Estimation of Aboveground Biomass of Picea schrenkiana Forests Considering Vertical Zonality and Stand Age

Abstract

The aboveground biomass (AGB) of forests reflects the productivity and carbon-storage capacity of the forest ecosystem. Although AGB estimation techniques have become increasingly sophisticated, the relationships between AGB, spatial distribution, and growth stages still require further exploration. In this study, the Picea schrenkiana (Picea schrenkiana var. tianschanica) forest area in the Kashi River Basin of the Ili River Valley in the western Tianshan Mountains was selected as the research area. Based on forest resources inventory data, Gaofen-1 (GF-1), Gaofen-6 (GF-6), Gaofen-3 (GF-3) Polarimetric Synthetic Aperture Radar (PolSAR), and DEM data, we classified the Picea schrenkiana forests in the study area into three cases: the Whole Forest without vertical zonation and stand age, Vertical Zonality Classification without considering stand age, and Stand-Age Classification without considering vertical zonality. Then, for each case, we used eXtreme Gradient Boosting (XGBoost), Back Propagation Neural Network (BPNN), and Residual Networks (ResNet), respectively, to estimate the AGB of forests in the study area. The results show that: (1) The integration of multi-source remote-sensing data and the ResNet can effectively improve the remote-sensing estimation accuracy of the AGB of Picea schrenkiana. (2) Furthermore, classification by vertical zonality and stand ages can reduce the problems of low-value overestimation and high-value underestimation to a certain extent.

1. Introduction

Forest biomass stands as a pivotal indicator of the productive capacity of forest ecosystems, consistently serving as a crucial evaluation factor in assessments of forest carbon budgets [1]. The biogeochemical cycling characteristics of Xinjiang exert a profound influence on the evolution of terrestrial ecosystems in the Central Asian region, shaping regional climate patterns and impacting greenhouse gas source and sink totals [2]. At the heart of the Central Asian mountain range, the Tianshan Mountains are home to vast stands of Picea schrenkiana forests, which represent the most significant terrestrial ecosystems in temperate arid regions. Picea schrenkiana, as the dominant tree species in the mountain forests of the Tianshan Mountains, plays a vital role in water conservation, soil retention, and maintaining the ecological environment of the mountains, occupying an extremely important ecological and geographical position. By conducting research on the biomass of the mountain ecosystems in the Tianshan Mountains, with a particular focus on Picea schrenkiana forests, we will lay a foundation for studies on the carbon cycle of mountain forest ecosystems in Central Asia. Furthermore, this research can provide valuable insights into the evolution of terrestrial ecosystems, regional climate dynamics, and greenhouse gas emissions and sequestration in the Central Asian region, thereby possessing significant scientific importance.

Biomass is categorized into aboveground and underground components. Given the difficulties in gathering underground biomass data, researchers mainly center their studies on aboveground biomass (AGB) [3]. Traditional measurement methods, while accurate, are labor-intensive and time-consuming, making them impractical for large-scale forest biomass monitoring [4]. The advent of remote-sensing technology addresses this limitation [5]. In the realm of remote-sensing-based forest biomass estimation, scholars from both domestic and international backgrounds have extensively utilized optical remote-sensing data since the last century, achieving significant advancements [6,7]. For instance, Chen et al. pioneered the use of HJ-1 satellite data to monitor Picea schrenkiana forest AGB in the western Tianshan Mountains, revealing spatial heterogeneity and spatiotemporal differentiation patterns of AGB and productivity in this region [8]. However, optical remote sensing has inherent limitations, particularly its weak spectral penetrability, which restricts its ability to capture vertical vegetation information beneath the canopy [9]. To overcome these limitations, radar data, with its longer wavelengths, has been increasingly adopted. Radar can penetrate vegetation layers and provide more detailed stand structure and stand factor information, making it a valuable complement to optical data [10]. Synthetic Aperture Radar (SAR), as an advanced active remote-sensing technology, operates continuously under all weather conditions, further enhancing its utility [11]. Studies have demonstrated the effectiveness of SAR in AGB estimation. For example, Cartus et al. utilized multi-frequency radar backscatter observations to estimate tropical forest AGB, highlighting the structural discriminative capacity of SAR data [12]. Wei et al. explored the potential of GF-3 polarization decomposition components for estimating artificial forest canopy AGB, finding that multiple polarization components exhibit high sensitivity to AGB variations [13]. Multi-source remote-sensing data can fully capitalize on the complementary aspects of information derived from various data sources. By skillfully integrating remote-sensing data from diverse sensors collected at different times and with varying resolutions, the benefits of remote sensing can be maximized. Researchers such as Zhang et al. [14] and Forkuor et al. [15] have developed forest AGB estimation models using optical data and SAR data. Wang et al., in their study, utilized GF-6 and GF-3 dual-polarization data to estimate the AGB of natural forest land in Gongliu County, Xinjiang. The study revealed that the integration of optical and SAR data can remarkably enhance the accuracy of AGB estimation [16]. In their assessment of Yichun City’s AGB, Liu et al. employed both optical and SAR data, concluding that AGB estimation models specifically designed for different forest types are more precise and stable compared to generalized approaches [17]. These findings underscore the importance of integrating multi-source data and considering forest-specific characteristics in AGB estimation.

