Article

Canopy Height Integration for Precise Forest Aboveground

Biomass Estimation in Natural Secondary Forests of Northeast
China Using Gaofen-7 Stereo Satellite Data
Caixia Liu 1 , Huabing Huang 2,3 , Zhiyu Zhang 1 , Wenyi Fan 4,5 and Di Wu 6,7, *

1 International Research Center of Big Data for Sustainable Development Goals, State Key Laboratory of
Remote Sensing Science, Aerospace Information Research Institute, Chinese Academy of Sciences,
Beijing 100101, China; [email protected] (C.L.); [email protected] (Z.Z.)
2 School of Geospatial Engineering and Science, Sun Yat-Sen University, Guangzhou 510275, China;
[email protected]
3 Key Laboratory of Comprehensive Observation of Polar Environment (Sun Yat-Sen University), Ministry of
Education, Zhuhai 519082, China
4 School of Forestry, Northeast Forestry University, Harbin 150040, China; [email protected]
5 Key Laboratory of Sustainable Forest Ecosystem Management-Ministry of Education, Northeast Forestry
University, Harbin 150040, China
6 School of Surveying and Mapping Engineering, Heilongjiang Institute of Technology, Harbin 150050, China
7 Heilongjiang Geomatics Centre of Ministry of Natural Resources, Harbin 150081, China
* Correspondence: [email protected]

Abstract: Accurate estimates of forest aboveground biomass (AGB) are necessary for the
accurate tracking of forest carbon stock. Gaofen-7 (GF-7) is the first civilian sub-meter three-
dimensional (3D) mapping satellite from China. It is equipped with a laser altimeter system
and a dual-line array stereoscopic mapping camera, which enables it to synchronously
generate full-waveform LiDAR data and stereoscopic images. The bulk of existing research
has examined how accurate GF-7 is for topographic measurements of bare land or canopy
height. The measurement of forest aboveground biomass has not received as much attention
as it deserves. This study aimed to assess the GF-7 stereo imaging capability, displayed
as topographic features for aboveground biomass estimation in forests. The aboveground
Academic Editor: Xiaoyang Zhang biomass model was constructed using the random forest machine learning technique,
Received: 18 October 2024
which was accomplished by combining the use of in situ field measurements, pairs of GF-7
Revised: 20 December 2024 stereo images, and the corresponding generated canopy height model (CHM). Findings
Accepted: 21 December 2024 showed that the biomass estimation model had an accuracy of R2 = 0.76, RMSE = 7.94 t/ha,
Published: 27 December 2024 which was better than the inclusion of forest canopy height (R2 = 0.30, RMSE = 21.02 t/ha).
Citation: Liu, C.; Huang, H.; Zhang, These results show that GF-7 has considerable application potential in gathering large-scale
Z.; Fan, W.; Wu, D. Canopy Height high-precision forest aboveground biomass using a restricted amount of field data.
Integration for Precise Forest
Aboveground Biomass Estimation in
Keywords: Gaofen-7 (GF-7); stereophotogrammetry; canopy height; aboveground biomass
Natural Secondary Forests of
Northeast China Using Gaofen-7
Stereo Satellite Data. Remote Sens.
2025, 17, 47. https://doi.org/
10.3390/rs17010047 1. Introduction
Copyright: © 2024 by the authors. Forests play a crucial role in the terrestrial carbon sink system, acting as the largest
Licensee MDPI, Basel, Switzerland. carbon store among terrestrial ecosystems [1]. They absorb carbon dioxide from the atmo-
This article is an open access article sphere through photosynthesis, storing it in biomass (trees, plants) and soil, thus mitigating
distributed under the terms and climate change [2]. The accurate estimation of forest biomass is crucial for ecological stud-
conditions of the Creative Commons
ies and global change research [3,4]. The advancements in remote sensing technologies,
Attribution (CC BY) license
particularly LiDAR [5], multi-angle photogrammetry [6], and synthetic aperture radar
(https://creativecommons.org/
licenses/by/4.0/).

Remote Sens. 2025, 17, 47 https://doi.org/10.3390/rs17010047

Remote Sens. 2025, 17, 47 2 of 18

(SAR) [7], have significantly enhanced the capabilities of high-resolution mapping for forest
aboveground biomass (AGB).
The evolution of spaceborne remote sensing for forest AGB monitoring began in the
1980s and 1990s, characterized by the emergence of improved sensors and platforms such
as SPOT and RADARSAT, which provided enhanced vegetation monitoring capabilities
compared with earlier satellites like Landsat [8–11]. As a result, spaceborne remote sensing
has thus become crucial for accurately estimating forest AGB. In the 2000s, the launch
of high-resolution commercial satellites like QuickBird and IKONOS brought about a
revolution in sub-meter resolution imagery [6,12]. At the same time, advances in LiDAR
technology led to a significant improvement in biomass estimation accuracy through the
integration of optical, radar, and LiDAR data using sophisticated algorithms and machine
learning [13,14]. Such modern remote sensing capabilities are essential for effective forest
conservation, climate change studies, and management practices.
The application of advanced remote sensing technologies significantly enhances the
measurement of forest three-dimensional structures, thereby improving the estimation of
forest AGB. Satellite observation currently employs the following three primary techniques
to achieve this: LiDAR technology [5,15], synthetic aperture radar (SAR) [7], and stereopho-
togrammetry [6]. One of the most significant advancements in this domain has been the
development of spaceborne LiDAR systems, which can acquire terrain and height data for
extensive forest areas at reduced costs. Historically, satellite-based measurements of forest
canopy height were largely dependent on NASA’s Geoscience Laser Altimeter System
(GLAS) [16–18]. However, recent innovations, including the NASA Ice, Cloud, and land
Elevation Satellite-2 (ICESat-2) LiDAR instrument and the Global Ecosystem Dynamics
Investigation (GEDI) system, have markedly enhanced global forest monitoring capabili-
ties [15,19]. The upcoming BIOMASS mission, scheduled for launch in 2025, exemplifies
the integration of advanced P-band synthetic aperture radar (SAR) technology to monitor
changes in forest AGB and production [20]. Stereophotogrammetry has also seen consider-
able improvements through several satellite missions [21]. Notable examples include the
Chinese ZY-3 [22], the Japanese ALOS PRISM [23], SkySat satellites from Planet [24], and
the Worldview series from Maxar [25]. These missions have enhanced forest monitoring
by leveraging parallax range-dependent stereophotogrammetry. The launch of China’s
domestically developed Gaofen-7 (GF-7) satellite, which is equipped with advanced stereo
imaging capabilities, presents new opportunities for forest parameter investigation [26].
With a superior spatial resolution of 0.8 m for panchromatic imagery compared with ZY-3’s
2 m, GF-7 enables more precise assessments of forest structures and health [22,27].
Stereo imaging for forest height and biomass estimation presents distinct advantages
over spaceborne LiDAR systems like GEDI and ICESat. While GEDI and ICESat offer
precise measurements, they are costly to develop, launch, and operate. In contrast, stereo
imaging provides similar data at a fraction of the cost, making forest monitoring more fi-
nancially viable, especially for resource-constrained regions. Puliti et al. [28] demonstrated
the feasibility of predicting aboveground biomass in Norway by leveraging spaceborne
stereogrammetric digital surface models, specifically ArcticDEM data. Their findings likely
showcased a high degree of accuracy in biomass estimation, highlighting the potential of
stereogrammetric data for forest monitoring and management. Similarly, Zhang et al. [29]
illustrated the accuracy of estimating developing stem volume using forest canopy height
derived from ZY-3 stereoscopic data. While specific accuracy metrics may vary depending
on the methodologies employed, both studies likely reported promising results, indicat-
ing the reliability and precision of using stereoscopic satellite data for forest parameter
estimation. These findings underscore the importance and efficacy of spaceborne stere-
Remote Sens. 2025, 17, 47 3 of 18

ogrammetric data in advancing our understanding of forest ecosystems and supporting

