Estimation of Subtropical Forest Aboveground Biomass Using Active and Passive Sentinel Data with Canopy Height

Abstract

Forest biomass is closely related to carbon sequestration capacity and can reflect the level of forest management. This study utilizes four machine learning algorithms, namely Multivariate Stepwise Regression (MSR), K-Nearest Neighbors (k-NN), Artificial Neural Network (ANN), and Random Forest (RF), to estimate forest aboveground biomass (AGB) in Chenzhou City, Hunan Province, China. In addition, a canopy height model, constructed from a digital surface model (DSM) derived from Sentinel-1 Interferometric Synthetic Aperture Radar (InSAR) and an ICESat-2-corrected SRTM DEM, is incorporated to quantify its impact on the accuracy of AGB estimation. The results indicate the following: (1) The incorporation of multi-source remote sensing data significantly improves the accuracy of AGB estimation, among which the RF model performs the best (R² = 0.69, RMSE = 24.26 t·ha⁻¹) compared with the single-source model. (2) The canopy height model (CHM) obtained from InSAR-LiDAR effectively alleviates the signal saturation effect of optical and SAR data in high-biomass areas (>200 t·ha⁻¹). When FCH is added to the RF model combined with multi-source remote sensing data, the R² of the AGB estimation model is improved to 0.74. (3) In 2018, AGB in Chenzhou City shows clear spatial heterogeneity, with a mean of 51.87 t·ha⁻¹. Biomass increases from the western hilly part (32.15–68.43 t·ha⁻¹) to the eastern mountainous area (89.72–256.41 t·ha⁻¹), peaking in Dongjiang Lake National Forest Park (256.41 t·ha⁻¹). This study proposes a comprehensive feature integration framework that combines red-edge spectral indices for capturing vegetation physiological status, SAR-derived texture metrics for assessing canopy structural heterogeneity, and canopy height metrics to characterize forest three-dimensional structure. This integrated approach enables the robust and accurate monitoring of carbon storage in subtropical forests.

1. Introduction

As the main component of terrestrial ecosystems, forest ecosystems are characterized by a vast spatial coverage, complex structural stratification, and rich biodiversity. They play a critical role in regulating the global carbon cycle, with the majority of their carbon stored in the form of vegetation biomass. Therefore, accurately quantifying forest biomass reserves and their spatial distribution is of great significance for assessing carbon sink potential, estimating carbon emissions or losses, and informing climate change mitigation strategies [1]. Among all biomass components of forest ecosystems, forest aboveground biomass (AGB) accounts for approximately 70% to 90% of the total. It serves not only as a core parameter for forest carbon pool monitoring, but also as a key ecological indicator linking ecosystem functions with human activities [2]. Methods for estimating forest AGB primarily include field measurements and remote-sensing-based approaches. Traditional field-based approaches face inherent limitations, including a low efficiency and infrequent data updates, which hinder the timely estimation and dynamic monitoring of AGB at regional scales. By contrast, satellite remote sensing technologies have been widely applied in large-scale AGB estimation, the time series monitoring of land use changes, and thematic mapping due to their advantages of a broad spatial coverage, high temporal resolution, and rapid data acquisition [3,4,5,6].

In recent years, the rapid advancement of multi-platform collaborative observation technologies, along with a high spatiotemporal resolution and hyperspectral remote sensing, has significantly improved the accuracy of forest AGB estimation by enabling the synergistic inversion of active and passive remote sensing data [7]. Specifically, passive remote sensing acquires rich spectral information on the vegetation canopy through multispectral or hyperspectral sensors, while active remote sensing technologies, such as Light Detection and Ranging (LiDAR), with its laser ranging capabilities, and Synthetic Aperture Radar (SAR), with its microwave penetration characteristics, can effectively capture the three-dimensional vertical structural parameters of vegetation. These two approaches are highly complementary in retrieving vegetation parameters [8]. Passive remote sensing is susceptible to cloud interference and suffers from spectral saturation in high-biomass regions. In contrast, active remote sensing, particularly SAR and LiDAR, can penetrate the canopy and extract structural metrics closely related to AGB, such as tree height and canopy vertical profile, thereby compensating for the limitations of optical data [9]. In passive remote sensing systems, medium-resolution data (e.g., the Landsat series) have become the preferred data source for regional-scale AGB prediction due to long-term data continuity and favorable cost-effectiveness. Specifically, the Sentinel-2 satellite provides important data support for regional AGB dynamic monitoring with its high sensitivity to vegetation biomass in the red-edge bands (Bands 5, 6, and 7), 12-day revisiting cycle, and free and open-access data policy. Existing studies have shown that the estimation accuracy of forest AGB can be significantly improved by extracting biophysical parameters from Sentinel-2 data, such as Leaf Area Index (LAI) and canopy water content, and by incorporating texture features [10,11]. However, in complex terrain areas, such as forests in mountainous regions, optical remote sensing still faces significant challenges, such as spectral saturation and cloud interference, which require mitigation through active–passive data incorporation. In contrast, active microwave remote sensing (such as SAR), with its all-weather conditions and day-and-night observation capabilities, can directly obtain structural parameters closely related to AGB, including the backscatter coefficient (σ⁰), polarization decomposition characteristics (e.g., HH/HV ratio), and interferometric coherence (γ), which effectively compensates for the limitations of passive optical remote sensing in detecting vertical forest structure. Among SAR systems, long-wavelength sensors (e.g., L-band ALOS PALSAR) exhibit a greater sensitivity to forest AGB due to their greater canopy penetration capacity, and have become the key data source for AGB estimation in cloud-prone tropical rainforest regions [12]. Notably, Interferometric Synthetic Aperture Radar (InSAR), by quantifying the phase stability of radar signals, exhibits a greater sensitivity to biomass than backscatter intensity, and can increase the inversion saturation point to a higher biomass level [13]. Currently, Sentinel-1 is the only freely available and open-access source of C-band SAR data. Its cross-polarization (VH/VV) data have been shown to achieve AGB prediction accuracies ranging from R² = 0.54 to 0.76 when using machine learning methods (e.g., Random Forest (RF) or Deep Neural Networks (DNNs)) [14,15]. However, since C-band electromagnetic waves are easily affected by surface scattering from the forest canopy and their inversion accuracy is significantly restricted by multiple factors, such as forest type (e.g., differences between coniferous and broad-leaved forests), terrain undulation, and surface humidity, the spatial extrapolation capability of C-band-data-based models is significantly constrained [16,17,18]. Optical data such as Sentinel-2 can provide horizontal structural parameters (e.g., LAI and red-edge position (REIP)), while SAR data, including Sentinel-1 and ALOS-2, provides vertical structural information, such as canopy height. The incorporation of active and passive data can greatly mitigate the limitations of using a single data source, such as spectral saturation in optical data and speckle noise or terrain-induced distortions in SAR imagery. Empirical studies have shown that the forest AGB estimation accuracy achieved through the incorporation of Sentinel-1 and Sentinel-2 data (R² = 0.84, RMSE = 15.3 Mg·ha⁻¹) offers a notable improvement over single-sensor estimates—especially compared to Sentinel-1 (R² = 0.34)—and modestly improves upon Sentinel-2 (R² = 0.82). In addition, the incorporated data-based model reduces the stock volume estimation error to 49.70 m³·ha⁻¹ [18,19]. This multi-source data incorporation framework presents a promising approach for enhancing the accuracy of large-scale forest AGB monitoring.