As a key geographical factor, topography exerts a profound influence on forest ecosystems [18]. Variations in elevation, slope, and aspect create unique geomorphological units, driving the vertical stratification of environmental factors such as thermal conditions, precipitation patterns, and solar radiation intensity. These changes result in marked differences in the vertical distribution of climate conditions, soil properties, light intensity, and other environmental factors [19,20]. These differences, in turn, have profound effects on forest AGB. Studies have shown that the trends of AGB with elevation vary significantly across different regions and ecosystems. For example, Wang et al., using a 30-m resolution AGB dataset of the Tibetan Plateau, found that grassland AGB values exhibit a monotonically increasing trend below 4800 m [21]. Wu et al. found different biomass at different elevations by studying Qilian Mountain Picea schrenkiana, in which the growth rate of AGB was highest at low elevations [22]. Maza et al. focusing on the elevation gradient of AGB in the Andean-Amazonian forests, found a negative correlation between AGB and elevation [23]. These differences may stem from the interaction between elevation-induced climatic changes, such as temperature lapse rates and orographic precipitation, and differences in vegetation types. In addition to the influence of elevation, researchers have also focused on the impact of topographic factors on the accuracy of AGB estimation. For instance, Chen et al. found that the introduction of a digital elevation model (DEM) improved model accuracy by 13.5% [24], while Ye et al. demonstrated that models incorporating aspects effectively mitigate the overestimation of low values and underestimation of high values, further enhancing model precision [25]. In recent years, the application of high-resolution satellite data has further strengthened the advantages of topographic factors. Han et al., using GF-1 and Sentinel-1, found that the integration of multi-source remote-sensing data and topographic factors significantly improves AGB estimation accuracy in complex terrain regions [26]. Forest stand ages are also an important factor affecting forest AGB [27]. Huang et al., focusing on Chinese fir plantations, found that the spatial structure factors affecting biomass accumulation vary significantly with stand age and proposed corresponding management measures [28]. Lee et al. identified forest age as a key variable regulating AGB in alpine-subalpine forests, revealing that older stands exhibit higher biomass [29]. Wu et al. demonstrated that models integrating both aspect and stand age outperform other models, highlighting the importance of considering multiple factors, such as topography and stand age, in AGB estimation [30]. These findings emphasize the complexity of AGB estimation and the importance of integrating diverse variables for improved accuracy.

This study concentrates on the forest area situated within the Kashi River Basin of the Ili Valley, nestled in the western Tianshan Mountains, as its research area. The Picea schrenkiana (Picea schrenkiana var. tianschanica) in this region are categorized into three types: the Whole Forest without vertical zonation and stand age (hereinafter referred to as the whole forest), Vertical Zonality Classification without considering stand age (hereinafter referred to as vertical zonality), and Stand-Age Classification without considering vertical zonality (hereinafter referred to as stand age). Due to insufficient sample size when simultaneously considering stand age and vertical zonality, which cannot meet the basic requirements for statistical analysis, this scenario was not included in the study. For these three classification cases, the performance of three feature combinations, namely single optical data, single radar data, and opto-radar multi-source remote-sensing data, is comprehensively investigated in three models: XGBoost, BPNN, and ResNet. Moreover, a comparative analysis of the estimation accuracy is conducted. Through this multi-dimensional and multi-method integrated analysis, the study aims to (1) improve the accuracy of remote-sensing estimation for AGB of Picea schrenkiana forests in the western Tianshan Mountains of Xinjiang and (2) expand the application scope and depth of domestic high-resolution satellites in forest resource monitoring. The novelty of this study lies in its integration of multi-source remote-sensing data and the incorporation of key variables such as vertical zonality and stand age, proposing a more refined and accurate method for forest biomass estimation. Furthermore, the findings not only enhance the understanding of biomass distribution patterns in the forest ecosystem of the western Tianshan Mountains but also provide valuable methodological references and practical support for other regions globally with complex terrain and diverse forest structures.

2. Materials and Methods

2.1. Study Area

Nilka County (81°85′ E to 84°58′ E, 43°25′ N to 44°17′ N) is in the northwestern part of Xinjiang, China. It lies within the western hinterland of the Tianshan Mountains and within the Kashi River Basin, which is part of the Ili Valley. This area experiences a continental climate within the northern temperate zone, characterized by an average annual temperature of 6.8 °C and a total annual precipitation of 406.9 mm, with the majority of rainfall occurring from April to July. The region distinctly exhibits mountain climate characteristics. The terrain features a slope from the northeast to the southwest. The forest area has a simple composition of tree species, mainly natural Picea schrenkiana, a small number of Populus spp. and Betula spp., and a few species of understory vegetation. Sub-compartments, as the smallest spatial units for forest resource surveys, are divided based on differences in ecological attributes such as stand origin, age class, and site type. In this study, the sub-compartments of Picea schrenkiana forests in the Tianshan Mountains were used as the basic units. Pure Picea schrenkiana forests, i.e., sub-compartments with Picea schrenkiana as the tree species composition, were selected as the study area. The method of stratified sampling was adopted, and the Picea schrenkiana forests in the study area were divided according to two schemes: different vertical zonations and different stand ages. Finally, 540 sub-compartments were sampled from the forest resources inventory database of the Nilke Forest Farm as research samples. The average size of the sub-compartments is 18.36 hm². The soil in this area is rich in humus and highly fertile. The most prevalent and expansive soil type in the forest area is mountain gray-cinnamon soil, which is ideally suited for Picea schrenkiana growth. Figure 1 illustrates the schematic diagram of the study area.