sustainable resource management practices.
Additionally, stereo imaging achieves higher spatial resolutions, enabling detailed
characterization of forest structure and better delineation of individual trees and canopy
gaps [6,30]. This finer resolution facilitates more accurate estimates of forest height and
biomass at a finer spatial resolution. Moreover, satellites with stereo imaging sensors can
cover large areas in a single pass, providing comprehensive coverage of forested landscapes,
unlike the narrower swath widths of spaceborne LiDAR systems. The temporal consistency
of stereo imaging also enables continuous monitoring of forest dynamics over time, crucial
for accurately tracking changes. Overall, stereo imaging emerges as a cost-effective, high-
resolution, and globally accessible solution for monitoring and managing forest ecosystems,
complementing the capabilities of spaceborne LiDAR systems.
The GF-7 satellite’s high spatial resolution enables the creation of detailed three-
dimensional (3D) models of forested areas, providing valuable insights into forest structure
at a level of detail previously unseen. Although initially intended for other applications,
such as terrestrial elevation modeling [31] and urban planning [32–34], the rich information
captured by GF-7 imagery presents an opportunity to extract relevant forest structural
parameters. According to Du et al. [35], there has been a significant improvement in
the accuracy of canopy height estimation using stereo imagery from the GF-7 satellite
compared with data obtained from airborne laser scanning (ALS). Ni et al. [27] presented a
technique to accurately extract forest height using GF-7 very high-resolution stereoscopic
data, demonstrating the possibility for efficient regional and worldwide forest height
assessment without extensive fieldwork. The high spatial resolution of the GF-7 satellite
makes it easier to derive specific forest heights, which improves large-scale forest structures.
Despite these advancements, the potential of canopy height models (CHMs) derived
from GF-7 in conjunction with spectral data for biomass estimation remains largely unex-
amined. While existing research has predominantly focused on evaluating the accuracy
of GF-7 for topographic measurements on bare land or canopy height, there has been in-
sufficient attention to its application in estimating forest aboveground biomass. Therefore,
this study aims to address the following questions: (1) How effectively can GF-7 stereo
imaging capture topographic features for the estimation of aboveground biomass in forests?
(2) How can an aboveground biomass model be constructed using the random forest ma-
chine learning technique by integrating in situ field measurements, GF-7 stereo images, and
the corresponding generated digital surface model (DSM)? (3) What is the accuracy of the
biomass estimation model, in terms of R2 and RMSE, and how does GF-7 demonstrate its
potential for gathering large-scale high-precision forest aboveground biomass data? These
questions aim to determine the significant promise of GF-7 for large-scale and precise forest
AGB estimation, thereby enhancing our capabilities in forest monitoring and management.

2. Methods and Data

2.1. Study Area
The forest in this study is over 500 km2 in size, with dimensions of roughly 21.66 km
east to west and 24.14 km north to south. It is in Shangzhi City, Heilongjiang Province,
China (127◦ 20′ to 127◦ 50′ E, 45◦ 05′ to 45◦ 30′ N) (Figure 1). With a frost-free period of
approximately 125 days and an average temperature of 2.4 ◦ C, the area displays traits of a
subarctic monsoon climate. Rainfall is mostly concentrated from May to September, with
an average of 700 mm per year.
 pine trees. Spruce, larch, walnut, oak, elm, yew, birch, and poplar are among the major
tree species. The forest’s landform is part of a low-lying hilly region. The land has an av-
Remote Sens. 2025, 17, 47 erage elevation of 300 m and rises gradually from the south to the north. Maoer Mountain,
4 of 18
at 805 m above sea level, is the highest peak.

Figure 1. Location
Figure 1. Location mapmap of
of the
the study
study area
area (Shangzhi,
(Shangzhi, Heilongjiang,
Heilongjiang, China).
China). (a)
(a) The
The location
location of
of the
the
study area; (b) field plots over the GF-7 multispectral image on 20 August 2020 (R—red, G—green,
study area; (b) field plots over the GF-7 multispectral image on 20 August 2020 (R—red, G—green,
B—blue).
B—blue).

The study area’s forest resources make up a significant portion of northeastern China’s
2.2. Field Measurements
forest resources, making it an important ecological resource area. The mixed coniferous
Within this study region’s GF-7 image coverage area, 74 plots in total were created
and broad-leaved forests that make up most of the vegetation belt are home to pine trees.
in the year of 2019. The forest type and structure in the research region remained steady,
Spruce, larch, walnut, oak, elm, yew, birch, and poplar are among the major tree species.
even though the plot data and GF-7 satellite data were gathered one year apart. As a re-
The forest’s landform is part of a low-lying hilly region. The land has an average elevation
sult, the change in the forest AGB during this time was not seen as significant. All the plots
of 300 m and rises gradually from the south to the north. Maoer Mountain, at 805 m above
were randomly sampled as closely as possible to reflect the true distribution of the forest
sea level, is the highest peak.
given the remote access to the field plots. The plots had a 0.06-hectare area and were
square
2.2. FieldinMeasurements
shape. The GPS coordinates and relative positions within the stand were noted.
A Vertex
Within IVthis
device
studywas utilized
region’s to image
GF-7 measure the height
coverage area, of
74 trees
plots in
in each plot that
total were hadina
created
diameter at breast height (DBH) greater than 5 cm. For live standing trees,
the year of 2019. The forest type and structure in the research region remained steady, even measurements
of the trees’
though species,
the plot data DBH, crown
and GF-7 width,
satellite andwere
data height were made.
gathered The apart.
one year AGB of Aseach plot was
a result, the
determined by adding the values of the relevant species-specific allometric
change in the forest AGB during this time was not seen as significant. All the plots were equations,
which were
randomly derivedasusing
sampled closelytree
asheight
possibleandtoDBH.
reflectThe
theDBH-based allometric
true distribution of theequations for
forest given
calculating AGB of main tree species are listed in Table S2. AGB was measured
the remote access to the field plots. The plots had a 0.06-hectare area and were square in in the plot
with an
shape. Theaverage value of roughly
GPS coordinates 128.2positions
and relative t/ha, a minimum
within theofstand
29.4 t/ha,
wereand a maximum
noted. A Vertex IVof
276.5 t/ha.
device was utilized to measure the height of trees in each plot that had a diameter at breast
height (DBH) greater than 5 cm. For live standing trees, measurements of the trees’ species,
2.3. Data
DBH, and Methods
crown width, and height were made. The AGB of each plot was determined by
2.3.1. GF-7
adding Stereo of
the values Images and CHM
the relevant Retrieval allometric equations, which were derived
species-specific
usingGF-7,
tree height and
China’s DBH.
first The DBH-based
sub-meter allometric
high-resolution Earthequations for and
observation calculating
remote AGB of
sensing
main tree species are listed in Table S2. AGB was measured in the plot with an average
satellite, is meant to be used for industrial monitoring, natural resource monitoring, and
value of roughlyIt128.2
land surveying. was t/ha, a minimum
launched of 29.4 t/ha,
on 3 December 2019, and
froma Taiyuan,
maximum of 276.5
China, andt/ha.
it was put
into a 500 km sun-synchronous orbit, with a temporal resolution of 59 days. Unlike
2.3. Data and Methods
2.3.1. GF-7 Stereo Images and CHM Retrieval
GF-7, China’s first sub-meter high-resolution Earth observation and remote sensing
satellite, is meant to be used for industrial monitoring, natural resource monitoring, and
land surveying. It was launched on 3 December 2019, from Taiyuan, China, and it was
put into a 500 km sun-synchronous orbit, with a temporal resolution of 59 days. Unlike
traditional optical remote sensing satellites, the GF-7 satellite contains double-line cameras
Remote Sens. 2025, 17, 47 5 of 18

(DLCs) and could offer elevation data via stereo mapping technology. The GF-7 satellite
images used in this study were provided by the Chinese Land Satellite Remote Sensing
Application Center (http://www.lasac.cn/ access on 12 April 2024). To explore the rela-
tionship between DSM and forest canopy structure in different seasonal stereo pairs, we
selected two periods of images in August and November, 2020, respectively. Three images
make up each GF-7 imaging scene: two panchromatic images (with viewing angles of
−5◦ in the backward direction and +26◦ in the forward direction) and one multispectral
image containing bands of blue, green, red, and near-infrared. Furthermore, the four-band
backward multispectral imagery has a spatial resolution of 2.6 m, and the GF-7 forward
panchromatic and backward panchromatic images have respective spatial resolutions of
0.65 m and 0.8 m (see Table 1).

Table 1. Basic information of GF-7 DLC imagery.

