
3.4. Fine-resolution AGB estimation of NEC using
RF-EBK model
3.4.1. The AGB estimation based on RF model
A total of 50 feature variables were extracted, including
45 features (i.e. six original bands, 30 texture features, six
vegetation indices, and three orthogonal variables)
extracted from Landsat 8 OLI imagery, four terrain fea-
tures (i.e. elevation, slope, aspect, and hillshade) were
extracted from SRTM data, and forest cover data (per-
centage). The recursive feature elimination (RFE) was
used to rank and select important variables (Guyon
et al. 2002). Low-contributing features were gradually
removed until the model performance (i.e., R
2
) showed
no signicant improvement. Ultimately, the remaining
feature variables were applied to develop an RF model
for spatially continuous forest AGB estimation. The RF
model, a machine learning algorithm capable of hand-
ling large numbers of predictors that are neither inde-
pendent nor linearly separable, has been widely applied
in forest biomass estimation (Belgiu and Drăguţ 2016).
3.4.2. Bias correction based on EBK
EBK, as a geostatistical interpolation method, diers from
traditional kriging in that it can automatically simulate
variogram parameters, reducing reliance on interactive
modeling (Gribov and Krivoruchko 2020). The EBK opti-
mizes interpolation by simulating multiple variograms and
calculating their weights, making it particularly eective in
handling irregularly distributed data or cases with strong
Figure 4. Sample augmentation procedure for the large-scale
AGB estimation using the EIV model (model II) and canopy
height data extracted from ICESat-2/ATL08.
GISCIENCE & REMOTE SENSING 7

spatial autocorrelation. In this study, EBK was applied to
interpolate the residuals (footprint-level) of the RF model
for AGB estimation, aiming to obtain a continuous residual
distribution and correct systematic biases in RF estimates.
By interpolating residuals, this study compensates for sys-
tematic errors in RF estimates in specic regions, making
the nal AGB estimates more accurate. To meet the funda-
mental assumptions of kriging, a normality test (e.g.
Anderson−Darling test) should be performed on the inter-
polated data (Wang, Zhang, and Li 2012).
The EBK interpolation was performed using the
Geostatistical Analyst module in ArcGIS 10.4, with the
power variogram model (POWER) selected (Krivoruchko
2012). Considering the trade-o between computational
eciency and interpolation accuracy, a standard circular
neighborhood search method was employed. The number
of neighboring points was limited by setting the minimun
and maximum adjacent points to 10 and 15, respectively.
Additionally, a single-sector search mode was selected, and
the output surface type was set to PREDICTION to ensure
a continuous prediction surface.
4. Results
4.1. Uncertainty analysis of AGB estimation at the
individual tree and plot scale
4.1.1. Individual tree- and plot-level AGB estimation
Based on U-HLS data, the PCS method achieved high-
accurate individual tree segmentation, with average recall
(r), precision (p), and F-score values of 0.84, 0.76, and 0.79,
respectively. A total of 6,902 accurately segmented indivi-
dual trees (1:1 matched trees) will be used for AGB estima-
tion. As demonstrated in Table 3, the DBH estimation
accuracy is particularly high when using fused point cloud
data, mainly because HLS data eectively captures tree
trunks. Although the accuracy of tree height estimation is
slightly lower than that of DBH estimation, it still achieved
a relatively high accuracy (rRMSE = 14.8%; Bias = −0.16 m).
The estimation accuracy for both individual tree and
plot-level AGB is shown in Figure 5. A slight decrease in
accuracy was observed when scaling from individual trees
to the plot level. Since only 1:1 matched trees (total: 6,902)
were used for the AGB estimation, there is a tendency for
AGB to be underestimated, particularly in plots where AGB
exceeds 150 Mg/ha, where this underestimation is more
pronounced. The underestimation of the plot-level AGB is
mainly due to the omission error from the small trees.
4.1.2. Uncertainty in plot AGB estimation
According to the uncertainty analysis of plot AGB esti-
mation (Figure 6), the uncertainty in AGB estimation
caused by individual tree parameter estimation is stabi-
lized at 10.7 Mg/ha (Figure 6(a)); after 1,000 simulations,
the plot-scale AGB estimates follow a normal distribu-
tion, with a mean of 109.16 Mg/ha and a standard devia-
tion of 10.7 Mg/ha.
4.2. Establishment of EIV model and augmentation
of reference AGB
An EIV model was developed to estimate the stand-level
AGB using ULS data as a medium. As shown in Table 4,
Table 3. Accuracy assessment of single tree parameter
extraction.
Parameters R
2
RMSE rRMSE(%) Bias
H(m) 0.67 2.45 14.8 −0.16
DBH(cm) 0.93 1.84 9.6 −0.26
Figure 5. Accuracy assessment of (a) individual tree AGB (b) plot AGB.
8Z. ZHEN ET AL.