2.2. Data Processing

2.2.1. GF-1 PMS and GF-6 PMS Data

All the remote-sensing data used in this paper are sourced from the National Remote-Sensing Data and Application Service Platform (https://www.cpeos.org.cn/, accessed on 12 April 2024). GF-1 and GF-6 are primarily used to acquire optical remote-sensing image data of the Earth’s surface. They are equipped with two distinct sensors. The high-resolution camera (PMS) is capable of capturing panchromatic black-and-white images at a resolution of 2 m and multispectral color images at a resolution of 8 m. The multispectral sensor covers four bands: blue, green, red, and near-infrared (NIR), which are utilized to obtain basic spectral information of ground objects and play a crucial role in various fields, including land resource surveying and ecological environment monitoring [31]. The 16 m multispectral medium-resolution wide-swath camera (WFV) boasts a large imaging swath width, enabling the acquisition of large-area ground images in a short period. We performed optical data preprocessing in ENVI software 5.6 (Exelis Visual Information Solutions, Inc., Boulder, CO, USA). To maintain consistency with the resolution of SAR data, no fusion operations were performed on the optical data, and only multispectral data were utilized.

2.2.2. GF-3 PolSAR Data

The GF-3 satellite, China’s first independently developed C-band synthetic aperture radar (SAR) satellite, features 12 imaging modes and provides rich polarization data. The full-polarization image data can simultaneously capture the transmission and reception information of horizontal polarization (H) and vertical polarization (V), encompassing four polarization modes: horizontal transmit-horizontal receive (HH), horizontal transmit-vertical receive (HV), vertical transmit-horizontal receive (VH), and vertical transmit-vertical receive (VV). In this study, the 20 pieces of SAR data utilized are all Quad-pol stripmap (QPSI) images with a resolution of 8 m. The preprocessing of radar data encompasses radiometric calibration, multi-look processing, filtering, and geocoding [32]. After undergoing radiometric calibration, the radar data can more accurately reflect the echo characteristics of ground objects. Multi-look processing and filtering can enhance the visual quality of the image, making the geometric features of the image more akin to the actual situation on the ground while also reducing speckle noise to a certain extent [33]. Since the complex image acquired by the SAR sensor is a slant range image, geocoding is necessary to convert the SAR data from the slant range coordinate system to the geographic coordinate system [34]. The preprocessing of GF-3 data was conducted using PIE-SAR 6.3, a software developed by Piesat Information Technology Co., Ltd. in Beijing, China.

2.2.3. Auxiliary Data

Seven pieces of ASTER GDEM V3 30 m data were downloaded from the geospatial data cloud platform (https://www.gscloud.cn/, accessed on 20 April 2024). The terrain factors, including slope, aspect, and elevation, redistribute the energy of solar radiation, thereby influencing vegetation growth and the accumulation of forest AGB. Use ArcGIS 10.8 (Environmental Systems Research Institute, Inc., Redlands, CA, USA) to extract terrain factors such as elevation, slope, and aspect. Based on the classification of forest vertical zones in the Yili area by Zhang et al. [35], the forest vertical zones in the study area are categorized according to elevation. The classification criteria are outlined in Table 1. Figure 2 is the topographic feature map of the study area.

2.2.4. Resource Data

The forest resources inventory data for management encompasses fields such as sample plot number, average diameter at breast height (DBH), stand age, crown density, soil type, soil layer thickness, and standing stock volume per hectare. Projection coordinate transformation is applied to these resource data to ensure alignment with the projection coordinates of remote-sensing images. The geographic coordinate system is WGS-1984, and the projected coordinate system is Universal Transverse Mercator (UTM), with the projection zone number being UTM Zone 44N. The sub-compartments of the forest farm, categorized by different altitudes and forest ages, are delineated, and only the fields relevant to this study are retained. By referring to the regression equation of Picea schrenkiana forest biomass and stock volume established by Fang et al. [36], the Picea schrenkiana biomass is estimated using these resource data. The formula is as follows.

(1)$B=0.4642V+47.499,$

Note: Biomass (B) is measured in t·hm⁻², stem volume (V) is measured in m³·hm⁻².

According to the forestry industry standards of the People’s Republic of China, “Regulations for age-class and age-group division of main tree species”, with 20 years as the stand-age interval, Tianshan Picea schrenkiana forests are divided into young forest, middle-aged forest, near-mature forest, mature forest, and over-mature forest according to their different age stages. The classification criteria are presented in Table 2, and the Stand-Age Classification map of the Picea schrenkiana forest is shown in Figure 3. In the study area, natural forests account for the vast majority. The age class of a sample plot is determined by the average age of the trees within the plot.

Picea schrenkiana forests in the western Tianshan Mountains are mainly distributed in three key vertical zones: middle mountain forest-meadow belt, upper middle mountain forest-meadow belt, and subalpine open-forest belt. The combined conditions of climate and soil in these three vertical zones engender a highly suitable environment conducive to the growth of Picea schrenkiana and cultivate an ideal ecological niche tailored to it. However, the distribution of Picea schrenkiana forest is very scanty in the middle and low mountain forest-steppe belt and below. This is mainly because the precipitation in this area is insufficient to meet the water demand of Picea schrenkiana growth [37]. More importantly, human activities have caused some interference with the growth of Picea schrenkiana, making it difficult for Picea schrenkiana forests to form and develop in this area. In the area above the subalpine open-forest belt, the extreme climatic conditions exceed the ecological adaptation range of Picea schrenkiana, preventing Picea schrenkiana forest from establishing itself here. Overall, small classes of Picea schrenkiana in these two unsuitable areas account for only 0.05% of the total.