GF-7 DLC Spectral Range Spatial Viewing Viewing Angle

Name
Imagery (nm) Resolution (m) Direction (◦ )
Blue: 450–520
Multispectral Green: 520–590
MUX 2.60 Backward −5.00
image Red: 630–690
NIR: 770–890
Panchromatic FWD 0.80 Forward 26.00
450–900
image BWD 0.65 Backward −5.00

The preprocessing steps for each GF-7 image include radiometric correction, atmo-
spheric correction, orthorectification, and BRDF correction. Initially, the raw images,
provided as digital numbers (DNs), are converted to physical quantities of radiance or
reflectance using the sensor’s calibration coefficients, with gain and offset values detailed
in the accompanying table. Atmospheric correction is performed to remove effects such
as aerosol and molecular scattering, allowing us to obtain surface reflectance using the
FLAASH module in ENVI 4.8 software. Orthorectification addresses geometric errors due
to terrain effects or sensor distortions by correcting the images to a uniform geographic
coordinate system through the RPC orthorectification module in ENVI. Lastly, to mitigate
BRDF effects caused by variations in solar elevation, azimuth, and observation angles, we
normalize the reflectance to standardized observation conditions using the Ross–Li model,
incorporating NASA’s MODIS MCD43A1 data in SNAP software. We will clarify in the
manuscript that both GF-7 images underwent these preprocessing steps, as they are critical
for enhancing the accuracy of model predictions.
Using the stereo image product in the rational polynomial coefficients (RPCs) file,
which provides the image coordinates and the ground field coordinates of the conversion
relationship, we used the image point with the same name using the matching method.
The RPC model was then corrected using ground control points (10–15 points/scene, this
study), effectively removing systematic errors from the model. Finally, we generated the
epipolar image by acquiring feature point and performing intensive matching using the
semi-global matching (SGM) algorithm, resulting in the production of the digital surface
model (DSM). Forest heights are then calculated as the difference between the stereoscopic
DSMs and the digital terrain model (DTM), which is a 1:10,000 digital elevation model
acquired by airborne LiDAR and is adequate for the requirements of our research. The
retrieval procedure of CHM can be referenced in Figure 2a.
Remote
RemoteSens.
Sens.2025,
2025,17,
17,x 47
FOR PEER REVIEW 6 of
6 of1818

Figure 2. The procedure for calculating forest canopy height and biomass from GF-7 stereoscopic
Figure 2. The procedure for calculating forest canopy height and biomass from GF-7 stereoscopic
imagery.
imagery.
2.3.2. Random Forest Model for AGB Estimation
2.3.2. Random Forest Model for AGB Estimation
To investigate whether DSM or CHM can be capitalized on to improve mixed forest
To investigate
AGB retrieval, whether regional
we mapped DSM or CHM can be
AGB maps incapitalized on tocompeting
2020 with four improve mixed forest
scenarios that
AGB retrieval, we mapped regional AGB maps in 2020 with four competing scenarios that
varied the incorporation of DSM or CHM as the model input, as illustrated in Table 2.
varied the incorporation of DSM or CHM as the model input, as illustrated in Table 2.
Table 2. Overview of scenarios for AGB mapping using different inputs from GF-7 data.
Table 2. Overview of scenarios for AGB mapping using different inputs from GF-7 data.
Scenario Description Inputs
Scenario Description Inputs
Traditional prediction
Traditional prediction algorithm Original GF-7 spectral
algorithm forfor
AGBOriginal GF-7 spectral bands + 33
AGB mapping using spectral features bands indices
multispectral + 33 multispectral
from GF-7
S1 mapping using spectral
S1
from GF-7 MUX images obtained indices from GF-7 MUX
features from during
GF-7 MUX MUX bands (see Supplement Ta-
bands (see Supplement
the growing season.
images obtained during the ble S1).
Table S1).
growing
Same as S1, but incorporates CHM season.
de- Inputs from S1 + CHM from DTM
S2 rived from digital terrainSame
modelas (DTM)
S1, but and GF-7 stereo image pairs (Au-
and GF-7 stereo image pairs in August.
incorporates CHM derived gust).
Inputs from S1 + CHM
S2 from digital terrain model from
Same as S1, includes CHM derived from Inputs from S1 + CHM DTM and GF-7
from stereo
DTM
S3 (DTM) and GF-7 stereo image pairs
DTM and GF-7 stereo image pairs in and GF-7 stereo image pairs (No- (August).
image pairs in August.
November. vember).
Remote Sens. 2025, 17, 47 7 of 18

Table 2. Cont.

Scenario Description Inputs

Same as S1, includes CHM
Inputs from S1 + CHM
derived from DTM and
S3 from DTM and GF-7 stereo
GF-7 stereo image pairs in
image pairs (November).
November.
Uses the full set of inputs
from S1 and different Inputs from S1 + DSM
digital surface model from GF-7 stereo image
S4
(DSM) derived from GF-7 pairs (August and
stereo image pairs in November).
August and November.

We selected the random forest (RF) regression model [36] implemented in a Matlab
environment to create all AGB maps. Compared with conventional regression algorithms,
the RF model possesses several advantages that make it well suited for our study. Other
than being computationally efficient and sophisticated at handling a high dimensionality
of input data, the RF demonstrates high prediction accuracy across a wide range of forest
ecosystems [13,16]. All the prediction algorithms (S1–S4) were trained and validated
with the same reference data. In addition to reserving seven independent samples for
AGB model comparison, all remaining field plots were utilized for model development.
The prediction procedure is briefly illustrated in Figure 2b. When building the model, the
dataset was randomly divided into two subsets, with 90% allocated for training and 10% for
model testing. We set the number of decision trees to 500 to obtain an unbiased estimate of
the generalization error and used the default number of variables to be tested for each split
(i.e., the square root of the number of input features) for all scenarios. The prediction results
were evaluated using the coefficient of determination (R2 ) for the relationship between
field-measured and predicted AGB, and the root mean square error (RMSE). This process
was repeated 50 times for each input scenario, with sample indexing randomized each
time to ensure that the division between training and testing samples was not influenced
by the original order of samples. We recorded R2 and RMSE values during each iteration,
selecting the model with the highest R2 and the lowest RMSE as the final inversion model.
It should be noted that the feature values of the plots, including spectral indices and canopy
height, were obtained by calculating the mean of all pixels covering the plots based on the
geographic coordinates of the plot’s center.
The flowchart in Figure 2 presents a detailed methodology for estimating forest
aboveground biomass (AGB) using digital surface model (DSM) and canopy height model
(CHM) data. The process begins with the collection of multitemporal satellite data and field
data, both of which are critical for ensuring accurate analysis. It encompasses two primary
stages: GF-7 forward (FWD) and GF-7 backward (BWD), with an emphasis on precise geo-
registration to facilitate effective elevation retrieval. This procedure produces seasonal DSM
and CHM products for both summer and fall, while field canopy height measurements are
utilized to validate the remote sensing data. Subsequently, the data undergoes downscaling
to align with the resolution of MUX imagery. Finally, this refined dataset informs the
development of a prediction model aimed at estimating forest AGB. It should be noted that
the CHM was downscaled to 2.6 m to ensure it matches the resolution of the MUX images,
facilitating subsequent analyses.
 Remote Sens. 2025, 17, 47 8 of 18

3. Results
mote Sens. 2025, 17, x FOR PEER REVIEW3.1. Forest Height Prediction 8 of 18
The August surface model and canopy height model displayed higher topographic
characteristics and more accurate estimations of forest height. Figure 3 displays the DSM
and CHM for August and November. The DSM for August displays more topographic
that 51.8% of the variance in canopy heights predicted by the August canopy height model
characteristics, such as ridges, valleys, and hilltops, than the DSM for November. We can
is explained by the field measurements. The August regression line closely follows the 1:1
also see that the ground elevation is higher in the north and lower in the south. August
line, suggesting a reasonable
CHM better captures fit withofa forest
the distribution slightheight.
underestimation
The scatter plotatinhigher
Figure 4tree heights. In
illustrates
contrast, the November
the relationship betweendata exhibited
canopy heightsa lower R using
predicted2 valueaof 0.45 and
canopy height a higher RMSE
model and field-of 2.96
m, showing
measured that 44.9%
heights foroftwothedifferent
variance inpoints:
time predictions
Augustfrom the November
and November. canopy
For August, theheight
regression analysis yielded an R 2 value of 0.52 and an RMSE of 2.14 m, indicating that
model is explained by the field data. The November regression line showed more devia-
51.8% of the variance in canopy heights predicted by the August canopy height model
tion from the 1:1 line, particularly underestimating tree heights more significantly.
is explained by the field measurements. The August regression line closely follows the
These results highlight the seasonal variability in the predictive performance of the
1:1 line, suggesting a reasonable fit with a slight underestimation at higher tree heights.
canopyIn height
contrast,model, indicating
the November datathe need for
exhibited further
a lower R2calibration
value of 0.45 toand
improve accuracy,
a higher RMSE par-
ticularly form,
of 2.96 November
showing that data.
44.9%Theof overall model,
the variance however,from
in predictions demonstrates
the November practical
canopyutility
for estimating
height modelcanopy heightsbyinthe
is explained different seasons,
field data. informing
The November forest management
regression line showed more practices
deviation from
and ecological studies. the 1:1 line, particularly underestimating tree heights more significantly.