the EIV model achieved satisfactory tness indices, with
the canopy height exhibiting better tness than the
AGB. The model II (in Equation (6)) was applied to esti-
mate the stand-level AGB (B
st,ICE
) using the ICESat-2/
ATL08 footprints for developing RF model. The number
of AGB samples was augmented from 109 to 16,430
samples, including 8400 samples for the rst period (-
2019–2020) and 8030 samples for the second period
(2021–2022).
4.3. Fine-resolution AGB estimation of NEC using
RF-EBK model
4.3.1. Feature selection and RF model development
Using the RFE method, the most important features for
model prediction were identied, with 28 and 31 fea-
tures (see Table 5) retained for the rst and second
periods, respectively (see Figure 7). These features
were then used to develop RF models of the AGB
estimation for the two periods. The studentized resi-
dual plots in Figures 8(b,d) indicated that the residuals
for both periods predominantly fall within the range of
[−2, + 2], with slightly clustered patterns. For lower
predicted values (<80 Mg/ha), negative residuals
exceeded positive residuals; conversely, for higher pre-
dicted values (>120 Mg/ha), positive residuals pre-
vailed. In the intermediate ranges of predicted
values, most residual points were evenly distributed
near the zero-residual line. This suggests the pre-
sence of overestimation for low values and under-
estimation for high values, which is consistent with
the results shown in Figures 8(a,c).
4.3.2. Bias correction based on EBK
The Anderson–Darling test was utilized to assess the
normality of the residuals to meet the assumption of
Kriging. The test statistics for the two research periods
were 0.77 and 0.68, indicating that Kriging’s assumption
was satised, and the residuals were subsequently spa-
tially interpolated using EBK.
As shown in Figure 9, the spatial distribution trends of
residuals from the RF model were similar across dierent
periods. Positive residuals (underestimation) were
mainly concentrated in high-altitude areas (i.e. the
Greater Khingan Mountains, Lesser Khingan Mountains,
and Changbai Mountains) with dense forests and large
AGB values, likely attributable to the saturation of
Landsat 8 OLI imagery in dense forests. The degree of
underestimation gradually decreased with decreasing
elevation. Positive residuals demonstrate a wider spatial
impact, while negative residuals (overestimation) display
a more limited distribution, predominantly located in
the forest–grassland transition zones in the southwest
Figure 6. Uncertainty analysis of the plot AGB estimation. (a) uncertainty caused by individual tree parameter estimation; (b) violin
plot of simulated plot-scale AGB.
Table 4. The parameter estimates and model fitting of the EIV
model.
Model a
i
b
i
c
i
R
2
RMSE rRMSE(%)
Model I 0.89 1.04 0.11 0.81 1.23m 6.4
Model II 0.15 2.27 – 0.69 19.59Mg/ha 15.1
GISCIENCE & REMOTE SENSING 9

and northwest, as well as in low-altitude coastal regions,
characterized by relatively low forest cover and dimin-
ished AGB values.
Finally, the AGB values were corrected using the resi-
duals interpolated by EBK. As shown in Figure 10(a,c),
the prediction accuracy of the RF-EBK model for both
phases improved signicantly, compared to that of the
RF model (see Figure 8). The R
2
values grew by 49% and
37%, while the RMSE reduced by 25% for both times and
the rRMSE also decreased by 25% for both periods,
respectively. The studentized residuals of the RF-EBK
model, as illustrated in Figures 10(b,d) and 8(b,d),
demonstrated a more random distribution among the
predicted values compared to the RF model, indicating
a reduction in the clustering pattern.
4.3.3. Fine-resolution mapping forest AGB in
northeast China using RF-EBK
The continuous forest AGB maps of NEC with a 30-m
resolution were generated for the two periods using the
RF-EBK model, and the recent AGB change was illustrated
(see Figure 11). In Figure 11(c), the forest AGB in NEC
exhibited an increasing trend from 2019 to 2022.
Figure 11(a,b) indicated that higher AGB values were
mainly concentrated in the NEC’s three major mountain
ranges, with the Changbai Mountains having the highest
forest AGB, followed by the Lesser Khingan Mountains and
Greater Khingan Mountains. The AGB in the Lesser
Khingan Mountains and Changbai Mountains exhibited
more pronounced growth, whereas that in the Greater
Khingan Mountains experienced a considerable reduction.
Table 5. The features selected in the RF model development for the two periods (Period 1:28 features and
Period 2:31 features).