From the perspective of stand age, Picea schrenkiana occupies a large proportion in the near-mature forest, mature forest, and over-mature forest stages. Specifically, the Picea schrenkiana population in these stages is more stable in the ecosystem and has stronger adaptability to the environment. In contrast, the number of Picea schrenkiana sub-compartments in young and middle-aged forests only accounts for 8.61% of the total, which provides limited reference for this study. This is not only because the number of individuals in the whole Picea schrenkiana forest ecosystem is small but also because it is difficult to reflect the overall ecological characteristics and biomass distribution of Picea schrenkiana forest. More importantly, young and middle-aged forests are still in the dynamic stage of growth and development, with unstable ecological characteristics and large data fluctuations, which impedes the construction of accurate models and in-depth analysis and research.

Based on the previous analysis, this study centers on three main vertical zones, namely the middle mountain forest-meadow belt, the upper middle mountain forest-meadow belt, and the subalpine open-forest belt, as well as three growth stages: near-mature forest, mature forest, and over-mature forest. Through the modeling and analysis of these key areas and stages, the aim is to accurately reveal the internal relationship between the distribution of Picea schrenkiana forest, vertical zonality, and stand age and provide a solid theoretical basis for the ecological protection and scientific management of Picea schrenkiana forest in the West Tianshan Mountains.

2.3. Feature Extraction

2.3.1. Optical Image Features

• (1). The Surface Reflectance

The original bands of remote-sensing images contain abundant object information and are the cornerstone of object recognition. In optical remote-sensing data, different bands correspond to different reflection characteristics of objects. The spectral information is highly correlated with vegetation parameters. The band designs of GF-1 and GF-6 are identical, although there are slight differences in the specific band ranges: the blue band (0.45~0.52 μm for both), the green band (0.52~0.59 μm for GF-1 and 0.52~0.60 μm for GF-6), the red band (0.63~0.69 μm for both), and the near-infrared band (0.77~0.89 μm for GF-1 and 0.76~0.90 μm for GF-6). Although there are slight differences, these discrepancies can be effectively mitigated through appropriate data processing. Therefore, this study combines the data from both satellites to form continuous and consistent optical satellite imagery, from which the spectral values of these four bands are extracted.

• (2). Vegetation Index

Analyzing and comparing only individual bands or multiple single bands is insufficient to comprehensively understand the complexity of vegetation remote sensing. As a widely distributed object type on the Earth’s surface, vegetation demonstrates complex and diverse spectral characteristics. To more effectively utilize satellite remote-sensing data to characterize vegetation status based on the spectral properties of vegetation, various linear or nonlinear combinations of satellite bands are performed, such as addition, subtraction, multiplication, and division. These operations are used to extract a total of eight vegetation indices, including the normalized difference vegetation index (NDVI), green normalized difference vegetation index (GNDVI), ratio vegetation index (RVI), enhanced vegetation index (EVI), difference vegetation index (DVI), infrared percentage vegetation index (IPVI), soil adjusted vegetation index (SAVI), and normalized difference mountain vegetation index (NVDMVI) [38,39]. The formula for calculating the vegetation index is shown in Table 3.

• (3). Texture Features

Texture features contain and reflect the spatial variation of the grayscale of the image and can better reflect the real structure of ground objects. In areas with more complex stand structures, they have a positive effect on AGB estimation [40]. Many scholars have confirmed that the inclusion of spectral texture features in the model significantly improves estimation accuracy [41]. After performing principal component analysis on the optical data in this study, it was found that the contribution rate of the first principal component is 97.5%. Based on the gray-level co-occurrence matrix (GLCM), eight texture features were extracted from six windows (3 × 3, 5 × 5, 7 × 7, 9 × 9, 11 × 11, 13 × 13) of the first principal component, including mean (MEAN), variance (VAR), homogeneity (HOM), contrast (CON), dissimilarity (DIS), entropy (ENT), second moment (ASM), and correlation (COR). The texture feature formula is shown in Table 4.

2.3.2. Radar Image Features

• (1). Backscatter Coefficient

The backscatter coefficient of radar data is a complex physical quantity, representing the result of the comprehensive influence of multiple factors within the forest stand. In the forest environment, factors such as ground roughness, vegetation coverage, and vegetation water content are intertwined and jointly determine the magnitude of the backscatter coefficient [42]. In this study, four polarized backscatter coefficients, HH, HV, VH, and VV, are extracted as the characteristic variables of the model.

• (2). Polarization Decomposition

The forest represents an incoherent target and exhibits a distributed characteristic with a universal time-varying effect [43]. Based on PIE-SAR 6.3 software, this study aims to extract the four incoherent polarization decomposition parameters (the Freeman decomposition, the Pauli decomposition, the H/A/Alpha decomposition, the Yamaguchi decomposition) as well as the one coherent polarization decomposition parameter (the Krogager decomposition) in order to deeply explore the forest feature information in radar data [32]. The related parameters of polarization decomposition are shown in Table 5.

• (3). Radar Vegetation Index

The radar vegetation index is a parameter derived from eigenvector decomposition. As vegetation density increases, the radar vegetation index will change accordingly, providing important reference data for the investigation and management of forest resources [44]. The radar vegetation index, utilizing the Freeman-Durden decomposition, takes into account the proportion of the three scattering mechanism components: dihedral scattering ($F_d$), surface scattering ($F_s$), and volume scattering ($F_v$). The radar vegetation index incorporates this information related to vegetation scattering by employing a specific algorithm, thereby enabling it to quantitatively describe the density of vegetation. The formula is as follows.