As a typical region where precipitation and temperature seasonally coincide under the regulation of the western Pacific monsoon, the central subtropical region of China hosts the world’s largest area of evergreen broad-leaved forest ecosystems. The high vegetation productivity and carbon density of evergreen broad-leaved forest ecosystems in China make this area a key region for China’s terrestrial carbon sink. For the accurate estimation of AGB in this region, special attention should be paid to the frequent cloud contamination of optical data under the monsoon climate, geometric distortion in SAR imagery caused by complex terrain (e.g., hilly landforms), and the influence of seasonal phenological variation in evergreen vegetation on spectral reflectance characteristics, underscoring the importance of applying an integrated active–passive remote sensing framework in this region. Chenzhou City, situated in the southeastern part of Hunan Province, lies within the subtropical monsoon climate zone and is rich in flora and fauna resources. The city is designated as one of the nationally key collective forestry cities [20]. Therefore, accurately estimating forest aboveground biomass in Chenzhou City provides a scientific basis for the formulation of regional carbon peaking and neutrality targets.

This study uses Chenzhou City, Hunan Province, China, as the case study area and integrates Sentinel-1A SAR data, Sentinel-2A multispectral imagery, and InSAR-derived forest canopy height (FCH) to construct a multi-source dataset for AGB estimation. Currently, there is limited research on the remote sensing estimation of AGB in subtropical forests using active and passive Sentinel data and FCH, which hampers the broader application of Sentinel data in subtropical forest carbon sink monitoring. Accordingly, this study is conducted based on the following two hypotheses: (1) the incorporation of active and passive remote sensing data from Sentinel-1 and Sentinel-2 can significantly improve the estimation accuracy of AGB and reduce saturation effects in high-biomass regions and (2) introducing the InSAR-derived FCH as an input variable of the inversion model can enhance the model’s ability to capture the vertical structure of the forest and further improve the AGB estimation accuracy. The main objective of this study is to develop and validate an effective AGB estimation framework for subtropical forests by integrating active and passive remote sensing data with FCH information. This research aims to provide a scientific basis for forest carbon sink assessments in subtropical regions using Sentinel data under complex terrain and persistent cloud cover conditions.

2. Materials and Methods

2.1. Study Area

Chenzhou City (112°13′–114°14′E, 24°53′–26°50′N) is located in southeastern Hunan Province (Figure 1), at the junction of the Nanling and Luoxiao Mountains, serving as the watershed between the Yangtze and Pearl River. The city covers an area of 1.94 × 10⁴ km², with a complex terrain dominated by mountains (42.3%) and hills (32.7%). Elevation in the city ranges from 70 m to 2061.3 m, forming a steep gradient from the southeastern highlands to the northwestern lowlands. The region has a subtropical monsoon humid climate, with an annual average temperature of 17.4 °C, a frost-free period of 286 days, and annual precipitation of 1452.1 mm (with 45% occurring in the April–June rainy season). Annual sunshine averages 1440.4 h. These favorable hydrothermal conditions support diverse forest ecosystems.

As one of Hunan Province’s four key forest regions, Chenzhou City has 1.36 million hectares of forest land, with a forest coverage rate of 62.96% and a total standing stock volume of 102 million m³. The forest vegetation exhibits a obvious vertical distribution: the evergreen broad-leaved forest is located below 650 m, the evergreen–deciduous broad-leaved mixed forest is located between 650 m and 1000 m, the deciduous broad-leaved forest is distributed between 1000 m and 1500 m, and the shrub grassland is distributed above 1500 m. The region harbors rich biodiversity (1409 tree and shrub species from 385 genera and 100 families), making it a Nanling Mountains biodiversity hotspot. As a critical water conservation area for the Xiangjiang and Pearl River headwaters and a key carbon sink, Chenzhou’s forests play a vital role in regional ecological security and carbon cycling [21,22,23].

2.2. Data Collection and Processing

2.2.1. Fixed Sample Plots Data of National Forest Continuous Inventory (NFCI)

This study used fixed plot data from the ninth NFCI in Hunan Province (2019) as ground verification data. Although the remote sensing images were acquired in the second half of 2018, approximately half a year prior to the field plots survey, considering that the study area is located in the mid-subtropical region, where tree growth nearly ceases during the winter months (November–February), leading to minimal biomass accumulation, the biomass data from 2018 and 2019 are considered comparable, with the temporal discrepancy exerting negligible influence on overall biomass estimation due to seasonal dormancy [24,25]. To further verify temporal consistency, this study compared the Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI) from the same period in 2018 and 2019. The results showed that the mean differences in the NDVI and EVI between the two years were minimal, with a mean NDVI difference of 0.03 and a mean EVI difference of 0.02. The coefficients of determination (R²) for the NDVI and EVI between the two years both exceeded 0.95, indicating a strong consistency in vegetation conditions. These findings support the feasibility of using data from different years in the model building. The study used data from 592 fixed plots in Chenzhou City, arranged in a systematic grid with a spacing of 6 km × 8 km. Each plot covered an area of 0.067 hectares (25.82 m × 25.82 m). Each plot recorded more than 60 variables, including land use type and forest stand characteristics, with detailed measurements of key structural parameters such as diameter at breast height (DBH), tree height, and canopy density. Forest AGB was estimated using an allometric biomass model developed by Zeng Weisheng et al. (2011) [26], which was calibrated for regional forest species. The AGB for each plot was obtained by applying the model to field-measured tree data (DBH ≥ 5 cm) and summing the estimated biomass of all individual trees [26]. The allometric model used to estimate individual tree biomass is expressed as follows:

$$M_{a1}=\alpha \times D^{{\displaystyle \frac{7}{3}}}$$

(1)

$$M_{a2}=N_{bamboo}\times {CF}_{bamboo}$$

(2)

$$M_{a3}=A_{shrub}\times {CF}_{shrub}$$

(3)

where $M_{a1}$ is the AGB of arbor trees (kg), $M_{a2}$ is the AGB of bamboo forest (kg), $M_{a3}$ is the AGB of shrubs (kg), $D$ is the diameter at DBH of trees (cm), $\alpha$ is a species-specific parameter defined as $\alpha$ = 0.3$\rho$, and $\rho$ is the basic wood density (g/cm³). $N_{Bamboo}$ is the number of bamboo plants. In Hunan Province, Moso bamboo (Phyllostachys edulis) accounts for over 97% of the total bamboo area. Other bamboo species are classified as shrubs in this study. ${CF}_{bamboo}$ is the average AGB per bamboo plant, taken as 22.5 kg/n (n represents the number of plants) [27], while $A_{shrub}$ is the area of shrubland and ${CF}_{shrub}$ is the average AGB per unit area, set at 19.76 t·ha⁻¹. The total biomass per plot was converted to AGB per hectare (t·ha⁻²) based on the plot area. Summary statistics of the forest AGB by major forest types are presented in

Table 1. AGB statistics by major forest types in fixed sample plots in Chenzhou City.

Vegetation Type	Plot Number	AGB/(t/ha−1)
		Minimum	Maximum	Mean	Standard Deviation
Coniferous forest	123	1.43	158.28	54.64	35.29
Broadleaf forest	46	1.82	271.19	73.19	56.27
Coniferous–mixed forest	14	25.83	128.21	64.33	29.41
Broadleaf–mixed forest	36	9.85	208.53	79.28	50.37
Coniferous–broadleaf mixed forest	35	4.89	198.13	60.21	43.52
Bamboo forest	26	1.79	66.53	40.14	17.96
Shrubland	62	0.58	59.83	26.16	28.69
Economic forest	30	1.26	82.26	24.64	18.64
Others	220

.