FigureFigure
3. The3. August and
The August November
and November DSM andCHM.
DSM and CHM.ThisThis figure
figure showsshows only
only the DSMthe
andDSM
CHMand
for CHM
for thethe commonregions
common regions between
betweenAugust
Augustandand
November, highlighted
November, by the read
highlighted box.read
by the (a,b)box.
DSMs for DSMs
(a,b)
August and November, respectively, and (e,f) show the larger detail plots in the red boxes. (c,d) CHMs
for August and November, respectively, and (e,f) show the larger detail plots in the red boxes. (c,d)
for August and November, respectively, and (g,h) show the larger detail plots in the red boxes.
CHMs for August and November, respectively, and (g,h) show the larger detail plots in the red
boxes.
 Remote Sens. 2025, 17, x FOR PEER REVIEW 9 of 18
Remote Sens. 2025, 17, 47 9 of 18

Figure The
4. 4.
Figure scatter
The plot
scatter illustrates
plot thethe
illustrates relationship between
relationship canopy
between heights
canopy predicted
heights usingusing
predicted a canopy
a can-
height model and field-measured heights for two different time points: August and November.
opy height model and field-measured heights for two different time points: August and November. The
light green points and corresponding regression line represent the August data, while the light blue
The light green points and corresponding regression line represent the August data, while the light
points and their regression line represent the November data. The 1:1 line (grey dashed) indicates
blue points and their regression line represent the November data. The 1:1 line (grey dashed) indi-
perfect concordance between predicted and measured heights.
cates perfect concordance between predicted and measured heights.
These results highlight the seasonal variability in the predictive performance of the
3.2. Feature
canopy height Importance in Biomass the
model, indicating Prediction
need for further calibration to improve accuracy,
Random
particularly forests were
for November employed
data. to assess
The overall model,thehowever,
relative importance
demonstrates of various
practicalvariables
utility
forfor predicting
estimating AGB, heights
canopy using the in “mean decrease
different seasons,ininforming
MSE” metric (see
forest Figure 5). The
management S1 sce-
practices
nario
and included
ecological multispectral indices computed from GF-7 MUX bands and the original
studies.
GF-7 spectral bands, while the S2 scenario added CHM data derived from DTM and GF-
3.2. Featureimages
7 stereo Importance in Biomass
collected Prediction
in August. In Scenario S1, TGI showed the highest importance,
Random
followed by forests
SGI and were employed
class, while BAI to was
assessthethe relative
least importance
sensitive variable. ofThe
various
finalvariables
predictive
for predicting AGB, using the “mean decrease in MSE” metric
variables selected were EVI, GEMI, GARI, GLI, IOR, LAI, RGRI, SGI, TGI, VARI, and (see Figure 5). Theclass,
S1
scenario
resultingincluded
in an Rmultispectral
2 of 0.71 and an indices
RMSEcomputed fromInGF-7
of 47.60 t/ha. MUXS2,
Scenario bands
CHM and the original
emerged as the
GF-7 spectral bands, while the S2 scenario added CHM data derived from
most critical variable, along with class and SGI, with BAI again being the least important. DTM and GF-7
stereo images
The final collected
variable in August.
set included In GLI,
EVI, Scenario
IOR, S1,
LAI, TGI showed
RGRI, RDVI,theSGI,
highest
TGI, importance,
VARI, CHM,
followed by SGI and class, while BAI was the least sensitive variable.
and class, significantly improving the model’s accuracy to an R of 0.90 and an 2 The final predictive
RMSE of
variables selected
20.01 t/ha. Thesewere EVI, GEMI,
findings GARI,
highlight theGLI, IOR, LAI,
substantial RGRI,
role SGI, TGI,
of CHM VARI, andbiomass
in enhancing class,
resulting in an R 2 of 0.71 and an RMSE of 47.60 t/ha. In Scenario S2, CHM emerged as the
prediction accuracy, emphasizing the need to incorporate structural information along-
most
sidecritical
spectral variable,
indices,along
and the with class andlow
consistent SGI, with BAI again
importance of BAIbeing
across the least
both important.
scenarios.
The final variable set included EVI, GLI, IOR, LAI, RGRI, RDVI, SGI, TGI, VARI, CHM,
and class, significantly improving the model’s accuracy to an R2 of 0.90 and an RMSE of
20.01 t/ha. These findings highlight the substantial role of CHM in enhancing biomass
Remote Sens. 2025, 17, 47 10 of 18

prediction
Remote Sens. 2025, 17, x FOR PEER REVIEW accuracy, emphasizing the need to incorporate structural information alongside
10 of 18
spectral indices, and the consistent low importance of BAI across both scenarios.