Features
Period 1
2019–2020
Period 2
2021–2022 Features
Period 1
2019–2020
Period 2
2021–2022
B2 √ √ B6_contrast √ –
B2_contrast – – B6_dvar √ √
B2_dvar – – B6_inertia √ √
B2_inertia – – B6_savg √ √
B2_savg √ √ B6_var √ √
B2_var – – B7 √ √
B3 √ √ B7_contrast – –
B3_contrast – – B7_dvar – –
B3_dvar – – B7_inertia – –
B3_inertia – – B7_savg √ √
B3_savg √ √ B7_var – –
B3_var – √ DVI – –
B4 √ √ EVI √ √
B4_contrast – – LAI √ √
B4_dvar – – NDVI – √
B4_inertia – – RVI √ √
B4_savg √ √ SAVI – –
B4_var – – TCB √ √
B5 – √ TCG √ √
B5_contrast √ √ TCW – –
B5_dvar √ √ Tree cover √ √
B5_inertia √ √ DEM √ √
B5_savg √ √ slope √ √
B5_var √ √ aspect – –
B6 √ √ hillshade √ √
Bi (i=2,3,4,5,6,7): The band information of Landsat 8 OLI; contrast: Measures the local contrast of an image; dvar: Difference
variance; Inertia: Inertia; savg: Measures how spread out the distribution of gray-levels; var: variance; NDVI: Normalized
Difference Vegetation Index; DVI: Difference Vegetation Index; RVI: Ratio Vegetation Index; SAVI: Soil-Adjusted Vegetation
Index; EVI: Enhanced Vegetation Index; LAI: Leaf Area Index; TCB: Brightness; TCW: Wetness; TCG: Greenness; DEM: Elevation
above sea level in meters; Tree cover: the percentage of pixel area covered by trees.
Figure 7. Variation in the performance of the RF model with
changes in the number of features.
10 Z. ZHEN ET AL.

Figure 8. Accuracy assessment: (a) and (c) represent scatter plots of the prediction accuracy of RF models for periods 1 and 2,
respectively; (b) and (d) represent the corresponding residual plots for the two periods (Period 1: 2019–2020; Period 2: 2021–2022).
Figure 9. The AGB residual maps of RF model interpolated by EBK for (a) period 1: 2019–2020;(b) period 2: 2021–2022.
GISCIENCE & REMOTE SENSING 11

5. Discussion
5.1. Individual tree parameters estimation and its
uncertainty
This study presents a framework for scaling up AGB
estimation from individual trees to stand level, and
then to the regional level, utilizing multi-source remote
sensing data and an error-in-variable model.
The accuracy of individual tree segmentation is essen-
tial for extracting tree parameters and estimating above-
ground biomass. This study demonstrated that the
combination of ULS and HLS data using a PCS algorithm
yielded superior accuracy in individual tree segmenta-
tion (r = 0.84, p = 0.76, F-score = 0.79) compared to the
results obtained using ULS and backpack laser scanning
(BLS) (r = 0.64, p = 0.64, F-score = 0.64) (Zhen et al. 2022),
and also surpassed the accuracy achieved with ULS and
TLS data (r = 0.76, p = 0.7, F-score = 0.73) (Zhao et al.
2023). Furthermore, the diameter at breast height
(DBH) accuracy of individual trees is analogous to that
derived from ULS-BLS and ULS-TLS data, with R
2
and
RMSE values of 0.9 and 2 cm, respectively. Nonetheless,
the accuracy of tree height (R
2
= 0.67, RMSE = 2.45 m)
obtained with U-HLS in this study is comparable to that
(R
2
= 0.78, RMSE = 2.78 m) achieved with ULS-BLS (Zhen
et al. 2022), yet signicantly inferior to that (R
2
= 0.72,
RMSE = 0.65 m) derived from ULS-TLS (Zhao et al. 2023).
This is because TLS, by more capturing extensive point
cloud data, can more eciently and eectively detect
sub-canopy structures across dierent forest types than
Figure 10. Accuracy assessment: (a) and (c) represent scatter plots of the prediction accuracy of RF-EBK models for period 1 and 2,
respectively; (b) and (d) represent the corresponding residual plots for the two periods (Period 1: 2019–2020; Period 2: 2021–2022).
12 Z. ZHEN ET AL.

BLS and HLS. Despite the signicantly lower point cloud
density of the HLS compared to the TLS (3 × 10
4
− 6 ×
10
4
versus 1.9 × 10
5
− 2.7 × 10
5
points/m
2
), HLS still
achieves comparable accuracy in measuring DBH. In
the study by Bauwens et al. (2016), a comparison
between HLS and TLS demonstrated that HLS yielded
the best DBH estimation results, with a bias of 0.08 cm
and an RMSE of 1.11 cm. Furthermore, HLS requires only
a single operator while maintaining data accuracy, thus
enhancing data collection eciency and portability, par-
ticularly in remote or complex environments, such as
natural forests or burned areas. Therefore, estimating
individual tree aboveground biomass using a univariate
allometric equation based on HLS-derived DBH appears
to be a feasible approach for future studies.