2.2.2. Remote Sensing Data

The multi-source remote sensing datasets used in this study include: Sentinel-1 C-band SAR data (GRD products), which provides dual-polarization (VV and VH) backscatter coefficients for forest structure and moisture analysis; Sentinel-2 Level-2A multispectral imagery, which has 13 spectral bands with spatial resolutions ranging from 10 to 60 m, used primarily for vegetation and spectral feature extraction; Global Canopy Height (GCH) data used for vertical structure characterization; digital elevation data derived from the Shuttle Radar Topography Mission (SRTM v3, 30-m resolution), utilized for terrain feature extraction such as slope, aspect, and elevation gradients; and ICESat-2 ATL08 photon-counting LiDAR data, which provides footprint-level canopy height estimates along satellite ground tracks, used to calibrate and validate aboveground biomass estimates. The acquisition dates and spatial resolutions of datasets mentioned above are summarized in

Table 2. A brief description of remote sensing datasets used in the study.

Dataset	Description	Source	Spatial Resolution	Date
Sentinel-1	Sentinel-1 IW mode Level-1 GRD product	European Space Agency	10 m × 10 m	4 October 2018 9 October 2018
	Sentinel-1 IW Level-1 SLC product		2.3 m × 14.1 m	29 August 2018, 3 September 2018, 10 September 2018, 15 September 2018
Sentinel-2	MAI Level-1C product	European Space Agency	10 m × 10 m	7 October 2018
SRTM	SRTM Version 4.1 DEM	United States Geological Survey	30 m × 30 m	11 February 2000– 22 February 2000
ICESat-2	ICESat-2 ATL08 Version 5 Land and Vegetation Height Product	National Snow and Ice Data Center	17 m (Footprint size)	1 January 2019– 13 December 2019
GCH	NASIA and SASIA regional data coverage	Global Land Analysis and Discovery	30 m × 30 m	2019

.

(1) Sentinel-1

① Ground Range Detected

This study utilizes the Sentinel Application Platform (SNAP), developed by the European Space Agency (ESA, Paris, France), to preprocess Sentinel-1 Level-1 GRD data in accordance with the ES-recommended workflow [28]. Using the 30 m resolution SRTM digital elevation model (DEM) as auxiliary data, orbital correction is first performed on three GRD scenes by applying precise orbit ephemerides to eliminate satellite positioning errors. Next, thermal noise correction is applied to reduce the internal electronic noise inherent to the SAR system. Digital number values are then radiometrically calibrated to obtain the backscatter coefficient (σ⁰), expressed in linear units or decibels (dB), depending on application needs [29]. Speckle noise is suppressed using a Refined Lee filter with a 7 × 7 moving window. Thirdly, Range-Doppler terrain correction is applied using the Shuttle Radar Topography Mission C-band DEM (SRTM DEM) to correct geometric distortions due to terrain variations, and the images are geocoded to a UTM projection. The calibrated backscatter coefficients are then converted to dB scale, and multi-temporal mosaicking is performed to generate seamless SAR composites. The mosaicked images are resampled to a spatial resolution of 25.82 m to match the field sample plot dimensions, generating backscatter coefficient products covering the entire study area [30,31,32].

② Single Look Complex

This study utilizes the SNAP to preprocess six Sentinel-1 Level-1 SLC scenes. First, precise orbit ephemerides (Apply Orbit File) are used to correct satellite orbit errors and improve geolocation accuracy. Subsequently, radiometric calibration is applied to convert the complex SAR data into absolute backscatter coefficients (σ⁰), enabling quantitative analysis of radar returns. To address the burst discontinuities inherent in Sentinel-1 TOPS-mode SLC data, a de-bursting step is applied to eliminate radiometric discontinuities at burst boundaries. To improve the signal-to-noise ratio and reduce speckle, multi-looking is performed by spatially averaging pixels across range and azimuth directions. This is followed by Range-Doppler terrain correction, which compensates for terrain-induced geometric distortions and enables precise geocoding of the imagery. The resulting products are standardized and georeferenced to make them suitable for further spatial analysis. Key preprocessing steps—including orbit correction, radiometric calibration, and terrain correction—are consistent with those used for GRD products, ensuring methodological standardization and result comparability.

(2) Sentinel-2

This study adopted an integrated preprocessing framework to systematically process multi-source remote sensing data, including Sentinel-2 and ICESat-2. For Sentinel-2A MSI data, Level-1C top-of-atmosphere (TOA) reflectance images are first processed using the Sen2Cor algorithm to perform atmospheric correction, producing Level-2A surface reflectance products derived via a radioactive transfer model. The surface reflectance data are subsequently mosaicked, clipped to the study area extent, and resampled to generate atmospherically corrected surface reflectance data with a spatial resolution of 25.82 m, matching the size of the fixed sample plots.

(3) ICESat-2

This study processes ICESat-2 ATL08 land elevation data collected throughout 2019, focusing on quality enhancement and spatial regularization. Given the high sensitivity of ICESat-2’s photon-counting LiDAR system to weak signal returns, a multi-stage quality control framework is implemented. First, elevation points from strong beam tracks are extracted using the ATL08 photon classification algorithm, which isolates ground photons. Next, a multi-parameter filtering approach is employed—incorporating 30 m SRTM DEM data—to detect and remove outliers based on the 3σ criterion. Finally, given the ~90 cm along-track footprint spacing, spatial thinning is applied to achieve a uniform distribution of elevation control points across the study area. To preserve representativeness, a stratified sampling strategy is used, comprising 8.28% of the original data. The spatial distribution of screened ICESat-2 elevation points is illustrated in Figure 2.

2.3. Extracting Feature Variables

Considering the complex interactions between forest AGB and other factors such as vegetation growth, atmospheric conditions, soil properties, and terrain characteristics, this study extracted predictor variables from multi-source remote sensing data for AGB estimation [33,34]. From Sentinel-2 optical data, 17 vegetation indices were calculated, including the atmospherically resistant vegetation index (ARVI), ratio vegetation index (RVI), NDVI, soil-adjusted vegetation index (SAVI), and chlorophyll index (PSSRa). In addition, reflectance values from 12 original spectral bands were extracted. To capture spatial heterogeneity in canopy structure, eight texture features—contrast, dissimilarity, mean, homogeneity, angular second moment, entropy, variance, and correlation—were derived using gray-level co-occurrence matrices (GLCMs) across three window sizes (3 × 3, 5 × 5, and 7 × 7), generating 288 texture variables. Furthermore, seven terrain variables—such as elevation (H), slope (β), and aspect—were extracted from the SRTM DEM. Altogether, a remote sensing feature set consisting of 330 predictor variables was constructed to support the estimation of forest AGB.

For active remote sensing variables, Sentinel-1 SAR data were used to extract calibrated backscatter coefficients (σ⁰) under VV and VH polarizations, along with derived polarization ratio (VV/VH) and difference (VV–VH). A total of 48 texture features were computed using gray-level co-occurrence matrices (GLCMs) based on backscatter intensity. In addition, three polarimetric decomposition features and two interferometric coherence metrics were incorporated. Combined with the 7 terrain-derived variables, a total of 64 active remote sensing features were extracted. Summary statistics for all extracted features are provided in

Table 3. List of extracted active and passive remote sensing feature factors.