Figure5.
Figure 5. Feature importance
importancescores
scoresfor predicting
for AGB
predicting AGBusing random
using forest
random models
forest underunder
models two scenar-
two
scenarios: S1 and S2. Features are ranked in decreasing order of importance based on the meanin
ios: S1 and S2. Features are ranked in decreasing order of importance based on the mean decrease
mean squared
decrease in mean error (MSE).error
squared The feature
(MSE). “class” refers“class”
The feature to landrefers
coverto
classification
land cover data, distinguishing
classification data,
between forested and non-forested areas, derived from geographic national condition data.
distinguishing between forested and non-forested areas, derived from geographic national condi-
tion
3.3.data.
Continuous AGB Mapping
Incorporating CHM and DSM data significantly enhances biomass prediction accuracy.
3.3. Continuous AGB Mapping
The spatial distribution of the predicted biomass is shown in Figure 6. The scatter plot in
Incorporating
Figure 7 highlightsCHM and DSM data
the relationship significantly
between enhances
field-measured biomass
biomass andprediction
model-predictedaccu-
racy. The spatial distribution of the predicted biomass is shown
biomass across four scenarios (S1–S4). Regression analysis indicates a marked improvement in Figure 6. The scatter
plot in Figure
in model 7 highlights
performance the relationship
metrics (R2 and RMSE) between when field-measured
CHM and DSM biomass
data areandincluded
model-
predicted biomass across four scenarios (S1–S4). Regression analysis
alongside traditional spectral indices. Specifically, scenarios S2 and S3, which incorporate indicates a marked
improvement
CHM derivedinfrom model performance
different periods, metrics (R2 andenhanced
demonstrate RMSE) when CHMover
accuracy and scenario
DSM data S1,
are included alongside traditional spectral indices. Specifically, scenarios
which only utilizes spectral data. Scenario S4, integrating DSM, also shows significant S2 and S3, which
incorporate CHM derived from
predictive improvements. These different
results periods,
underscore demonstrate
the potential enhanced
of CHM accuracy
and DSM overin
scenario S1, which only utilizes spectral data. Scenario S4, integrating
refining AGB retrieval models, suggesting that structural parameters from remote sensing DSM, also shows
significant predictive
data critically augment improvements.
conventional These results
spectral underscore the potential of CHM and
approaches.
DSM in refining AGB retrieval models, suggesting
The models across the four scenarios are summarized that structural in parameters
Table 3, which frompresents
remote
sensing dataintercepts,
the slopes, critically augment
minimum conventional spectral approaches.
(Min) and maximum (Max) values, and percentiles (Q1,
The models across the four scenarios
median, Q3) of each model. This table highlights are summarized
that theinincorporation
Table 3, which ofpresents
the CHM thein
slopes,
modelsintercepts,
S2 and S3 minimum
significantly (Min) and maximum
improves (Max) values,
biomass estimates. and percentiles
Specifically, Model S2 (Q1, me-
utilizes
dian, Q3) of each model. This table highlights that the incorporation
CHM derived from DTM and GF-7 stereo images captured in August, achieving enhanced of the CHM in models
S2 and S3 significantly
prediction accuracy with improves
a slope ofbiomass
0.86 andestimates. Specifically,
a stable intercept Model
of 15.58. S2 utilizes
Looking at the CHM
results
derived from DTM and GF-7 stereo images captured in August,
from different models, Model S2 exhibits a minimum value of 120.81 t/ha and a maximum achieving enhanced pre-
diction
of 171.31accuracy withpercentiles
t/ha, with a slope of (Q1: 0.86 133.06,
and a stable
median:intercept
138.48,ofQ3:15.58. Looking
156.33) at therelatively
reflecting results
from different models, Model S2 exhibits a minimum value of 120.81
tight clustering around the median, which highlights consistent predictions across the t/ha and a maximum
ofdataset.
171.31 t/ha, with percentiles
In contrast, Model S3, (Q1: 133.06,
despite median:
also including138.48, Q3: yields
CHM, 156.33)areflecting
lower slope relatively
of 0.38,
tight
with a minimum of 110.71 t/ha and a maximum of 158.13 t/ha. Its percentiles (Q1:the
clustering around the median, which highlights consistent predictions across da-
139.11,
taset. In contrast,
median: 155.46, Q3:Model S3, despite
157.23) suggestalso including
increased CHM,in
variability yields a lower likely
predictions, slope influenced
of 0.38, with by
aseasonal
minimum of 110.71 t/ha and a maximum of 158.13 t/ha. Its percentiles
changes affecting biomass assessment. Interestingly, Model S4, which incorporates (Q1: 139.11, me-
dian: 155.46, Q3: 157.23)
a comprehensive set of suggest
inputs, increased variability
shows a notable in predictions,
decline in predictive likely influenced
capability, withbya
seasonal changes affecting biomass assessment. Interestingly,
slope of 0.19. Its minimum value of 129.67 t/ha and maximum of 152.84 t/ha, along Model S4, which incorpo-
rates
withapercentiles
comprehensive set of inputs,
indicating shows a predictions
more uniform notable decline (Q1:in137.16,
predictive capability,
median: 140.83,with
Q3:
a142.85),
slope ofsuggest
0.19. Itsthat,
minimum value of 129.67 t/ha and maximum of 152.84
while it captures a broader range, the complexity of inputs may dilute t/ha, along with
percentiles indicating more uniform predictions (Q1: 137.16, median: 140.83, Q3: 142.85),
suggest that, while it captures a broader range, the complexity of inputs may dilute pre-
dictive power. Overall, the findings emphasize the critical role of CHM in enhancing the
accuracy of forest biomass estimation.
Remote Sens. 2025, 17, 47 11 of 18

Remote Sens. 2025, 17, x FOR PEER REVIEW 11 of 18

predictive power. Overall, the findings emphasize the critical role of CHM in enhancing
the accuracy of forest biomass estimation.

Figure 6. Predicted
Figure 6. Predicted biomass
biomass maps
maps in
in different
different scenarios:
scenarios: (a)
(a) for
for S1
S1 scenario
scenario and
and (b)
(b) for
for S2
S2 scenario.
scenario.
Detailed drawings of the red-framed area are shown in Figure
Detailed drawings of the red-framed area are shown in Figure 8. 8.

Table 3. Statistical metrics of biomass estimation models.

Table 3. Statistical metrics of biomass estimation models.
Scenario
Scenario Slope
Slope Intercept
Intercept Min
Min Max
Max Q1
Q1 Median
Median Q3
Q3
S1
S1 0.84
0.84 21.23
21.23 92.79
92.79 171.97
171.97 136.36
136.36 154.14
154.14 165.68
165.68
S2
S2 0.86
0.86 15.58
15.58 120.81
120.81 171.31
171.31 133.06
133.06 138.48
138.48 156.33
156.33
S3
S3 0.38
0.38 88.72
88.72 110.71
110.71 158.13
158.13 139.11
139.11 155.46
155.46 157.23
157.23
S4
S4 0.19
0.19 112.14
112.14 129.67
129.67 152.84
152.84 137.16
137.16 140.83
140.83 142.85
142.85

Figure 8 illustrates the impact of CHM on biomass estimation across various scenarios.
The results highlight the significance of incorporating CHM in improving biomass predic-
tions. Scenario (a) shows that disturbances from water bodies and soil moisture in river
valley deltas lead to spectral confusion, causing biomass underestimation when relying
solely on spectral properties. Scenario (b) demonstrates that, in the flat eastern region
with lower canopy heights and predominant coniferous species, CHM results incorporate
more accurate biomass estimates. Scenario (c) reveals that, during the growing season,
spectral features alone cause an overestimation of farmland vegetation height, with the
average farmland biomass being less than 5 t/ha; however, predictions improve with CHM
inclusion. Scenario (d) emphasizes that at logging sites with low biomass, the use of CHM
leads to more precise biomass distribution predictions. These findings underscore the criti-
Remote Sens. 2025, 17, 47 12 of 18

Remote Sens. 2025, 17, x FOR PEERcal role

REVIEW of canopy height information in enhancing the accuracy of biomass assessments,
12 of 18
particularly in areas with topographic and spectral variabilities.

Figure Scatterplots
Figure7.7. Scatter plotsdepicting
depicting thethe relationship
relationship between
between predictedpredicted
biomassbiomass
(t/ha) and (t/ha) and field-
field-meas-
measured biomass
ured biomass (t/ha)(t/ha) for different
for four four different scenarios
scenarios (S1, S2,(S1,
S3, S2,
andS3, S4)and S4) inDetailed
in 2020. 2020. Detailed
scenario scenario
de-
descriptions
scriptions are areprovided
provided in in Table
Table 2. Scatter
2. (a) (a) Scatter plot includes
plot includes a regression
a regression line,
line, with with annotations
annotations dis-
displaying theregression
regressionequation,
equation, coefficient 2 ), and root mean square error
playing the coefficient of of determination
determination (R2),(Rand root mean square error
(RMSE)
(RMSE) totoquantitatively
quantitatively assess modelperformance.
assess model performance.(b)(b) Residuals
Residuals for for
each each model’s
model’s prediction
prediction com- com-
pared
paredwith
withfield
fieldbiomass.
biomass. The results
resultsdemonstrate
demonstrateincremental
incremental improvements
improvements in biomass
in biomass prediction
prediction
accuracy
accuracyfrom
fromS1 S1totoS4,
S4, highlighting
highlighting the thesignificant
significantimpact
impact of of incorporating
incorporating CHM CHMand and
DSM DSM
data. data.
Remote Sens. 2025, 17, x FOR PEER REVIEW 13 of 18
Scenarios S2 and S3 show enhanced prediction accuracy due to the inclusion
Scenarios S2 and S3 show enhanced prediction accuracy due to the inclusion of detailed canopy of detailed canopy
height information.
height information.

Figure 8 illustrates the impact of CHM on biomass estimation across various scenar-
ios. The results highlight the significance of incorporating CHM in improving biomass
predictions. Scenario (a) shows that disturbances from water bodies and soil moisture in
river valley deltas lead to spectral confusion, causing biomass underestimation when re-
lying solely on spectral properties. Scenario (b) demonstrates that, in the flat eastern re-
gion with lower canopy heights and predominant coniferous species, CHM results incor-
porate more accurate biomass estimates. Scenario (c) reveals that, during the growing sea-
son, spectral features alone cause an overestimation of farmland vegetation height, with
the average farmland biomass being less than 5 t/ha; however, predictions improve with
CHM inclusion. Scenario (d) emphasizes that at logging sites with low biomass, the use
of CHM leads to more precise biomass distribution predictions. These findings under-
score the critical role of canopy height information in enhancing the accuracy of biomass
assessments, particularly in areas with topographic and spectral variabilities.