Unlike previous studies that focused on assessing
model uncertainty (Chen, Laurin, and Valentini 2015),
this study explored the uncertainty in AGB estimation
based on remote sensing processes within the ITA fra-
mework at the individual tree-to-plot scale. The results
indicate that the uncertainty in individual tree parameter
estimation is relatively low (10.7 Mg/ha), which can be
attributed to the integrated application of multiple near-
ground LiDAR datasets. Additionally, accurate tree
height estimation from ULS and comprehensive trunk
scanning from HLS provide a solid foundation for DBH
estimation (Zhao et al. 2023), further reducing the uncer-
tainty in individual tree parameter estimation.
5.2. Consistency of the canopy height extracted
from ICESat-2 and ULS data
Based on the preliminary ltering results of ICESat-2/
ATL08 data described in Section 2.2.2.2., this study
further selected ATL08 data that spatially overlapped
with the ULS data for consistency evaluation. To assess
the accuracy of ICESat-2/ATLAS data, the canopy height
model (CHM) and digital terrain model (DTM) were
derived from ULS data. As shown in Figure 12(a), the
RH98 forest canopy height extracted from ICESat-2/
ATLAS under the nighttime strong-beam mode exhib-
ited a high level of agreement with the ULS-derived
CHM, with a coecient of determination (R
2
) of 0.72,
RMSE of 2.27 m, and rRMSE of 19.83%. Furthermore, as
illustrated in Figure 12(b), an even stronger agreement
was observed between ICESat-2/ATLAS and ULS-derived
DTM, with an R
2
of 0.99, an RMSE of 1.50 m, and an
exceptionally low rRMSE of only 0.1%.
5.3. Sample augmentation
Comparing to individual tree-based approach, one of
the major disadvantages of area-based approaches for
AGB estimation is the limited number of reference AGB
plots. However, the area-based approach is convenient
and indispensable for large scale (i.e. region, country,
and continent) AGB estimation. This study presents
Figure 11. Estimation of forest AGB in NEC based on RF-EBK: (a) period 1: 2019–2020; (b) period 2: 2021–2022; and (c) the AGB change
during the two periods.
GISCIENCE & REMOTE SENSING 13

a hybrid method for regional AGB estimation and con-
ducts sample augmentation based on the EIV model II
and ICESat-2/ATL08 data, using ULS data as a medium.
There are two assumptions of this hybrid method: (1) the
98
th
percentile canopy height extracted from ICESat-2/
ATL08 data has high consistency with that derived from
ULS (Rai et al. 2024; Neuenschwander et al. 2019b;
Neuenschwander et al. 2020); (2) the footprint-level
AGB estimated by the EIV Model II can be applied as
reference AGB for establishing RF model.
The reference AGB at the footprint scale serves as
a compromise due to the lack of sucient reference
data for large-scale AGB estimation and has been applied
in previous studies (Hall et al. 2011). The reference AGB,
although not the actual value, is derived from an indivi-
dual-tree-based methodology and the EIV model in this
study. The total number of reference AGB samples has
increased by 150 times, with sample plots dispersed
throughout Northeast China (see Figure 2). The utilization
of numerous samples facilitates an extensive evaluation
of forest biomass across diverse environmental variables,
encompassing climate, soil, topography, and vegetation
species. This method successfully reduces biases resulting
from inadequate sample size, therefore enhancing the
accuracy of AGB estimation.
5.4. The AGB mapping using EBK-RF model and its
short-term dynamics
Although random forest model is free of any assump-
tions (i.e. independent reference AGB), the residuals of
RF model with a Moran’s index of 0.31 indicate obvious
positive spatial autocorrelation, which would impact the
accuracy of continuous AGB mapping. In this study, the
biases were corrected using EBK, and the overall predic-
tion accuracy was further improved (see Figure 10(a,c)).
At the same time, the Moran’s index of the corrected
residuals becomes 0.11, indicating that the spatial auto-
correlation is weakened. The error metrics for AGB esti-
mation of NEC using EBK-RF are below the international
standards for AGB estimation, with satellite remote sen-
sing data (RMSE <20 Mg/ha or relative RMSE < 20%) (Hall
et al. 2011; Zolkos, Goetz, and Dubayah 2013).
Nonetheless, it is challenging to detect changes in
natural forests over a span of a few years. The change
of AGB may arise not alone from forest growth or
decline, but also from the model or data, notwithstand-
ing our endeavors to uphold optimal model perfor-
mance and data quality. Future studies could attempt
to integrate Global Ecosystem Dynamics Investigation
(GEDI) data with ICESat-2 data for cross-validation to
reduce the eects of uncertainties in spaceborne LiDAR
data. This approach can also compensate for the limited
penetration capability of ICESat-2’s photon-counting
LiDAR in densely vegetated regions. Furthermore, data
integration may augment the sample size, thereby
diminishing uncertainty in the modeling process result-
ing from inadequate training samples. The framework
eectively reduces the time required for large-scale (i.e.
national or regional) AGB assessment and eciently
identies forest AGB dynamics.