Feature Category		Remote Sensing Factor
Active remote sensing feature	Normalized backscatter coefficients and derived polarization metrics	VV, VH, (VH − VV)/(VH + VV), VV/VH
	Texture features (3 × 3, 5 × 5, 7 × 7 windows)	VV/VH_CON, VV/VH_DIS, VV/VH_MEA, VV/VH_HOM, VV/VH_ASM, VV/VH_ENT, VV/VH_VAR, VV/VH_COR
	Polarimetric decomposition features (Cloude-Pottier)	Alpha, Anisotropy, Entropy
	InSAR coherence	VV/VH, Coherence coefficient
Passive remote sensing factor	Spectral reflectance	B1, B2, B3, B4, B5, B6, B7, B8, B8a, B9, B11, B12
	Information enhancement features	ALBEDO, VIS234, RED5678A, PCA_B1, PCA_B2, PCA_B3
	Texture features (3 × 3, 5 × 5, 7 × 7 windows)	B1-12_CON, B1-12_DIS, B1-12_MEA, B1-12_HOM, B1-12_ASM, B1-12_ENT, B1-12_VAR, B1-12_COR
	Vegetation indices	RVI, DVI, WDVI, IPVI, NDVI, NDI45, GNDVI, SAVI, TSAVI, MSAVI, ARVI, PSSRa, MTCI, MCARI, S2REP, REIP, GEMI
Topographic factor		H, β, α, sinα, cosα, Cv, Ch

Notes: B1-12__CON: Contrast (GLCM-based texture feature, measures local intensity variation), B1-12__DIS: Dissimilarity (GLCM-based, reflects difference between neighboring pixel values), B1-12__MEA: Mean (GLCM-based average reflectance value in the window), B1-12__HOM: Homogeneity (GLCM-based, evaluates closeness of element distribution to the diagonal), B1-12__ASM: Angular Second Moment (GLCM-based, reflects textural uniformity), B1-12__ENT: Entropy (GLCM-based, measures randomness in texture patterns), B1-12__VAR: Variance (GLCM-based, measures pixel value dispersion), B1-12__COR: Correlation (GLCM-based, describes linear dependency of gray levels), Alpha (α): Mean scattering angle, indicates dominant scattering mechanism, Anisotropy: The directional variation of scattering/reflection, influenced by primary and secondary mechanisms, Entropy: Polarimetric entropy, reflects the number of effective scattering mechanisms, ALBEDO: Multi-band albedo integrated from Sentinel-2 B1-B12, VIS234: Composite reflectance from visible bands (Bands 2, 3, 4), PCA_B: Principal Component Analysis (PCA) bands derived from original reflectance bands, RVI: Ratio Vegetation Index, DVI: Difference Vegetation Index, WDVI: Weighted Difference Vegetation Index, IPVI: Infrared Percentage Vegetation Index, NDVI: Normalized Difference Vegetation Index, NDI45: Optimized Normalized Difference Index (Bands 4 and 5), GNDVI: Green Normalized Difference Vegetation Index, SAVI: Soil Adjusted Vegetation Index, TSAVI: Transformed Soil Adjusted Vegetation Index, MSAVI: Modified Soil Adjusted Vegetation Index, ARVI: Atmospherically Resistant Vegetation Index, PSSRa: Pigment Specific Simple Ratio (chlorophyll-related index), MTCI: MERIS Terrestrial Chlorophyll Index, MCARI: Modified Chlorophyll Absorption in Reflectance Index, S2REP: Sentinel-2 Red Edge Position Index, REIP: Red Edge Inflection Point, GEMI: Global Environmental Monitoring Index, H: Elevation (meters above sea level), β (beta): Slope (degrees), α (alpha): Aspect (degrees), Cv: Profile curvature, and Ch: Plan curvature.

. This multi-source feature set integrates spectral, structural, and topographic information, forming a robust input for the subsequent aboveground biomass estimation model.

2.4. Generating FCH

This study employs a digital elevation differencing approach (DSM–DEM subtraction) to extract FCH. The method derives vertical vegetation structure by subtracting DEM from the digital surface model (DSM), which includes vegetation and building height [35]. Given that the C-band radar of Sentinel-1 has limited penetration into dense forest canopies (typically <5 m), its interferometric elevation measurements predominantly characterize the upper canopy surface. Therefore, the InSAR-derived DSM can be reasonably interpreted as a canopy surface model for forested areas. To mitigate elevation bias caused by vegetation-induced scattering in the SRTM C-band DEM, this study utilizes ICESat-2 ATL03 photon data and ATL08 terrain elevation products to extract accurate ground elevation references. High-precision ground control points (GCPs) are identified through photon classification and the filtering of sparsely distributed, high-confidence ground returns. A linear regression model is constructed between ICESat-2-derived ground elevations and the corresponding SRTM DEM, and this model is applied to correct vegetation-related elevation errors, producing an optimized DEM. Subsequently, an FCH distribution map of the study area is generated by subtracting the corrected DEM from the InSAR-derived DSM. This approach reduces the typical overestimation of ground elevation in forested regions observed in traditional SRTM DEM products and provides reliable vertical structure parameters for subsequent biomass estimation.

2.5. Screening Remote Sensing Variables

In this study, a multi-stage feature selection strategy was implemented to optimize the remote-sensing-based modeling of forest AGB [36,37]. The main objective of feature selection was to identify the most informative subset from the original set of 394 candidate variables, while mitigating overfitting risks caused by multicollinearity and feature redundancy. The process began with univariate Pearson correlation analysis, where variables not significantly correlated with AGB (p > 0.05) were excluded. Subsequently, model-specific variable selection strategies were applied. For the multivariate stepwise regression (MSR) model, a stepwise selection procedure based on F-tests was applied, with an entry threshold of p = 0.05 and a removal threshold of p = 0.1. For machine learning models, variables were selected based on their importance scores. Specifically, the following two indicators were used: the percentage increase in mean squared error after permutation (%IncMSE), which reflects the impact of each variable on model prediction accuracy, and the total decrease in node impurity (IncNodePurity), which represents the contribution of the variable to the structural optimization of the model. In both cases, larger values indicate higher variable importance. Following this two-stage screening process, 12 core predictor variables were selected for model construction. Among these, texture features–particularly the mean value derived from 7 × 7 window GLCM statistics—exhibited a consistently high importance across both statistical and machine learning methods, highlighting the critical role of spatial structural characteristics in AGB estimation.

2.6. AGB Estimation Model

2.6.1. MSR

This study applied stepwise MSR to analyze the relationship between the AGB of sample plots and multi-source remote sensing variables. The stepwise procedure iteratively selected variables to minimize multicollinearity while optimizing model fit. Predictors were comprised of both active (Sentinel-1 backscatter coefficients and polarimetric features) and passive (Sentinel-2 vegetation indices and texture metrics) remote sensing variables. The stepwise regression based on F-test criteria in SPSS 26.0 was used, with variable entry and removal p-value thresholds set at 0.05 and 0.10, respectively. After variable elimination, all retained variables had a variance inflation factor (VIF) below 10 to minimize the influence of multicollinearity, thereby supporting reliable model estimation and interpretation.

2.6.2. ANN

An ANN based on the backpropagation (BP) algorithm was constructed using a three-layer feedforward architecture, consisting of an input layer, a hidden layer, and an output layer. The input layer contained 12 neurons, corresponding to the selected remote sensing predictor variables. The number of neurons in the hidden layer was set between 11 and 21 and optimized using a stepwise grid search approach with a step size of 1. The output layer comprised a single neuron representing the estimated AGB [38]. The neural network was constructed and trained using the neuralnet package in the R programming environment. To enhance model robustness and reduce variability caused by random weight initialization, the network was independently trained 100 times. Each training iteration employed ten-fold cross-validation, during which the average relative error between the training and test sets was calculated to evaluate generalization performance. The final model configuration was selected based on the minimum values of the composite error criteria—RMSE and MAE—thereby ensuring both a high predictive accuracy and strong generalization capability, while effectively avoiding overfitting and underfitting.