8. Biomass,
Figure
Figure 8. Biomass,spectrum,
spectrum, and canopyheight
and canopy heightspatial
spatial features
features at the
at the samesame location.
location. The biomass
The biomass
detail
detail mapmap
on on
thethe
farfar
leftleft shows
shows thethe result
result ofof
thethe
S1S1scenario,
scenario,while
whilethe
theone
oneon
on the
the right
right shows
shows the
the result
result of the S2 scenario. The biomass ramp is consistent with that shown in Figure 6, and the RGB
channel denotes the real color channel display of GF-7. The CHM ramp is similar to those shown in
Figure 3. The last column shows land cover, with black indicating forested area and white repre-
senting non-forest land. Figure (a): the disturbance of water bodies and soil moisture on the river
valley delta causes the forest vegetation spectra to be mistaken for bare soil and water bodies, which
Remote Sens. 2025, 17, 47 13 of 18

of the S2 scenario. The biomass ramp is consistent with that shown in Figure 6, and the RGB channel
denotes the real color channel display of GF-7. The CHM ramp is similar to those shown in Figure 3.
The last column shows land cover, with black indicating forested area and white representing non-
forest land. Figure (a): the disturbance of water bodies and soil moisture on the river valley delta
causes the forest vegetation spectra to be mistaken for bare soil and water bodies, which leads to an
underestimating of biomass forecast based solely on spectral properties. The regional variability of
forest species and height under various topographic circumstances is depicted in Figure (b). Because
of the region’s eastern side’s relative flatness, low forest heights, and predominance of coniferous
tree species, biomass estimations that consider CHM factors are more accurate in reflecting the real
distribution. Figure (c): due to the overestimation of the height of the farmland vegetation caused
by spectral features alone (August during the growing season), the farmland’s spectrum is viewed
as being spectral like the forest. In agriculture, the average biomass is less than 5 t/ha, although
biomass is more precisely anticipated because of the constraint that CHM is approximately 0 t/ha.
A logging site is in the region of figure (d), with low biomass. A more accurate prediction of the
biomass distribution is made by taking canopy height features into account.

4. Discussion
4.1. The Importance of Canopy Height in Biomass Estimation
The results presented in Figure 7 highlight the critical importance of incorporating
CHM data for accurate AGB predictions. Our analysis compares four scenarios with
varying input datasets: traditional spectral-only models (S1), models incorporating CHM
derived from August (S2) and November (S3), and models utilizing digital surface models
(DSMs) from different times (S4). Among these, Scenario S2, which includes CHM data
from August, exhibits the highest model performance, as evidenced by the highest R2
and the lowest root mean square error (RMSE). This indicates that the addition of precise
canopy height information, particularly during the growing season when the canopy is
fully developed, substantially enhances the accuracy of biomass predictions. Conversely,
Scenario S3, which uses CHM derived from November (a leaf-off period), shows reduced
performance, underscoring the influence of seasonal conditions on the effectiveness of CHM.
Although Scenario S4, incorporating DSM data, improves biomass prediction compared
with the spectral-only approach (S1), its performance does not surpass that of S2. This
comparison suggests that, while DSM data are beneficial, CHM from summer conditions is
more valuable in capturing canopy structure for biomass estimation.
These findings align with previous research. For example, Ni et al. [22] examines the
seasonal effects on AGB estimation in mountainous deciduous forests using ZY-3 stereo-
scopic imagery, demonstrating that data collected during the growing season provide
better AGB estimation due to a fully developed canopy structure. Our results similarly
emphasize the superior accuracy of CHM derived from summer imagery. Moreover, the
study by Gong et al. [37] on improving an oak canopy model using digital photogram-
metry further supports these findings, highlighting the enhancement in canopy structure
estimation when precise height information is used. While Ni et al. [22] focus specifically
on mountainous deciduous forests and the ZY-3 platform, and Gong et al. [37] address
oak canopy models with photogrammetry, our study extends these findings by comparing
multiple scenarios and incorporating data from the GF-7 satellite. Additionally, our study
highlights the relative benefits of CHM versus DSM data, underscoring that CHM collected
during optimal conditions (summer) provides more accurate biomass predictions than
DSM when included alongside traditional spectral indices. In summary, both our study
and the referenced works underscore the importance of seasonal timing and precise canopy
height information in remote sensing for biomass estimation. The inclusion of CHM during
the growing season is significantly more effective, as it captures detailed canopy structure
Remote Sens. 2025, 17, 47 14 of 18

crucial for accurate AGB retrieval. These findings collectively enhance our understanding
of the optimal methodologies for AGB estimation in diverse environments.

4.2. Seasonal Variability in Canopy Height Model Performance

The comparison between the predicted canopy heights and field-measured heights for
both August and November reveal several key aspects of model behavior and accuracy.
The regression analysis for August data yield an R2 value of 0.52 and an RMSE of 2.14 m,
whereas the November data exhibit an R2 value of 0.45 and an RMSE of 2.96 m. The
higher R2 and lower RMSE for the August data suggest that the canopy height model
performs better during this month. The light green points and their regression line for
August are more closely aligned with the 1:1 line, indicating a relatively higher accuracy in
height predictions during this period. Conversely, the light blue points and regression line
for November deviate more significantly from the 1:1 line, particularly underestimating
canopy heights. This seasonal variability in model performance may result from denser
foliage and more defined canopy structures in August, providing clearer signals for the
model, while the partial defoliation and structural changes in November introduce noise
and variability, degrading the model’s predictive performance. Similar seasonal effects on
remote sensing accuracy have been documented by Ni et al. [38], who found substantial
differences in ground surface elevation extraction over deciduous forests between seasons.
Ni et al. [22] further explain that the tops of the forest canopy are easily apparent in closed
forest stands throughout the leaf-on season; in open forest stands, the ground surface is
visible in mid-summer through canopy gaps, but it may be obscured in late summer or
early fall.
These results underscore the necessity for careful selection of stereo satellite images
when deriving canopy height models. The demonstrated seasonal variability suggests
that images captured during periods with fully developed canopy structures (e.g., late
summer) yield more accurate predictions. Therefore, for critical applications requiring high
precision, such as forest carbon stock assessment or habitat monitoring, it is advisable to
select stereo pairs from seasons that minimize foliage variability. Despite this variability,
the canopy height model shows practical utility for large-scale canopy height estimation,
critical for forest management and ecological monitoring. The promising performance
in August suggests prioritizing this period for high-accuracy surveys. For periods like
November, where the model’s performance diminishes, incorporating more field measure-
ments can validate and complement model outputs. Further integration of multi-temporal
datasets and advanced remote sensing technologies, such as LiDAR, could significantly
improve model robustness and accuracy across different seasons, facilitating better forest
assessments and ecological studies.

4.3. Enhancing AGB Mapping: The Role of Optical Stereo Images

The integration of optical stereo images for AGB mapping offers a promising com-
plement to traditional methods like LiDAR [13,39] and SAR [10,40], providing several
advantages and some challenges. Optical stereo images are cost-effective, accessible,
and provide high spatial resolution, which enhances the detail and accuracy of biomass
estimates by capturing fine-scale vegetation variations. They also offer high temporal
resolution, allowing for frequent monitoring and understanding of seasonal dynamics. Ad-
ditionally, when combined with LiDAR and SAR, optical images can significantly improve
biomass estimation accuracy [13]. However, they are dependent on weather conditions,
such as cloud cover, and require complex data processing to derive accurate CHM. More-
over, data saturation in dense forests can limit their effectiveness [41,42]. Prospects for
optical stereo images in biomass mapping include integrating multi-source data to leverage
Remote Sens. 2025, 17, 47 15 of 18

the strengths of each method, advancing image processing algorithms, and employing
machine learning and AI to refine analysis and predictive models. Establishing global
monitoring programs utilizing optical stereo imagery can provide consistent up-to-date
biomass data, supporting efforts in climate change mitigation, biodiversity conservation,
and sustainable forest management. Despite the challenges, rapid advancements are likely
to enhance the utility of optical stereo images, making them a crucial component in the
future of forest monitoring.