(2)$Freeman\_RVI={\displaystyle \frac{F_v}{F_v+F_d+F_s}},$

2.4. Model Introduction

The measured data and the factors, after having undergone feature optimization, under three combination forms of single optical data, single SAR data, and multi-source remote-sensing data combining optical with SAR are inputted into the eXtreme Gradient Boosting (XGBoost), Back Propagation Neural Network (BPNN), and Residual Network (ResNet) models, respectively, to estimate the AGB of Picea schrenkiana forests across three classification criteria: the whole forest, different vertical zonality, and different stand ages. In the model construction phase, Python 3.7 is selected as the development language, and the powerful PyCharm 3.8.3 editor is used for code writing. To ensure the reliability and generalization ability of the model, 30% of the data are randomly selected as the validation set to evaluate the performance of the model. During the model training process, through continuous attempts at different parameter combinations, a Picea schrenkiana forest biomass inversion model with the highest accuracy is finally established.

2.4.1. XGBoost

XGBoost is a powerful machine-learning algorithm. It demonstrates superior performance in both classification and regression tasks within the domain of weakly supervised learning in machine learning. The training process is optimized by introducing an innovative sparse-aware algorithm for sparse data processing and a weighted quantile sketch for approximate tree learning [45]. The XGBoost model aims to prevent overfitting and, at the same time, reduce computational cost by keeping predictions at optimal computational efficiency through simplification and regularization.

Suppose a dataset $\mathcal{D}=(x_i,y_i)(\left|\mathcal{D}\right|=n,x_i\in R^m,y_i\in R)$, where $n$ is the number of samples, $m$ is the feature dimension of each sample, and the additive model is defined as:

(3)$\begin{array}{c}{\widehat{y}}_i^{(0)}=0\\ {\widehat{y}}_i^{(1)}=f_1(x_i)={\widehat{y}}_i^{(0)}+f_1(x_i)\\ {\widehat{y}}_i^{(2)}=f_1(x_i)+f_2(x_i)={\widehat{y}}_i^{(1)}+f_2(x_i)\\ \cdots \\ {\widehat{y}}_i^{(k)}={\mathsf{\Sigma }}_{k=1}^K{f}_k(x_i)={\widehat{y}}_i^{(t-1)}+f_t(x_i)\end{array},$

The algorithm model of XGBoost can be expressed as:

(4)${\widehat{y}}_i=\varphi (x_i)={\mathsf{\Sigma }}_{k=1}^K{f}_k(x_i),f_k\in \mathcal{F},$

Here, $\mathcal{F}=f(x)=w_{q(x)}(q:R^m\to T,w\in R^T)$ represents the sample space of regression trees (typically CART trees are used). The function $q$ maps each sample to a leaf node index, where $T$ denotes the number of leaf nodes in each tree and $w$ represents the weights of the leaf nodes. Each $f_k$ corresponds to an independent tree defined by $q$ (tree structure) and $w$ (leaf node weights).

2.4.2. BPNN

The BPNN is one of the currently commonly used neural networks. It calculates the error through forward propagation and realizes the training of the neural network by backpropagating the error gradient. The structure consists of three components: the input layer, the hidden layer, and the output layer [46]. Increasing the number of neurons in the hidden layer can enhance the training accuracy of the BP neural network. If the number of neurons in the hidden layer is insufficient, the network may obtain too little information to solve problems effectively; if the number is too large, not only will the training time increase, but more importantly, an overfitting problem may arise if there are too many hidden layer nodes, resulting in an increase in test error and a decrease in generalization ability. The model framework is shown in Figure 4.

2.4.3. ResNet

In traditional deep CNNs, as the number of network layers increases, gradient disappearance (the gradient value gradually diminishes to zero) or explosion (the gradient value skyrockets) occurs in the backpropagation process, making it difficult to train the network. Residual connections enable gradients to propagate more effectively in the network by directly passing the input signal [47]. During backpropagation, the gradient can be directly passed back through the shortcut connection without passing through all the convolutional layers, thereby reducing the risk of gradient disappearance or explosion. The model framework is shown in Figure 5.

2.5. Accuracy Verification

A total of 30% of the data are randomly selected as the test set to evaluate the model’s applicability. The coefficient of determination (R²) and the root mean square error (RMSE) are employed as evaluation metrics to assess the accuracy of each model result. R² is utilized to measure the goodness of fit of the model to the data and to explain the proportion of variance in the target variable. The value range is from 0 to 1. R² = 1 indicates that the model perfectly fits the data. RMSE is an index used for evaluating regression models and is used to measure the deviation between the model’s predicted value and the true value.

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

(6)$R_{\mathrm{mse}}=\sqrt{{\displaystyle \frac{{\displaystyle \sum _{i=1}^n{\left(y_i-{\mathrm{y}}_i\right)}^2}}{n}}},$

Types (5) and (6): $y_i$ represents the actual observed value, ${\widehat{y}}_i$ represents the predicted value, ${\overline{y}}_i$ represents the mean of the actual observations, and n represents the number of samples.

3. Results

3.1. Determination of Characteristic Factors

The selection of modeling factors plays a crucial role in determining the accuracy of the model. If numerous modeling factors are introduced arbitrarily, many of them may have little or no correlation with AGB. This excessive number of factors not only increases the computational burden but also prolongs the model’s training duration. To ensure the reliability of the model, prior to the modeling process, the normality of the data was first examined using Q-Q plots in the SPSS data analysis software (IBM SPSS Statistics v26, IBM Corp., Armonk, NY, USA), a widely used tool for statistical analysis in various research fields. After confirming the data’s normality, the Pearson correlation coefficient was employed to examine the correlation between characteristic factors and AGB. Based on the results of this analysis, factors that were significantly associated with AGB were carefully selected. This approach helps in identifying the most relevant factors, thereby enhancing the model’s accuracy and efficiency.