2.6.3. k-NN

The k-Nearest Neighbors (k-NN) algorithm predicts target values by aggregating responses from the k most similar training samples within the feature space. The algorithm was implemented with the k-NN class package in R, with Euclidean distance as the similarity metric. A ten-fold cross-validation approach was adopted according to Chen’s findings, with k in the k-NN model set to 10 in this study. The model performance was evaluated using ten-fold cross-validation [39].

2.6.4. RF

RF uses bootstrap resampling to generate multiple training subsets from an original dataset. A decision tree is built for each bootstrap sample, and the final prediction is obtained by aggregating the outputs of all trees via majority voting for classification or averaging for regression. Extensive theoretical and empirical studies have demonstrated that RF offers a high prediction accuracy for AGB estimation, a strong robustness to outliers and noise, and a low risk of overfitting [39]. In this study, the RF model was implemented using the randomForest package in R. The number of trees (ntree) was fixed at 500—a commonly used default value that balances model performance and computational efficiency. The mtry parameter was initially set to one-third of the total number of predictors (i.e., 4 for 12 variables). Additional mtry values corresponding to 0.5× and 1.5× of the default (i.e., 2 and 6) were tested for comparison [40]. The final model was selected based on the lowest prediction error observed during ten-fold cross-validation.

2.7. Model Accuracy Evaluation

Ten-fold cross-validation was implemented to evaluate the predictive performance of the models. The dataset was randomly stratified into 10 subsets of equal size, with each subset serving sequentially as the validation set while the remaining nine subsets served as the training set. This process was repeated across all folds, and the mean values of the evaluation metrics were used as the final performance indicators. To comprehensively evaluate model performance, the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE) were adopted as evaluation metrics. R² quantifies the proportion of variance in the dependent variable explained by the model. It ranges from 0 to 1, with values closer to 1 indicating a better fit. RMSE measures overall prediction accuracy as the square root of the average squared differences between predicted and observed values. Smaller RMSE values indicate a higher prediction accuracy. MAE represents the average magnitude of prediction errors, providing an intuitive measure of absolute error regardless of direction. An MAE of 0 indicates perfect prediction, and larger values correspond to greater average errors [41]. The formulas for these three metrics are as follows:

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

(4)

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

(5)

$$MAE={\displaystyle \frac{\sum _{i=1}^n|{\widehat{y}}_i - {\mathrm{y}}_{\mathrm{i}}|}{n}}$$

(6)

where $n$ is the number of samples, $y_i$ is the actual observation value, $\overline{y}$ is the mean of the actual observation value, and ${\widehat{y}}_i$ is the model prediction value.

3. Results and Analysis

3.1. Screening of Key Feature Variables

This study extracted optical characteristics from Sentinel-2 multispectral imagery, including surface reflectance across spectral bands, vegetation indices, and GLCM-based texture features, while deriving SAR features from Sentinel-1 data, encompassing dual-polarization backscatter coefficients, Cloude–Pottier polarimetric decomposition parameters, and interferometric coherence metrics, together with topographic parameters generated from auxiliary DEM data, ultimately compiling an initial feature set of 387 variables. To mitigate multicollinearity and feature redundancy that may adversely affect model performance, this study implemented a dual feature selection approach combining MSR and RF. MSR iteratively eliminated statistically insignificant variables using the F-test (p < 0.05), while RF ranked variable importance based on its impact on prediction accuracy. The final set of selected key variables is presented in

Table 4. Key feature variables of ABG estimation model from different data sources.

Data Source	Model	Characteristic Variable
Sentinel-1	MSR	H, β, VH, VH_3_CON, VH_5_DIS, VH_5_HOM\VH_7_ASM, VH_7_CON, VH_7_HOM, VV_7_HOM, VV_7_ENT, Cv
	ANN, k-NN, RF	H, β, VH_5_VAR, VH_5_HOM, VH_3_VAR, CCVH, VV, VH_3_CON, sinα, VV_5_VAR, VV_7_ASM, VH_7_HOM
Sentinel-2	MSR	B1, B1_7_MEA, B2_5_MEA, B2_7_MEA, B2, B3_3_MEA, B3_7_MEA, B4_5_MEA, B5_7_MEA, RVI, PSSRa
	ANN, k-NN, RF	B2_5_MEA, B3_7_MEA, B4, B2_7_MEA, B4_5_MEA, GNDVI, B4_7_MEA, B2, VIS234, B12_7_DIS, PSSRA, IPVI
Sentinel-1 & 2	MSR	B1, B1_7_MEA, B2_5_MEA, B2_7_MEA, B3_3_MEA, B3_5_MEA, B4_5_MEA, B5_7_MEA, RVI, PSSRA
	ANN, k-NN, RF	B3_5_MEA, VIS234, Cv, B3_7_MEA, VH_3_HOM, TSAVI, VH, B4_7_MEA, B4, VV_5_ASM, PSSRA, B2_7_MEA

Notes: CC_VH denotes the interferometric coherence coefficient derived from VH polarization. Texture features are labeled in the format “Band_WindowSize_TextureMetric”, where “Band” indicates the spectral band, “WindowSize” refers to the moving window dimensions (e.g., 3 × 3, 5 × 5), and “TextureMetric” represents the corresponding statistical descriptor (e.g., contrast, entropy, homogeneity).

.

(1) Key active remote sensing features

The feature selection results indicate that topographic parameters—specifically slope, aspect, and elevation—show consistently high importance scores in both selection methods. First, topographic features are closely related to the spatial distribution pattern of forest AGB by affecting ecological processes such as light distribution, water conditions, and nutrient spatial heterogeneity. In addition, terrain variation alters the propagation path and scattering behavior of microwave signals. Due to its side-looking geometry, SAR data is particularly sensitive to topographic conditions. As shown in

Table 5. Accuracy of AGB estimation models by different remote sensing data.

Model	Evaluating Indicator	Sentinel-1	Sentinel-2	Sentinel-1 & 2
MSR	R2	0.39	0.44	0.46
	RMSE/(t·ha−1)	24.29	23.75	24.33
	MAE/(t·ha−1)	35.02	34.51	34.92
k-NN	R2	0.51	0.55	0.57
	RMSE/(t·ha−1)	28.76	29.78	29.40
	MAE/(t·ha−1)	48.91	47.65	48.54
ANN	R2	0.51	0.57	0.62
	RMSE/(t·ha−1)	24.67	22.68	27.42
	MAE/(t·ha−1)	34.52	29.44	35.78
RF	R2	0.63	0.65	0.69
	RMSE/(t·ha−1)	24.67	25.25	24.26
	MAE/(t·ha−1)	36.52	36.23	36.08

, VH polarization and its associated texture features (e.g., VH_5_HOM and VH_7_CON) are selected more frequently than VV-polarization-derived features. Additionally, the coherence coefficient for VH polarization (CC_VH) is included. This difference can be attributed to the design of the Sentinel-1 sensor. Because cross-polarization (VH) typically exhibits a lower signal intensity, the Sentinel-1 sensor employs a higher receiver bandwidth to enhance signal reception. Consequently, VH polarization exhibits a greater sensitivity to canopy surface roughness and vegetation volume scattering, enhancing its utility for forest AGB estimation. Therefore, when using Sentinel-1 data alone for forest biomass estimation, VH polarization should be prioritized due to its higher sensitivity to vegetation structural characteristics.