4.4. Accuracy Comparisons with LiDAR and Other Remote Sensing Methods
In this study, we recognize the importance of discussing the accuracy of canopy
height and biomass estimates derived from remote sensing techniques. While our results
demonstrate that the incorporation of the CHM from GF-7 significantly enhances biomass
estimation accuracy (R2 = 0.76, RMSE = 7.94 t/ha), it is crucial to consider how these
results align with accuracy standards set by other technologies. Previous studies have
indicated that LiDAR systems, such as NASA’s GEDI, typically achieve high accuracy
levels, with R2 values often exceeding 0.8 in forest environments [43]. Furthermore, studies
have shown that airborne LiDAR can provide detailed canopy structure information, often
resulting in lower RMSE values compared with non-LiDAR approaches [44]. However,
we must recognize that geographical variability, the specific forest types examined, and
the limited number of ground truth measurements may introduce uncertainties into our
biomass estimates, potentially limiting the broader applicability of our findings. To frame
our results within a broader context, we refer to the literature, indicating that “good”
accuracy in AGB prediction is typically characterized as an R2 value of 0.70 or higher [15,41].
Although direct comparisons at the plot level may not be feasible with our current dataset,
exploring aggregate comparisons at larger spatial scales can yield valuable insights into the
relative efficiency and accuracy of satellite photogrammetry versus LiDAR methods such
as ICESat-1/2 [16,39] or GEDI [15], which provide point sampling data for AGB prediction.
Regarding the differences in results between spaceborne photogrammetry and air-
borne methods, while our study focuses primarily on spaceborne stereo imagery for esti-
mating AGB, we acknowledge that understanding these differences is important. Airborne
methods for predicting AGB offer several advantages, including higher spatial resolution
and detailed canopy structure information, which enhance accuracy in AGB estimates.
However, these methods are limited by smaller coverage areas and often higher costs
associated with data acquisition [45]. By situating our findings within the larger frame-
work of remote sensing technologies that integrate spaceborne LiDAR for sampling with
stereo photogrammetry or SAR imagery for comprehensive AGB mapping, we aim to offer
a clearer understanding of the strengths and limitations inherent to each method. This
knowledge will ultimately support informed decision making in forest management and
carbon stock assessments.

5. Conclusions
In this study, we have shown that the GF-7 very high-resolution stereo satellite is
a useful tool for AGB and forest canopy height estimation. Our findings demonstrate
that the accuracy of biomass prediction is much increased by the incorporation of canopy
height model (CHM) data, especially when the data are collected from summer images
when the canopy is fully developed. When compared with spectral-only models, the
random forest model performs better in scenarios involving CHM, producing an R2 = 0.76,
RMSE = 7.94 t/ha. This high degree of accuracy highlights the value of GF-7 in recording
specific forest structural metrics, offering a high-resolution, globally accessible, and reason-
ably priced tool for extensive forest monitoring. Overall, the GF-7 satellite demonstrates
Remote Sens. 2025, 17, 47 16 of 18

significant potential in advancing remote sensing applications in forestry. Our approach

offers significant improvements in the precision of AGB estimates by incorporating struc-
tural information, which is crucial for ecological studies, forest management, and climate
change mitigation efforts.

Supplementary Materials: The following supporting information file can be downloaded at: https:
//www.mdpi.com/article/10.3390/rs17010047/s1.

Author Contributions: Conceptualization, D.W. and C.L.; methodology, D.W. and C.L.; validation,
W.F. and H.H.; writing—original draft preparation, C.L.; writing—review and editing, Z.Z., H.H.
and W.F.; visualization, C.L. and D.W. All authors have read and agreed to the published version of
the manuscript.

Funding: This research was supported by the National Key Research and Development Program of
China (grant no. 2023YFB3907702), the National Natural Science Foundation of China (42371339),
and the Young Elite Scientist Sponsorship Program by Heilongjiang Province (2022QNTJ011).

Data Availability Statement: The data that support the findings of this study can be obtained from
the corresponding author upon reasonable request.

Conflicts of Interest: The authors declare no conflict of interest.

References
1. Bonan, G.B. Forests and Climate Change: Forcings, Feedbacks, and the Climate Benefits of Forests. Science 2008, 320, 1444–1449.
[CrossRef] [PubMed]
2. Shugart, H.; Saatchi, S.; Hall, F. Importance of structure and its measurement in quantifying function of forest ecosystems.
J. Geophys. Res. Biogeosci. 2010, 115. [CrossRef]
3. Fang, J.-y.; Wang, G.G.; Liu, G.-h.; Xu, S.-l. Forest biomass of China: An estimate based on the biomass–volume relationship. Ecol.
Appl. 1998, 8, 1084–1091.
4. Yang, H.; Ciais, P.; Frappart, F.; Li, X.; Brandt, M.; Fensholt, R.; Fan, L.; Saatchi, S.; Besnard, S.; Deng, Z. Global increase in biomass
carbon stock dominated by growth of northern young forests over past decade. Nat. Geosci. 2023, 16, 886–892. [CrossRef]
5. Dubayah, R.O.; Drake, J.B. Lidar remote sensing for forestry. J. For. 2000, 98, 44–46. [CrossRef]
6. St-Onge, B.; Hu, Y.; Vega, C. Mapping the height and above-ground biomass of a mixed forest using lidar and stereo Ikonos
images. Int. J. Remote Sens. 2008, 29, 1277–1294. [CrossRef]
7. Minh, D.H.T.; Le Toan, T.; Rocca, F.; Tebaldini, S.; d’Alessandro, M.M.; Villard, L. Relating P-band synthetic aperture radar
tomography to tropical forest biomass. IEEE Trans. Geosci. Remote Sens. 2013, 52, 967–979. [CrossRef]
8. Sader, S.A.; Waide, R.B.; Lawrence, W.T.; Joyce, A.T. Tropical forest biomass and successional age class relationships to a vegetation
index derived from Landsat TM data. Remote Sens. Environ. 1989, 28, 143–198.
9. Castillo-Santiago, M.A.; Ricker, M.; de Jong, B.H. Estimation of tropical forest structure from SPOT-5 satellite images. Int. J.
Remote Sens. 2010, 31, 2767–2782. [CrossRef]
10. Stelmaszczuk-Górska, M.A.; Urbazaev, M.; Schmullius, C.; Thiel, C. Estimation of above-ground biomass over boreal forests in
siberia using updated in situ, ALOS-2 PALSAR-2, and RADARSAT-2 data. Remote Sens. 2018, 10, 1550. [CrossRef]
11. Shi, Y.; Wang, Z.; Zhang, G.; Wei, X.; Ma, W.; Yu, H. Evaluating the Research Status of the Remote Sensing-Mediated Monitoring
of Forest Biomass: A Bibliometric Analysis of WOS. Forests 2024, 15, 524. [CrossRef]
12. Chen, G.; Hay, G.J.; St-Onge, B. A GEOBIA framework to estimate forest parameters from lidar transects, Quickbird imagery and
machine learning: A case study in Quebec, Canada. Int. J. Appl. Earth Obs. Geoinf. 2012, 15, 28–37. [CrossRef]
13. Huang, H.; Liu, C.; Wang, X.; Zhou, X.; Gong, P. Integration of multi-resource remotely sensed data and allometric models for
forest aboveground biomass estimation in China. Remote Sens. Environ. 2019, 221, 225–234. [CrossRef]
14. Simard, M.; Pinto, N.; Fisher, J.B.; Baccini, A. Mapping forest canopy height globally with spaceborne lidar. J. Geophys. Res.
Biogeosci. 2011, 116. [CrossRef]
15. Duncanson, L.; Kellner, J.R.; Armston, J.; Dubayah, R.; Minor, D.M.; Hancock, S.; Healey, S.P.; Patterson, P.L.; Saarela, S.; Marselis,
S. Aboveground biomass density models for NASA’s Global Ecosystem Dynamics Investigation (GEDI) lidar mission. Remote
Sens. Environ. 2022, 270, 112845. [CrossRef]
16. Huang, H.; Liu, C.; Wang, X.; Biging, G.S.; Chen, Y.; Yang, J.; Gong, P. Mapping vegetation heights in China using slope correction
ICESat data, SRTM, MODIS-derived and climate data. ISPRS J. Photogramm. Remote Sens. 2017, 129, 189–199. [CrossRef]
Remote Sens. 2025, 17, 47 17 of 18