The correlation coefficients between different vertical zonality and characteristic factors from various data sources are presented in Table 6. When categorized by vertical zonality, Band3, DVI, HH polarization, HV polarization, Freema_Rvi, and SLOPE, all exhibit notably high correlation coefficients. In the middle mountain forest-meadow belt, MEAN_3 shows the strongest correlation with a coefficient of 0.432. Subsequently, Band3 and HV polarization have correlation coefficients of −0.372 and 0.363, respectively. The correlation coefficient between Band3 and the upper middle mountain forest-meadow belt is −0.453. In this belt, the correlations with MEAN_3 and Band2 are 0.504 and −0.473, respectively. In SAR, the highest correlation coefficients of HH polarization and HV polarization with the upper middle mountain forest-meadow belt are −0.378 and −0.344, respectively. The correlation coefficient between the vegetation index DVI and the subalpine open-forest belt is −0.374. Moreover, in the subalpine open-forest belt, the correlation of SAR data is higher than that of optical data, and the correlation coefficients of HH polarization and HV polarization are the highest, being 0.406 and 0.383, respectively.

The correlation coefficients of different stand ages and characteristic factors of different data sources are shown in Table 7. In near-mature forests, MEAN_3 and Band2 have the highest correlation, with correlation coefficients of 0.568 and −0.571, respectively. Second, VH polarization, HV polarization, and Pauli_Odd in SAR data have correlation coefficients of −0.424, −0.413, and −0.403, respectively. Among vegetation indices, the correlation coefficients of RVI, DVI, and NDVI are 0.346, 0.340, and 0.312, respectively. In mature forests, MEAN_5, Band3, and Band2 have the highest correlation, which is 0.515, −0.500, and −0.492, respectively. In SAR data, the correlation coefficients of HH, VH, and Freeman_Rvi are −0.431, −0.358, and 0.266, respectively. In over-mature forests, VH polarization has the highest correlation, with a correlation coefficient of −0.423. Second, Band3, DVI, and Freeman_Vol have correlation coefficients of −0.412, 0.399, and −0.390, respectively. Among topographic factors, SLOPE and DEM have high correlations with all three stand ages. The highest value of SLOPE appears in mature forests, with a correlation coefficient of 0.427.

3.2. Estimation Results of Different Vertical Zonality

Based on the in-depth correlation analysis in the early stage, this study divides the data sources into single optical data (Optical data + Topographic Factor), single radar data (SAR data + Topographic Factor), and multi-source remote-sensing data (Optical data + SAR data + Topographic Factor). The aim is to thoroughly investigate the specific impact of various data sources on the accuracy of the Picea schrenkiana forest biomass estimation model under different classification methods.

The accuracy of different models under different vertical zonality and different data sources is shown in Table 8. In the upper middle mountain forest-meadow belt, when only optical data are used, the R² of the ResNet is 0.74, which is much higher than the R² of the XGBoost of 0.57 and the R² of the BPNN of 0.66. Due to the lesser number of input values in the subalpine open-forest belt, the highest inversion accuracy combination is multi-source remote-sensing data plus ResNet, with R² = 0.69 and RMSE = 12.63 t·hm⁻². The lowest accuracy appears in the estimation using a single data source with XGBoost, where R² = 0.51 and RMSE = 24.56 t·hm⁻². It can be seen that the performance of the ResNet in combinations involving three data sources outperforms that of the XGBoost and BPNN. In addition, the performance of different data sources under the application of the same model also varies. In the middle mountain forest-meadow belt, when only optical data are input into the XGBoost, the obtained R² = 0.52 and RMSE = 23.40 t·hm⁻². Based on the SAR data source, R² = 0.46 and RMSE = 24.87 t·hm⁻². The R² value of Picea schrenkiana forest estimation utilizing multi-source data is increased to 0.62, and the RMSE value is reduced to 20.69 t·hm⁻², indicating that integrating various data sources can further improve the accuracy of the model. In the upper middle mountain forest-meadow belt, for the BPNN, the R² of the optical data source is 0.66 and RMSE = 19.65 t·hm⁻², while the R² of the SAR data source is 0.65 and RMSE = 22.21 t·hm⁻². When both optical and SAR data sources are used simultaneously, R² = 0.71 and RMSE = 19.21 t·hm⁻², which indicates that combining multiple data sources has a positive impact on model accuracy. The accuracy of each model using multi-source data in the case of vertical zonality is shown in Figure 6.

3.3. Estimation Results of Different Stand Ages

The accuracy of different models under different stand ages and different data sources is shown in Table 9. Compared to near-mature forest and mature forest, over-mature forest have a smaller area and fewer input values, resulting in lower estimation accuracy than that of near-mature forest and mature forest. In mature forest, when only optical data are used, for XGBoost, R² = 0.60 and RMSE = 16.47 t·hm⁻². For BPNN, R² = 0.63 and RMSE = 15.37 t·hm⁻². For ResNet, R² = 0.67 and RMSE = 14.91 t·hm⁻². In near-mature forest, when multi-source remote-sensing data are used, for XGBoost, R² = 0.60 and RMSE = 13.40 t·hm⁻². For BPNN, R² = 0.69 and RMSE = 12.60 t·hm⁻². For ResNet, R² = 0.77 and RMSE = 10.06 t·hm⁻². Under the same data source and different model conditions, ResNet demonstrates the highest accuracy. When the model is fixed as using ResNet, in over-mature forest, when only optical remote-sensing images are used, R² = 0.74 and RMSE = 21.58 t·hm⁻². When only SAR data are used, R² = 0.72 and RMSE = 22.44 t·hm⁻². When combining optical and SAR data, R² = 0.77 and RMSE = 22.68 t·hm⁻². This once again confirms that combining two data sources can further improve the accuracy of ResNet and enable it to achieve optimal performance. The accuracy of each model using multi-source data in the case of stand ages is shown in Figure 7.