5.5. Limitations and future work
Although this study provides a highly accurate hybrid
framework incorporating ITA and ABA for regional-scale
Figure 12. Consistency evaluation of ICESat-2/ATLAS canopy height and ground elevation (a) CHM; (b) DTM. The blue dashed line
represents the 1:1 line, while the red solid line represents the fitted line between the measured and estimated values.
14 Z. ZHEN ET AL.

estimation of forest AGB, it still has certain limitations.
Firstly, meteorological data were not included as feature
variables in the model since this study focused on the
short-term AGB changes (2019–2022). Given the com-
plex mechanisms driving AGB variations, climatic factors,
such as temperature and precipitation play a crucial role
in inuencing photosynthetic eciency, vegetation
growth cycles, and carbon sequestration capacity (Liu
et al. 2024b), thereby profoundly aecting the long-term
dynamics of forest AGB. Thus, the meteorological data
should be considered in the long-term AGB estimation.
Second, the observed short-term uctuations in forest
AGB may not accurately reect actual changes due to
forest growth or decline, but rather be aected by train-
ing data, model performance, remote sensing data qual-
ity, and even forest management practices, and so on.
However the implementation of the National Plan for
Forest Protection and Development (2016–2050) by the
National Forestry and Grassland Administration of China,
it has eectively maintained a stable forest environment
in Northeast China. Thus, this study only focused on
enhancing the training data, improving the estimation
model, and validating the data quality. In the future,
GEDI data can be combined with ICESat-2 data for cross-
validation and sample size augmentation, hence redu-
cing uncertainty in the modeling process and improving
the accuracy of AGB estimation.
6. Conclusion
This study introduces a highly accurate hybrid frame-
work incorporating ITA and ABA for regional-scale esti-
mation of forest AGB through the combination of
handheld, UAV, and satellite LiDAR, together with
Landsat 8 imagery. The results showed that by utilizing
various near-ground LiDAR data (i.e. U-HLS), both indivi-
dual tree and plot-level AGB estimates achieved satisfac-
tory accuracy in the forests of NEC, with an RMSE of
29.88 kg, an rRMSE of 22.98%, and bias of −3.9 kg at
the individual tree level, and an RMSE of 20.60 Mg/ha,
an rRMSE of 15.58%, and bias of −16,56 Mg/ha at the
plot level. At the stand level, this study constructed an
error-in-variable (EIV) model based on plot-level AGB
estimates derived from the ITA method (R
2
= 0.69,
RMSE = 19.59 Mg/ha, rRMSE = 15.1%) and used this
model to estimate footprint-level AGB based on canopy
height (RH98) data extracted from ICESat-2/ATL08. This
approach signicantly expanded the sample size for
regional AGB estimation, increasing it by 150 times.
Based on the extensional reference AGB, the RF-EBK
model was established for AGB estimation and demon-
strated a substantial improvement in accuracy compared
to the RF model, with an increase in R
2
of approximately
37%–49% and a reduction in RMSE and rRMSE by 25%.
The framework proposed in this study provides technical
support for scaling AGB estimation from individual trees
to the regional level, making it suitable for large-scale
forest AGB assessments based on both ITA and ABA, thus
improving the temporal interval of the national or regio-
nal forest resource inventory.
Disclosure statement
No potential conict of interest was reported by the author(s).
Funding
This research was funded by the Key Project of the China
National Key Research and Development Program,
[2021YFD2200401]; the National Natural Science Foundation
of China “Multi-scale forest aboveground biomass estimation
and its spatial uncertainty analysis based on individual tree
detection techniques” [32071677], National Forestry and
Grassland Data Center-Heilongjiang platform
[2005DKA32200-OH].
Data availability statement
The data that support the ndings of this study are openly
available in [gshare] at [10.6084/m9.28007714]. The reference
data are available on request from the corresponding author
(Y.Z.).
References
Bauwens, S., H. Bartholomeus, K. Calders, and P. Lejeune. 2016.
“Forest Inventory with Terrestrial Lidar: A Comparison of
Static and Hand-Held Mobile Laser Scanning.” Forests 7 (6):
127. https://doi.org/10.3390/f7060127 .
Belgiu, M., and L. Drăguţ. 2016. “Random Forest in Remote
Sensing: A Review of Applications and Future Directions.”
ISPRS Journal of Photogrammetry and Remote Sensing
114:24–31. https://doi.org/10.1016/j.isprsjprs.2016.01.011 .
Cao, L., N. C. Coops, J. L. Innes, S. R. J. Sheppard, L. Fu, H. Ruan,
and G. She. 2016. “Estimation of Forest Biomass Dynamics in
Subtropical Forests Using Multi-Temporal Airborne Lidar
Data.” Remote Sensing of Environment 178:158–171. https://
doi.org/10.1016/j.rse.2016.03.012 .