17. Wang, X.; Huang, H.; Gong, P.; Liu, C.; Li, C.; Li, W. Forest canopy height extraction in rugged areas with ICESAT/GLAS data.
IEEE Trans. Geosci. Remote Sens. 2013, 52, 4650–4657. [CrossRef]
18. Liu, C.; Wang, X.; Huang, H.; Gong, P.; Wu, D.; Jiang, J. The importance of data type, laser spot density and modelling method for
vegetation height mapping in continental China. Int. J. Remote Sens. 2016, 37, 6127–6148. [CrossRef]
19. Tang, H.; Huang, H.; Zheng, Y.; Qin, P.; Xu, Y.; Ding, S. Improved GEDI canopy height extraction based on a simulated ground
echo in topographically undulating areas. IEEE Trans. Geosci. Remote Sens. 2023, 61, 1–15. [CrossRef]
20. Quegan, S.; Le Toan, T.; Chave, J.; Dall, J.; Exbrayat, J.-F.; Minh, D.H.T.; Lomas, M.; D’alessandro, M.M.; Paillou, P.; Papathanassiou,
K. The European Space Agency BIOMASS mission: Measuring forest above-ground biomass from space. Remote Sens. Environ.
2019, 227, 44–60. [CrossRef]
21. Ni, W.; Zhang, Z.; Sun, G.; Liu, Q. Modeling the stereoscopic features of mountainous forest landscapes for the extraction of
forest heights from stereo imagery. Remote Sens. 2019, 11, 1222. [CrossRef]
22. Ni, W.; Yu, T.; Pang, Y.; Zhang, Z.; He, Y.; Li, Z.; Sun, G. Seasonal effects on aboveground biomass estimation in mountainous
deciduous forests using ZY-3 stereoscopic imagery. Remote Sens. Environ. 2023, 289, 113520. [CrossRef]
23. Huang, H.; Chen, P.; Xu, X.; Liu, C.; Wang, J.; Liu, C.; Clinton, N.; Gong, P. Estimating building height in China from ALOS
AW3D30. ISPRS J. Photogramm. Remote Sens. 2022, 185, 146–157. [CrossRef]
24. Bhushan, S.; Shean, D.; Alexandrov, O.; Henderson, S. Automated digital elevation model (DEM) generation from very-high-
resolution Planet SkySat triplet stereo and video imagery. ISPRS J. Photogramm. Remote Sens. 2021, 173, 151–165. [CrossRef]
25. Meddens, A.J.; Vierling, L.A.; Eitel, J.U.; Jennewein, J.S.; White, J.C.; Wulder, M.A. Developing 5 m resolution canopy height and
digital terrain models from WorldView and ArcticDEM data. Remote Sens. Environ. 2018, 218, 174–188.
26. Xie, J.; Huang, G.; Liu, R.; Zhao, C.; Dai, J.; Jin, T.; Mo, F.; Zhen, Y.; Xi, S.; Tang, H. Design and data processing of China’s first
spaceborne laser altimeter system for earth observation: GaoFen-7. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2020, 13,
1034–1044.
27. Ni, W.; Li, Z.; Wang, Q.; Zhang, Z.; Liu, Q.; Pang, Y.; He, Y.; Li, Z.; Sun, G. Forest heights extraction using GF-7 very high-resolution
stereoscopic imagery and Google Earth multi-temporal historical imagery. J. Remote Sens. 2024, 4, 0158. [CrossRef]
28. Puliti, S.; Hauglin, M.; Breidenbach, J.; Montesano, P.; Neigh, C.; Rahlf, J.; Solberg, S.; Klingenberg, T.; Astrup, R. Modelling
above-ground biomass stock over Norway using national forest inventory data with ArcticDEM and Sentinel-2 data. Remote Sens.
Environ. 2020, 236, 111501. [CrossRef]
29. Zhang, T.; Lin, H.; Long, J.; Zhang, M.; Liu, Z. Analyzing the Saturation of Growing Stem Volume Based on ZY-3 Stereo and
Multispectral Images in Planted Coniferous Forest. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2021, 15, 50–61. [CrossRef]
30. Maack, J.; Kattenborn, T.; Fassnacht, F.E.; Enßle, F.; Hernández, J.; Corvalán, P.; Koch, B. Modeling forest biomass using Very-
High-Resolution data—Combining textural, spectral and photogrammetric predictors derived from spaceborne stereo images.
Eur. J. Remote Sens. 2015, 48, 245–261. [CrossRef]
31. Zhu, X.; Tang, X.; Zhang, G.; Liu, B.; Hu, W. Accuracy comparison and assessment of DSM derived from GFDM satellite and GF-7
satellite imagery. Remote Sens. 2021, 13, 4791. [CrossRef]
32. Chen, P.; Huang, H.; Liu, J.; Wang, J.; Liu, C.; Zhang, N.; Su, M.; Zhang, D. Leveraging Chinese GaoFen-7 imagery for
high-resolution building height estimation in multiple cities. Remote Sens. Environ. 2023, 298, 113802. [CrossRef]
33. Chen, P.; Huang, H.; Ye, F.; Liu, J.; Li, W.; Wang, J.; Wang, Z.; Liu, C.; Zhang, N. A benchmark GaoFen-7 dataset for building
extraction from satellite images. Sci. Data 2024, 11, 187. [CrossRef]
34. Wu, J.; Meng, Q.; Gao, L.; Zhang, L.; Zhao, M.; Su, C. A deep learning framework for 3D vegetation extraction in complex urban
environments. Int. J. Appl. Earth Obs. Geoinf. 2024, 129, 103798. [CrossRef]
35. Du, L.; Pang, Y.; Ni, W.; Liang, X.; Li, Z.; Suarez, J.; Wei, W. Forest terrain and canopy height estimation using stereo images and
spaceborne LiDAR data from GF-7 satellite. Geo-Spat. Inf. Sci. 2024, 27, 811–821. [CrossRef]
36. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
37. Gong, P.; Mei, X.; Biging, G.S.; Zhang, Z. Improvement of an oak canopy model extracted from digital photogrammetry.
Photogramm. Eng. Remote Sens. 2002, 68, 919–924.
38. Ni, W.; Sun, G.; Ranson, K.J.; Pang, Y.; Zhang, Z.; Yao, W. Extraction of ground surface elevation from ZY-3 winter stereo imagery
over deciduous forested areas. Remote Sens. Environ. 2015, 159, 194–202. [CrossRef]
39. Travers-Smith, H.; Coops, N.C.; Mulverhill, C.; Wulder, M.A.; Ignace, D.; Lantz, T.C. Mapping vegetation height and identifying
the northern forest limit across Canada using ICESat-2, Landsat time series and topographic data. Remote Sens. Environ. 2024, 305,
114097. [CrossRef]
40. Musthafa, M.; Singh, G. Improving Forest Above-Ground Biomass Retrieval Using Multi-Sensor L- and C- Band SAR Data and
Multi-Temporal Spaceborne LiDAR Data. Front. For. Glob. Change 2022, 5, 822704. [CrossRef]
41. Cao, L.; Coops, N.C.; Sun, Y.; Ruan, H.; Wang, G.; Dai, J.; She, G. Estimating canopy structure and biomass in bamboo forests
using airborne LiDAR data. ISPRS J. Photogramm. Remote Sens. 2019, 148, 114–129. [CrossRef]
Remote Sens. 2025, 17, 47 18 of 18

42. 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. [CrossRef]
43. Shendryk, Y. Fusing GEDI with earth observation data for large area aboveground biomass mapping. Int. J. Appl. Earth Obs.
Geoinf. 2022, 115, 103108. [CrossRef]
44. Du, L.; Pang, Y.; Wang, Q.; Huang, C.; Bai, Y.; Chen, D.; Lu, W.; Kong, D. A LiDAR biomass index-based approach for tree-and
plot-level biomass mapping over forest farms using 3D point clouds. Remote Sens. Environ. 2023, 290, 113543. [CrossRef]
45. Tebaldini, S.; Ho Tong Minh, D.; Mariotti d’Alessandro, M.; Villard, L.; Le Toan, T.; Chave, J. The status of technologies to measure
forest biomass and structural properties: State of the art in SAR tomography of tropical forests. Surv. Geophys. 2019, 40, 779–801.
[CrossRef]

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.