(2) Key passive remote sensing features

The visible spectrum bands (B2, B3, and B4), along with their composite (VIS234), are consistently selected as important predictor variables across all models. This aligns well with established spectral response patterns of vegetation: healthy vegetation exhibits peak reflectance in the green band (B3, 560 nm) and strong chlorophyll absorption in the blue (B2, 490 nm) and red (B4, 665 nm) bands. While red-edge bands (B5–B7) are strongly correlated with chlorophyll content, their contribution to AGB estimation is limited in the subtropical mixed forests. This may be attributed to phenological variability among forest types in the study area and the red-edge spectrum’s sensitivity to soil moisture or canopy water content. Among the vegetation indices, the TSAVI improves vegetation signal extraction by incorporating soil adjustment parameters to minimize background interference in sparsely vegetated areas. The PSSRa, calculated as the ratio of Band 7 to Band 4 (B7/B4), leverages red-edge sensitivity to chlorophyll content and serves as an effective indicator of vegetation physiological condition. The RVI emphasizes vegetation cover by exploiting the contrast between near-infrared (NIR) and red reflectance. The IPVI highlights NIR dominance in healthy vegetation, improving sensitivity to biomass and canopy structure. Together, these indices characterize canopy structure, physiological condition, and coverage from complementary spectral perspectives, providing multi-dimensional inputs for AGB estimation.

(3) Incorporated key active and passive features

Texture features demonstrate a substantial importance in the integrated model, representing 64.7% of all selected variables. Among these, features based on the mean statistic (MEA) account for 77.8% of the texture predictors. Analysis of window size reveals that larger texture windows (5 × 5 and 7 ×7) yield a stronger predictive performance compared to 3 × 3 windows, with an average increase in R² of 18.6%. This improvement is primarily attributed to the enhanced capacity of larger windows to suppress random noise and better capture spatial heterogeneity. Notably, Band 1 (443 nm), which lies in the blue range of the spectrum, exhibits a significant correlation with AGB in the multivariate linear regression model (r = 0.32, p < 0.05). This may be related to its sensitivity to atmospheric scattering and canopy structure under complex topographic conditions. Therefore, it is advisable to include Band 1 reflectance in Sentinel-2-based biomass estimation in mountainous or topographically complex forested regions, where it may help to mitigate terrain-induced radiometric distortions.

3.2. Model Accuracy by Different Data Sources

The accuracy of forest AGB estimation models varied by Sentinel-1 data, Sentinel-2 data, and their combined. The RF model integrating active and passive remote sensing features (Sentinel-1 and Sentinel-2) achieved the highest predictive performance, with an R² of 0.69. Detailed performance metrics for each model are provided in Table 5.

The RF model integrating Sentinel-1 and Sentinel-2 features yielded the highest accuracy in AGB estimation (R² = 0.69, RMSE = 24.26 t·ha⁻¹, MAE = 36.08 t·ha⁻¹), followed by the Sentinel-2 feature based model (R² = 0.65, RMSE = 25.25 t·ha⁻¹, MAE = 36.23 t·ha⁻¹), while the Sentinel-1 feature based model showed the lowest accuracy (R² = 0.63, RMSE = 24.67 t·ha⁻¹, MAE = 36.52 t·ha⁻¹). Regarding the single-source models, the Sentinel-2-feature-based model yielded a higher accuracy in forest AGB estimation compared to Sentinel-1 SAR. This difference may be attributed to several reasons, as follows: (1) Feature dimensionality: Sentinel-2 offers a richer feature space, with 323 variables extracted from 12 spectral bands, compared to only 57 features derived from VV and VH polarizations in Sentinel-1 data. This highlights the potential for further improvement in active remote-sensing-based feature extraction for AGB estimation. (2) Data quality limitations: C-band SAR signals are susceptible to attenuation under subtropical monsoon conditions, particularly due to high atmospheric water vapor content. Additionally, geometric distortions from complex terrain degrade SAR data quality, negatively impacting AGB estimation accuracy. (3) Red-edge sensitivity: Sentinel-2’s four red-edge bands and their derived vegetation indices offer an enhanced sensitivity to vegetation physiological parameters, providing critical inputs for AGB models. (4) Spatial resolution benefit: The 10 m resolution of Sentinel-2 reduces mixed-pixel effects, thereby improving the spatial precision and accuracy of biomass estimation.

In terms of algorithm performance using Sentinel-1 and -2, machine learning models consistently outperformed traditional linear regression approaches in AGB estimation. The RF algorithm achieved the highest coefficient of determination (R² = 0.69), outperforming the ANN (R² = 0.62), k-NN (R² = 0.57), and MSR (R² = 0.48) models by 11.3%, 21.1%, and 43.8%, respectively. In addition, the RF model exhibited the lowest RMSE (24.26 t·ha⁻¹) and MAE (36.08 t·ha⁻¹), indicating a superior predictive reliability. This variation in performance is primarily attributed to differences in the model’s algorithmic principles. The RF algorithm enhances generalization by integrating multiple decision trees and introducing randomness in feature selection during training. This ensemble approach also reduces the risk of overfitting. In contrast, while the ANN possess strong nonlinear modeling capabilities, its complex parameter space and sensitivity to initialization can result in convergence to local AGB minimum during training. Traditional linear regression models, constrained by their assumption of linear relationships between predictors and response variables, are unable to capture the complex nonlinear interactions among feature variables, leading to a reduced estimation accuracy.

Using the RF model, AGB derived from different data sources was mapped and classified using the mean ± standard deviation classification method (Figure 3). For Sentinel-1-feature-based estimation, the mean AGB was 47.65 t·ha⁻¹, with a maximum of 281.38 t·ha⁻¹, a minimum of 1.91 t·ha⁻¹, and a standard deviation of 15.62 t·ha⁻¹. For Sentinel-2-feature-based estimates, the mean was 48.78 t·ha⁻¹, with a maximum of 238.08 t·ha⁻¹, a minimum of 3.51 t·ha⁻¹, and a standard deviation of 17.54 t·ha⁻¹. For the incorporated Sentinel-1- and -2-feature-based estimation, the mean AGB was 50.03 t·ha⁻¹, with a maximum of 253.00 t·ha⁻¹, a minimum of 3.36 t·ha⁻¹, and a standard deviation of 18.32 t·ha⁻¹. Regarding spatial distribution, Sentinel-2-feature-based estimation exhibited a broader extent of high-biomass regions (>70 t·ha⁻¹), whereas Sentinel-1-feature-based estimation demonstrated finer spatial details in areas with a very high biomass (>200 t·ha⁻¹). This difference arises from essential distinctions in sensor mechanisms: the penetration capability of C-band SAR allows it to detect structural biomass components such as stems and branches, whereas optical sensors are more responsive to foliar and canopy biomass. These findings confirm that incorporating active and passive remote sensing data leverages the complementary strengths of each sensor type, thereby improving the accuracy of forest AGB estimation.