Chen, L., Y. Wang, C. Ren, B. Zhang, and Z. Wang. 2019.
“Assessment of Multi-Wavelength Sar and Multispectral
Instrument Data for Forest Aboveground Biomass Mapping
Using Random Forest Kriging.” Forest Ecology and
Management 447:12–25. https://doi.org/10.1016/j.foreco.
2019.05.057 .
Chen, Q., G. V. Laurin, and R. Valentini. 2015. “Uncertainty of
Remotely Sensed Aboveground Biomass Over an African
Tropical Forest: Propagating Errors from Trees to Plots to
Pixels.” Remote Sensing of Environment 160:134–143. https://
doi.org/10.1016/j.rse.2015.01.009 .
Conners, R. W., M. M. Trivedi, and C. A. Harlow. 1984.
“Segmentation of a High-Resolution Urban Scene Using
GISCIENCE & REMOTE SENSING 15

Texture Operators.” Computer Vision, Graphics and Image
Processing 25 (3): 273–310. https://doi.org/10.1016/0734-
189X(84)90197-X .
Dong, L., L. Zhang, and F. Li. 2015. “Developing Additive
Systems of Biomass Equations for Nine Hardwood Species
in Northeast China.” Trees 29 (4): 1149–1163. https://doi.org/
10.1007/s00468-015-1196-1 .
Forkuor, G., J. B. Zoungrana, K. Dimobe, B. Ouattara,
K. P. Vadrevu, and J. E. Tondoh. 2020. “Above-Ground
Biomass Mapping in West African Dryland Forest Using
Sentinel-1 and 2 Datasets-A Case Study.” Remote Sensing of
Environment 236:111496. https://doi.org/10.1016/j.rse.2019.
111496 .
Fu, L., Q. Liu, H. Sun, Q. Wang, Z. Li, E. Chen, Y. Pang, X. Song,
and G. Wang. 2018. “Development of a System of
Compatible Individual Tree Diameter and Aboveground
Biomass Prediction Models Using Error-In-Variable
Regression and Airborne Lidar Data.” Remote Sensing
10 (2): 325. https://doi.org/10.3390/rs10020325 .
Gander, W., G. H. Golub, and R. Strebel. 1994. “Least-Squares
Fitting of Circles and Ellipses.” BIT Numerical Mathematics
34 (4): 558–578. https://doi.org/10.1007/BF01934268 .
Gribov, A., and K. Krivoruchko. 2020. “Empirical Bayesian
Kriging Implementation and Usage.” Science of the Total
Environment 722:137290. https://doi.org/10.1016/j.scito
tenv.2020.137290 .
Guyon, I., J. Weston, S. Barnhill, and V. Vapnik. 2002. “Gene
Selection for Cancer Classication Using Support Vector
Machines.” Machine Learning 46 (1/3): 389–422. https://doi.
org/10.1023/A:1012487302797 .
Hall, F. G., K. Bergen, J. B. Blair, R. Dubayah, R. Houghton,
G. Hurtt, J. Kellndorfer, M. Lefsky, J. Ranson, and S. Saatchi.
2011. “Characterizing 3d Vegetation Structure from Space:
Mission Requirements.” Remote Sensing of Environment
115 (11): 2753–2775. https://doi.org/10.1016/j.rse.2011.01.
024 .
Hansen, M. C., P. V. Potapov, R. Moore, M. Hancher,
S. A. Turubanova, A. Tyukavina, D. Thau, S. V. Stehman,
S. J. Goetz, and T. R. Loveland. 2013. “High-Resolution
Global Maps of 21st-Century Forest Cover Change.” Science
342 (6160): 850–853. https://doi.org/10.1126/science.
1244693 .
Huang, H., C. Liu, X. Wang, X. Zhou, and P. Gong. 2019.
“Integration of Multi-Resource Remotely Sensed Data and
Allometric Models for Forest Aboveground Biomass
Estimation in China.” Remote Sensing of Environment
221:225–234. https://doi.org/10.1016/j.rse.2018.11.017 .
Krivoruchko, K. 2012. “Empirical Bayesian Kriging.” Arcuser Fall
6 (10): 1145.
Li, W., Q. Guo, M. K. Jakubowski, and M. Kelly. 2012. “A New
Method for Segmenting Individual Trees from the Lidar Point
Cloud.” Photogrammetric Engineering & Remote Sensing 78 (1):
75–84. https://doi.org/10.14358/PERS.78.1.75 .
Li, W., Z. Niu, R. Shang, Y. Qin, L. Wang, and H. Chen. 2020.
“High-Resolution Mapping of Forest Canopy Height Using
Machine Learning by Coupling Icesat-2 Lidar with
Sentinel-1, Sentinel-2 and Landsat-8 Data.” International
Journal of Applied Earth Observation and Geoinformation
92:102163. https://doi.org/10.1016/j.jag.2020.102163 .