3.4. Estimation Results of the Whole Forest

It is easy to deduce from Table 8 and Table 9 that multi-source data have higher prediction accuracy than a single data source. Therefore, multi-source data are individually incorporated into the three models, as shown in Table 10 and Figure 8, respectively. It is readily apparent that the ResNet has the highest accuracy, with R² = 0.71 and RMSE = 20.95 t·hm⁻². The XGBoost model has the lowest accuracy, with R² = 0.55 and RMSE = 27.71 t·hm⁻². To sum up, there are differences in the accuracy performance of various data sources and models under different modeling methods. Generally, in terms of data, integrating optical and SAR data sources can significantly improve the model’s accuracy. The ResNet model demonstrates notably higher accuracy compared to the other two models. Among different modeling approaches, the accuracy of AGB estimation is higher for vertical zonality and stand-age scenarios than for the entire forest. Among them, the accuracy of estimation based on stand age is the highest.

3.5. AGB Mapping

Through the above comparison, it can be inferred that the combination of ResNet and multi-source remote-sensing data has the highest accuracy. Using this model and data source, we draw AGB distribution maps for three scenarios: the entire forest, classification by vertical zonality, and classification by stand ages. These maps are shown in Figure 9.

4. Discussion

In the realm of AGB estimation research, the acquisition and utilization of comprehensive and precise information are of paramount significance. This study is based on the forest AGB estimation outcomes obtained from three distinct data sources: single optical data, single radar data, and integrated optical-radar data. In this study, the estimation accuracy of multi-source remote-sensing data is significantly higher than that of single-source data. This conclusion has been rigorously verified by an extensive body of research. Attarchi et al. [48] evaluated 11 distinct multiple linear regression models using optical and SAR data. Their research findings clearly show that integrating multispectral and SAR data can effectively reduce uncertainties from data limitations, thus improving the accuracy of biomass estimation. The results of our current study are highly congruent with this view. This further validates the significant advantages of multi-source remote-sensing data in biomass estimation.

Given the challenge of acquiring forest AGB products with temporal and spatial resolutions fully consistent with those of this study, we selected the AGB product with the highest resolution and the closest acquisition time (the 30 m resolution AGB product in China in 2019) for a comparative analysis with the AGB estimated in this study. As shown in Figure 9, the biomass distribution range obtained from the whole-forest estimation in this study mainly fell within the range of 70–190 t·hm⁻², with an average biomass per unit area of 126.49 t·hm⁻². Based on 4789 ground-measured AGB data and multi-source remote-sensing data, Yang [49] derived a 30 m resolution biomass estimation for China in 2019. After converting the research findings, the average biomass per unit area in this study area is approximately 129.34 t·hm⁻², which is comparable to the results of this study. However, an in-depth analysis of Yang’s estimation results reveals that the biomass distribution range mainly ranges from 100 to 160 t·hm⁻². Taking the forest resources inventory data for management as the true value, the biomass in the study area is mainly distributed in the range of 70–190 t·hm⁻². As shown in Table 11, the RMSE of the estimation based on the entire forest in this study is 22.61 t·hm⁻², while the RMSE in Yang’s study is 30.20 t·hm⁻². From the comparison results, it can be seen that the estimated values in this study fit better with the true values across regions with different biomass levels, especially in areas with higher biomass, where Yang’s study exhibited significant underestimation. In contrast, this study more accurately reflects the actual biomass conditions. This is because the spatial resolution of the dataset used in this study exceeds 30 m, and a larger-scale measured dataset has been obtained in the study area. These improvements have enhanced the reliability of biomass estimation at the extremes. The comparison of AGB estimation results between this study and Yang’s study is presented in Figure 10, while the specific errors are shown in Table 11.

When multi-source remote-sensing data and ResNet were used to estimate whole-forest biomass, R² was 0.71, and RMSE was 20.95 t·hm⁻². In the vertical zonality modeling method, the estimation accuracy of the subalpine open-forest belt, which has a slightly lower sample number, is lower than that of the whole forest. Conversely, the estimation accuracies of the other two vertical belts are higher than that of the whole forest. In the case of stand-age modeling, the R² of the model reaches up to 0.83. According to Figure 6c,f,i, Figure 7c,f,i and Figure 8c, when the biomass is less than 100 t·hm⁻², the overestimation of the whole-forest biomass is the most significant, and the deviation value significantly exceeds the estimation results of vertical zonality and stand-age modeling. This overestimation is not accidental. When the biomass reaches 170 t·hm⁻² and above, the underestimation issue emerges again in the whole forest. Taking the biomass estimation approach integrating multi-source remote-sensing data and the ResNet model as an example, among all the estimation results, when the error between the estimated value and the actual biomass exceeds 30 t·hm⁻², the error proportion under different modeling methods exhibits distinct differences. When estimating the forest as a whole, the estimated values with an error exceeding 30 t·hm⁻² account for 23.09% of the total estimated values. When the vertical zonality modeling method is applied to estimate biomass, the error ratio is significantly reduced to 5.45%. In the case of stand-age modeling, the proportion of estimates with an error greater than 30 t·hm⁻² is the lowest, merely 4.72%. The climatic and other conditions in distinct vertical zones vary substantially, with a profound impact on vegetation growth and biomass distribution. The biomass characteristics of trees in different stand ages also differ significantly because of their disparate growth states. Dividing forests based on vertical zonation or age class allows for more precise identification of these differences. As a result, the accuracy of estimations is notably enhanced, lending support to more accurate biomass evaluations.