3.3. Effect of FCH on AGB Estimation

This study utilizes ICESat-2 photon-counting LiDAR data to correct the SRTM DEM, aiming to improve the accuracy of FCH estimation. Because the SRTM DEM is derived from C- and X-band InSAR data, its elevation values include canopy height bias, particularly in forested areas. Directly using it to calculate canopy height by subtraction will lead to significant overestimation. To address this issue, a linear regression model (R² = 0.99) is established between the ground elevation points extracted from ICESat-2 and the corresponding SRTM DEM values.

$$y=1.006x-21.295$$

(7)

where $y$ is the corrected elevation and $x$ is the original elevation of SRTM.

The corrected DEM exhibits a higher accuracy and provides a reliable terrain reference for FCH estimation. Subsequently, the DSM derived from Sentinel-1 interferometry is subtracted by the corrected DEM to extract FCH in the study area (Figure 4). This method effectively reduces the vegetation-induced elevation bias present in the original SRTM DEM and improves the accuracy of FCH estimation.

As shown in Figure 4, FCH in the study area exhibits an obvious gradient change from west to east, with lower values in the western region and higher values in the east. This pattern can be attributed to differences in topography and land use types. The northwestern region is relatively flat and urbanized, with concentrated settlements and limited forest cover. In contrast, the central and southern parts are largely mountainous and host key ecological areas, including national nature reserves, forest parks, and state-owned forest farms. Correlation analysis reveals a statistically significant relationship between the estimated canopy height and ground-measured average stand height from fixed sample plots, with a correlation coefficient of 0.64 (p < 0.01). This correlation is substantially higher than that of the GLAD global canopy height product, which exhibits a weaker relationship with the same reference data (r = 0.39). A significant correlation is also observed between the extracted canopy height and plot-level AGB (r = 0.42, p < 0.01). These results demonstrate that canopy height derived from the DEM differencing method provides valuable input for AGB modeling and can be effectively integrated into subsequent predictive frameworks.

FCH is a key parameter reflecting vertical forest structure. Incorporating canopy height into AGB estimation models may reduce model uncertainty and enhance biomass prediction accuracy [42]. In this study, FCH is introduced as an additional predictor variable in the RF model. As shown in

Table 6. Comparison of RF model performances before and after adding FCH.

Data Source	Without FCH			With FCH
	R2	RMSE/(t·ha−1)	MAE/(t·ha−1)	R2	RMSE/(t·ha−1)	MAE/(t·ha−1)
Sentinel-1	0.63	24.67	36.52	0.65	5.16	36.44
Sentinel-2	0.65	25.25	36.23	0.69	4.95	36.34
Sentinel-1 and -2	0.69	24.26	36.08	0.74	24.37	36.01

, the inclusion of canopy height features significantly improves AGB estimation accuracy in the three data source modes of Sentinel-1, Sentinel-2, and Sentinel-1 and -2.

As shown in Table 6, the inclusion of canopy height as a predictor variable improved AGB estimation performance across all remote sensing data models. Specifically, the R² of the Sentinel-1 model increased from 0.63 to 0.65, that of the Sentinel-2 model increased from 0.68 to 0.69, and the R² of the combined Sentinel-1 and -2 model showed the largest improvement, increasing from 0.69 to 0.74. These results demonstrate that FCH, as a critical structural parameter, substantially contributes to the prediction accuracy of remote-sensing-based AGB models. This improvement is attributed to the strong correlation between canopy height and forest biophysical attributes such as stand age, crown stratification, and biomass distribution. By introducing canopy height, the prediction model could compensate for the limited vertical structural information of traditional optical and radar data, thereby enhancing its ability to capture spatial variation in forest AGB.

Based on the analysis mentioned above, this study selected the optimal set of predictor variables extracted from Sentinel-1 and -2 along with canopy height to build a random forest model to estimate forest AGB in the study area. The spatial distribution of forest AGB is presented in Figure 5.

As shown in Figure 5, forest AGB in Chenzhou City exhibited significant spatial variation in 2019. The total forest AGB was estimated at 69.44 × 10⁶ t, with an average of 51.87 t·ha⁻¹. The maximum AGB value (258.41 t·ha⁻¹) was located south of the Dongjiang Lake in central Zixing City. Other places with higher AGB values were mainly distributed in mountainous areas with higher altitudes, especially in nature reserves, national forest parks, and state-owned forest farms. The minimum AGB value (6.38 t·ha⁻¹) was primarily observed in highly urbanized areas and in places near large water bodies. The overall spatial distribution of forest biomass showed a “low in the west and high in the east” pattern, which aligns with the regional topography, socio-economic development level, forest coverage, and population distribution. The western region is characterized by lower elevations, relatively flat terrain, extensive urban development, well-developed transportation infrastructure, low forest coverage, and a high population density. In contrast, the eastern region mainly comprises mountains with higher elevations, steeper slopes, and a high concentration of nature reserves and forest parks, resulting in less anthropogenic disturbance from urban expansion or industrial activities.

4. Discussion

4.1. The Role of Feature Variables in Biomass Estimation

From Sentinel-2 data, this study derived 323 feature variables, including vegetation indices, principal components, and texture metrics. For Sentinel-1, 57 feature variables were generated through texture analysis, polarization decomposition, coherence calculation, and backscatter coefficient extraction. Among the passive remote sensing variables, visible bands and their associated texture features emerged as the most influential predictors in the biomass estimation models. Notably, the blue band (B2) and red band (B4) were consistently ranked among the most important variables in this study, a finding that aligns with previous research [43,44,45]. Previous studies have shown that the vegetation index (VI) derived from Sentinel-2 data plays a critical role in biomass estimation models and often contributes more significantly to the estimation of AGB than raw spectral reflectance bands [46]. Vegetation indices exploit the spectral reflectance and absorption characteristics of vegetation in specific bands to derive spectral transformations, thereby enhancing vegetation signals while minimizing interference from non-vegetated surfaces [47]. This study selected vegetation indices including TSAVI, PSSRa, RVI, and IPVI as predictor variables for modeling forest AGB. Previous studies have also demonstrated a strong association between these indices and AGB, highlighting their predictive ability [48,49]. In the context of the active remote sensing of Sentinel-1, SRTM DEM-derived terrain variables played a key role in AGB estimation, contributing more significantly than SAR backscatter and texture features, which is a finding consistent with previous studies [50]. Among the active remote sensing predictors, backscatter coefficients and texture features derived from VH polarization were more frequently selected and exhibited a stronger predictive power than those from VV polarization. This is attributed to the fact that VH polarization is more sensitive to vegetation water content, whereas VV polarization primarily responds to soil moisture. Similar advantages of C-band cross-polarization for vegetation detection have also been reported in other radar-based studies [51].