Liu, Y., Y. Fan, J. Yue, X. Jin, Y. Ma, R. Chen, M. Bian, G. Yang, and
H. Feng. 2024b. “A Model Suitable for Estimating
Above-Ground Biomass of Potatoes at Dierent Regional
Levels.” Computers and Electronics in Agriculture
222:109081. https://doi.org/10.1016/j.compag.2024.109081 .
Liu, Y., H. Feng, J. Yue, X. Jin, Y. Fan, R. Chen, M. Bian, Y. Ma,
X. Song, and G. Yang. 2023. “Improved Potato AGB Estimates
Based on UAV RGB and Hyperspectral Images.” Computers
and Electronics in Agriculture 214:108260. https://doi.org/10.
1016/j.compag.2023.108260 .
Liu, Y., F. Yang, J. Yue, W. Zhu, Y. Fan, J. Fan, Y. Ma, et al. 2024a.
“Crop Canopy Volume Weighted by Color Parameters from
UAV-Based RGB Imagery to Estimate Above-Ground Biomass
of Potatoes.” Computers and Electronics in Agriculture
227:109678. https://doi.org/10.1016/j.compag.2024.109678 .
Lu, D., Q. Chen, G. Wang, L. Liu, G. Li, and E. Moran. 2016.
“A Survey of Remote Sensing-Based Aboveground Biomass
Estimation Methods in Forest Ecosystems.” International
Journal of Digital Earth 9 (1): 63–105. https://doi.org/10.
1080/17538947.2014.990526 .
Magruder, L., T. Neumann, and N. Kurtz. 2021. “Icesat-2 Early
Mission Synopsis and Observatory Performance.” Earth and
Space Science 8 (5): e2020EA001555. https://doi.org/10.
1029/2020EA001555 .
Markus, T., T. Neumann, A. Martino, W. Abdalati, K. Brunt,
B. Csatho, S. Farrell, H. Fricker, A. Gardner, and D. Harding.
2017. “The Ice, Cloud, and Land Elevation Satellite-2
(Icesat-2): Science Requirements, Concept, and
Implementation.” Remote Sensing of Environment
190:260–273. https://doi.org/10.1016/j.rse.2016.12.029 .
Neuenschwander, A., E. Guenther, J. C. White, L. Duncanson,
and P. Montesano. 2020. “Validation of Icesat-2 Terrain and
Canopy Heights in Boreal Forests.” Remote Sensing of
Environment 251:112110. https://doi.org/10.1016/j.rse.2020.
112110 .
Neuenschwander, A. L., and L. A. Magruder. 2019b. “Canopy
and Terrain Height Retrievals with Icesat-2: A First Look.”
Remote Sensing 11 (14): 1721. https://doi.org/10.3390/
rs11141721 .
Neuenschwander, A., and K. Pitts. 2019a. “The ATL08 Land and
Vegetation Product for the ICESat-2 Mission.” Remote
Sensing of Environment 221:247–259. https://doi.org/10.
1016/j.rse.2018.11.005 .
Pan, Y., R. A. Birdsey, J. Fang, R. Houghton, P. E. Kauppi,
W. A. Kurz, O. L. Phillips, A. Shvidenko, S. L. Lewis, and
J. G. Canadell. 2011. “A Large and Persistent Carbon Sink in
the World’s Forests.” Science 333 (6045): 988–993. https://
doi.org/10.1126/science.1201609 .
Rai, N., Q. Ma, K. P. Poudel, A. Himes, and Q. Meng. 2024.
“Evaluating the Uncertainties in Forest Canopy Height
Measurements Using Icesat-2 Data.” Journal of Remote
Sensing 4:0160. https://doi.org/10.34133/remotesensing.
0160 .
Reuter, H. I., A. Nelson, and A. Jarvis. 2007. “An Evaluation of
Void-Filling Interpolation Methods for SRTM Data.”
International Journal of Geographical Information Science
21 (9): 983–1008. https://doi.org/10.1080/136588106011
69899 .
Rodda, S. R., R. R. Nidamanuri, R. Fararoda, T. Mayamanikandan,
and G. Rajashekar. 2024. “Evaluation of Height Metrics and
Above-Ground Biomass Density from Gedi and Icesat-2 Over
Indian Tropical Dry Forests Using Airborne Lidar Data.”
Journal of the Indian Society of Remote Sensing 52 (4):
841–856. https://doi.org/10.1007/s12524-023-01693-1 .
16 Z. ZHEN ET AL.

Satoo, T., and H. A. Madgwick. 2012. Forest Biomass. Dordrecht:
Springer Science & Business Media.
Simon, F., A. Girard, M. Krotki, and J. Ordoñez. 2021. “Modelling
and Simulation of the Wood Biomass Supply from the
Sustainable Management of Natural Forests.” Journal of
Cleaner Production 282:124487. https://doi.org/10.1016/j.jcle
pro.2020.124487 .
Su, Y., Q. Guo, B. Xue, T. Hu, O. Alvarez, S. Tao, and J. Fang. 2016.
“Spatial Distribution of Forest Aboveground Biomass in
China: Estimation Through Combination of Spaceborne
Lidar, Optical Imagery, and Forest Inventory Data.” Remote
Sensing of Environment 173:187–199. https://doi.org/10.
1016/j.rse.2015.12.002 .
Tang, H., J. Stoker, S. Luthcke, J. Armston, K. Lee, B. Blair, and
M. Hofton. 2023. “Evaluating and Mitigating the Impact of
Systematic Geolocation Error on Canopy Height Measurement
Performance of GEDI.” Remote Sensing of Environment
291:113571. https://doi.org/10.1016/j.rse.2023.113571 .
Tang, S., and Y. Wang. 2002. “A Parameter Estimation Program
for the Error-In-Variable Model.” Ecological Modelling
156 (2–3): 225–236. https://doi.org/10.1016/S0304-3800(02)
00173-4 .
Vicharnakorn, P., R. P. Shrestha, M. Nagai, A. P. Salam, and
S. Kiratiprayoon. 2014. “Carbon Stock Assessment Using
Remote Sensing and Forest Inventory Data in Savannakhet,
Lao Pdr.” Remote Sensing 6 (6): 5452–5479. https://doi.org/
10.3390/rs6065452 .
Wang, K., C. Zhang, and W. Li. 2012. “Comparison of
Geographically Weighted Regression and Regression
Kriging for Estimating the Spatial Distribution of Soil
Organic Matter.” GIScience & Remote Sensing 49 (6):
915–932. https://doi.org/10.2747/1548-1603.49.6.915 .
Wu, Z., and F. Shi. 2022. “Mapping Forest Canopy Height at
Large Scales Using Icesat-2 and Landsat: An Ecological
Zoning Random Forest Approach.” IEEE Transactions on
Geoscience and Remote Sensing 61:1–16. https://doi.org/10.
1109/TGRS.2022.3231926 .
Xu, Q., A. Man, M. Fredrickson, Z. Hou, J. Pitkänen, B. Wing,
C. Ramirez, B. Li, and J. A. Greenberg. 2018. “Quantication
of Uncertainty in Aboveground Biomass Estimates Derived
from Small-Footprint Airborne Lidar.” Remote Sensing of
Environment 216:514–528. https://doi.org/10.1016/j.rse.
2018.07.022 .
Zhang, R., X. Zhou, Z. Ouyang, V. Avitabile, J. Qi, J. Chen, and
V. Giannico. 2019. “Estimating Aboveground Biomass in
Subtropical Forests of China by Integrating Multisource
Remote Sensing and Ground Data.” Remote Sensing of
Environment 232:111341. https://doi.org/10.1016/j.rse.2019.
111341 .
Zhao, M., J. Yang, N. Zhao, L. Liu, D. Ling, X. Xiao, T. Yue, and
J. P. Wilson. 2021. “Spatially Explicit Changes in Forest
Biomass Carbon of China Over the Past 4 Decades:
Coupling Long-Term Inventory and Remote Sensing Data.”
Journal of Cleaner Production 316:128274. https://doi.org/10.
1016/j.jclepro.2021.128274 .
Zhao, X., Q. Guo, Y. Su, and B. Xue. 2016. “Improved Progressive
Tin Densication Filtering Algorithm for Airborne Lidar Data
in Forested Areas.” ISPRS Journal of Photogrammetry and
Remote Sensing 117:79–91. https://doi.org/10.1016/j.
isprsjprs.2016.03.016 .
Zhao, Y., J. Im, Z. Zhen, and Y. Zhao. 2023. “Towards Accurate
Individual Tree Parameters Estimation in Dense Forest:
Optimized Coarse-To-Fine Algorithms for Registering UAV
and Terrestrial Lidar Data.” GIScience and Remote Sensing
60 (1): 2197281. https://doi.org/10.1080/15481603.2023.
2197281 .
Zhen, Z., L. Yang, Y. Ma, Q. Wei, H. I. Jin, and Y. Zhao. 2022.
“Upscaling Aboveground Biomass of Larch (Larix Olgensis
Henry) Plantations from Field to Satellite Measurements:
A Comparison of Individual Tree-Based and Area-Based
Approaches.” GIScience & Remote Sensing 59 (1): 722–743.
https://doi.org/10.1080/15481603.2022.2055381 .
Zolkos, S. G., S. J. Goetz, and R. Dubayah. 2013. “A
Meta-Analysis of Terrestrial Aboveground Biomass
Estimation Using Lidar Remote Sensing.” Remote Sensing of
Environment 128:289–298. https://doi.org/10.1016/j.rse.
2012.10.017.
GISCIENCE & REMOTE SENSING 17