4.2. Advantages and Limitations of Active and Passive Remote Sensing Data Incorporation

Among all modeling configurations (whether based on single-source data or the incorporation of active and passive remote sensing data source), more than half of the selected predictor variables were texture features. This finding supports the conclusion that texture metrics are key determinants of AGB estimation accuracy, irrespective of the remote sensing data used [52]. This finding is consistent with previous studies, which also emphasized the dominant role of texture features in forest biomass estimation. Notably, GLCM-based texture features derived from Sentinel-2 imagery showed a stronger correlation with AGB than traditional spectral indices. In some regions, AGB appeared to be more strongly associated with texture-based measures than with spectral reflectance variables [53]. Among all extracted texture-based predictors, more than half were mean texture features. These features reflect the average grayscale level within the moving window and demonstrated a strong relevance to AGB estimation, which aligns with findings from previous studies [15]. In this study, most of the high-performing texture features were extracted using larger window sizes (e.g., 5 × 5 or 7 × 7), which is consistent with the findings of Li, who recommended the use of larger windows for an performance improvement [54]. However, Liu reported that texture features derived from smaller 3 × 3 windows often yielded the best predictive performance for AGB [55]. The effectiveness of texture features can be influenced by factors such as geographic variability, sensor characteristics, and land use types. This highlights the importance of optimizing the texture window size based on local landscape characteristics and the spatial resolution of the imagery used. In forest AGB estimation, the incorporation of multispectral data and SAR data—such as Sentinel-2 and Sentinel-1—represents an emerging and effective approach to improve AGB estimation accuracy. SAR operates based on side-looking radar geometry, exhibiting strong penetration capabilities into forest canopies, and is unaffected by atmospheric conditions such as cloud cover and illumination variability. These characteristics make SAR data particularly valuable in regions with frequent cloud cover, where optical imagery is often limited by cloud contamination or shadow effects. In this study, Sentinel-1 data helped to compensate for the temporal and spatial limitations of Sentinel-2, enriching spatial structural information and extending the vertical sensitivity of biomass detection. The integration of spectral (Sentinel-2) and structural (Sentinel-1) features, especially when incorporated with canopy height information, provided a more comprehensive representation of forest structure and, thus, improved AGB estimation performance. These findings align with and reinforce conclusions from previous studies [56,57]. Therefore, Sentinel-1 is an indispensable tool for cloud-prone regions, significantly reducing dependency on clear skies and enabling the reliable monitoring of AGB where optical data fails. However, its role is primarily complementary or as a replacement for specific cloud-sensitive tasks, rather than a complete substitute for all optical data.

4.3. Contribution of FCH to Biomass Estimates

The average forest AGB in Chenzhou City was estimated at 51.87 t·ha⁻¹ in 2019, slightly lower than the 2014 estimate of 53.68 t·ha⁻¹ reported by Xu for Guidong County–a forest-rich region in western Chenzhou [58]. This aligns with the spatial pattern identified in this study, wherein lower AGB values were generally observed in the western parts of the city. The 2019 estimate also exceeds the average forest AGB of Hunan Province in 2014, reported as 47.40 t·ha⁻¹ by Li based on arborous forest data and 41.27 t·ha⁻¹ by Dai based on NFCI data [59,60]. These comparisons suggest that forest resources in Chenzhou City experienced effective biomass accumulation during the 2014–2019 period, possibly reflecting successful conservation efforts or reduced anthropogenic disturbance. The estimated average forest AGB of 51.87 t·ha⁻¹ is comparable to the provincial mean of 52.36 t·ha⁻¹ reported in the NFCI of Hunan Province (2019) for the same period, suggesting that the results of this study are consistent with official inventory data and may serve as a reliable input for forest carbon sink assessments in Chenzhou City [61]. However, the estimate is substantially lower than the national average forest biomass of 85.29 t·ha⁻¹ reported in the ninth NFCI for the same period. This discrepancy may be attributed to the predominance of young and middle-aged forest stands in Chenzhou City, along with comparatively limited investment in collective forest management. FCH is a key structural parameter in AGB estimation and helps to reduce uncertainty in remote-sensing-based biomass assessment, consistent with previous findings [62,63]. However, the FCH estimates derived in this study exhibited a limited accuracy. This phenomenon may have been due to decreased forest density and increased canopy gaps, which amplify variability in signal penetration depth and phase center displacement between polarization channels. Such structural heterogeneity can introduce significant errors into InSAR-derived height estimates, particularly when a large number of gaps are unevenly distributed [64]. SAR signals are highly sensitive to terrain effects, and their complex scattering mechanisms further complicate analysis. The estimation precision of InSAR is closely linked to site-specific conditions such as wind speed, humidity, and temperature, all of which may introduce uncertainties in forest height inversion [65]. Furthermore, the relatively low energy of ICESat-2′s single-pulse laser reduces its signal-to-noise ratio, making the measurements more susceptible to noise. Forward scattering caused by interactions between the laser beam and atmospheric particles can further reduce the accuracy of ICESat-2 elevation measurements. Despite these limitations, our results demonstrated that incorporating FCH as a predictor in the inversion model can improve forest AGB prediction performance [66].

5. Conclusions

This study selected Chenzhou City as the research area and employed NFCI data, Sentinel-1 and Sentinel-2 imagery, and ICESat-2 LiDAR data, along with other auxiliary datasets, to estimate regional AGB by using MSR, ANN, k-NN, and RF models. The results demonstrated that incorporating active and passive remote sensing data significantly improved AGB estimation accuracy. Among the four evaluated models, RF performed the best, achieving an R² of 0.74 and an RMSE of 24.37 t·ha⁻¹. Texture features—particularly mean gray-level metrics derived from 7 × 7 windows—were the most influential variables, which could effectively capture spatial heterogeneity within the forest canopy. FCH, derived from the difference between the Sentinel-1 InSAR DSM and ICESat-2-corrected DEM, was significantly correlated with AGB (r = 0.64). Incorporating FCH into the model reduced the RMSE in high-biomass areas (>200 t·ha⁻¹) by 18.6%. Among the evaluated machine learning models, RF exhibited the highest resistance to overfitting, owing to its ensemble learning framework and random feature selection strategy. The R² of the RF model exceeded that of ANN, k-NN, and MSR by 19.4%, 29.8%, and 54.2%, respectively. The spatial distribution of forest AGB in Chenzhou City in 2019 exhibited a distinct “low in the west and high in the east” pattern, with an average biomass density of 51.87 t·ha⁻¹ and a total AGB stock of 69.44 × 10⁶ t. High-biomass regions were primarily concentrated in the Dongjiang Lake Scenic Area, with a maximum value of 256.41 t·ha⁻¹, whereas low-biomass zones coincided with areas of urban expansion, with a minimum of 6.38 t·ha⁻¹. This spatial pattern was significantly associated with topographic variables such as elevation and slope, and also corresponded closely to the distribution of ecological protection zones.

Methodologically, this study introduced a DEM differencing approach based on an ICESat-2-corrected SRTM DEM, which reduced canopy height inversion error by 21.3% compared to the GLAD global canopy height product. In addition, a multi-dimensional feature integration strategy combining red-edge spectral bands, SAR-derived texture features, and FCH was developed, effectively mitigating the saturation effect commonly observed in traditional optical data in high-biomass regions. In terms of practical application, the study provides a reliable technical framework for subtropical forest carbon sink monitoring, supporting decision making for forest management under the implementation of the “dual carbon” national strategy in Chenzhou City. The study also demonstrated the cost efficiency of using freely available Sentinel series satellite data for large-scale AGB mapping, providing a valuable reference for forest carbon sink monitoring in subtropical regions of China and comparable ecological zones globally.

Although the study demonstrated promising results, several limitations remain to be addressed. From the perspective of data quality, the Sentinel-1 C-band is significantly affected by atmospheric water vapor during the rainy season (June–August), leading to temporal instability and a reduced backscatter consistency. From the perspective of model performance, RF is sensitive to spatial misalignment among heterogeneous data sources, with a reduction of up to 0.15 in R² observed when spatial offsets exceed 30 m. Future research may be conducted in the following two directions: (1) in terms of data incorporation, red-edge band time series analysis could be enhanced by incorporating Sentinel-3 OLCI data, while the integration of ALOS-2 PALSAR-2 L-band SAR may improve canopy penetration and structural sensitivity in high-biomass areas, and (2) for practical applications, an AGB–carbon sink conversion model could be developed by integrating forest age and species composition, thereby supporting carbon accounting in regional carbon markets.