5. Conclusions

In the biomass estimation study of Picea schrenkiana forests in the West Tianshan Mountains, this study integrated GF-1, GF-6, and GF-3 PolSAR data with dem data, extracted corresponding feature factors, and applied the ResNet method to conduct the estimation work. Eventually, high accuracy in biomass estimation was achieved. It is well-established that along altitude gradients, over short distances, there are systematic alterations in temperature and precipitation. Additionally, as forests age, biomass continues to build up. Based on these principles, this study further classified Picea schrenkiana forests according to two methods: Vertical Zonality Differentiation without considering stand age and Stand-Age Classification without considering vertical zonality. The results showed that the model accuracy was remarkably improved.

Starting from the middle and low mountain forest-steppe belt, as the altitude continues to increase, temperature and precipitation conditions become highly conducive to AGB accumulation in Picea schrenkiana forests. However, if the altitude exceeds a certain critical value, the temperature will drop significantly, and this low-temperature environment will no longer be suitable for AGB accumulation in Picea schrenkiana forests [50]. Conversely, below the middle and low mountain forest-steppe belt, due to intensive human activities, it is also not conducive to the large-scale growth of Picea schrenkiana forest. This further explains why Picea schrenkiana forest is mainly distributed in the area 1700–2700 m above sea level.

Due to limitations in data availability, in-depth studies on the long-time-series AGB of Picea schrenkiana forests remain scarce. In follow-up research, remote-sensing data from the Landsat series or other sources can be integrated at 10−20-year intervals, combined with environmental factors such as precipitation and temperature, to systematically analyze the dynamic changes in Picea schrenkiana forest AGB. Furthermore, the influencing factors considered in the current model construction are not comprehensive enough. In future studies, additional key stand characteristic factors, such as slope position and mean DBH, can be incorporated to continuously optimize the AGB estimation model for Picea schrenkiana forests.

Author Contributions

Conceptualization, G.Z. and D.C.; methodology, M.P.; software, Q.Z.; investigation, J.Z.; resources, H.Z.; data curation, Y.H.; writing—original draft preparation, G.Z.; writing—review and editing, D.C. and H.L.; visualization, H.Z.; supervision, D.C.; project administration, J.F. All authors have read and agreed to the published version of the manuscript.

Data Availability Statement

All data, models, or code generated or used during the study are available from the author by request ([email protected]).

Acknowledgments

Thanks to the data support provided by the websites of various data sources in the article. I would like to thank Donghua Chen, Hu Li et al. for their guidance on the paper, as well as their support for the projects related to this paper.

Conflicts of Interest

The authors declare no conflicts of interest.

Footnotes

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

Figures and Tables

Figure 1. Geographic location map of the study area. (a) Position of the study area within China. (b) Position of the study area within Xinjiang Province. (c) Elevation map of the study area, along with the forest sample locations.

Figure 2. Topographic feature map of the study area. (a) altitude map of the study area. (b) vertical zonality division map of the study area. (c) slope division map of the study area. (d) aspect division map of the study area.

Figure 3. Classification map of stand ages of Picea schrenkiana forests.

Figure 4. The framework of the Back Propagation Neural Network model.

Figure 5. The framework of the Residual Network model. k represents the convolution kernel, s represents the stride, and p represents the padding.

Figure 6. The accuracy of each model using multi-source data in the case of vertical zonality.

Figure 7. The accuracy of each model using multi-source data in the case of stand ages.

Figure 8. Accuracy of each model using multi-source data in the case of whole forest.

Figure 9. AGB estimation results under three different modeling methods. (a) shows the output map of estimated AGB for the whole forest. (b) based on the classification of vertical zonality, shows the output map of estimated AGB. (c) based on the classification of stand ages, shows the output map of estimated AGB.

Figure 10. The comparison of AGB estimation results between this study and Yang’s study. (a) the results of this experimental study; (b) The results of Yang’s study.

Table 1

Classification criteria for vertical zonality.

Vertical Zonality of Picea schrenkiana Forest	DEM
Below the middle and low mountain forest-steppe belt	Below 1500 m
Middle and low mountain forest-steppe belt	1500–1700 m
Middle mountain forest-meadow belt	1700–2250 m
Upper middle mountain forest-meadow belt	2250–2550 m
Subalpine open-forest belt	2550–2700 m
Above the subalpine open-forest belt	Above 2700 m

Table 2

Classification criteria for stand age of Picea schrenkiana forests.

Stand Age of Picea schrenkiana Forest	Age (Unit: Year)
Young forest	Below 60
Middle-aged forest	61–100
Near-mature forest	101–120
Mature forest	121–160
Over-mature forest	Above 161

Table 3

Commonly used vegetation index.

Source	Type	Variable	Equations
Optical data	Vegetation Index	NDVI	$NDVI={\displaystyle \frac{(NIR-R)}{(NIR+R)}}$
		RVI	$RVI={\displaystyle \frac{NIR}{R}}$
		GNDVI	$GNDVI={\displaystyle \frac{NIR-G}{NIR+G}}$
		SAVI	$SAVI=({\displaystyle \frac{NIR-R}{NIR+R+L}})\times (1+L)$
		DVI	$DVI=NIR-R$
		IPVI	$IPVI={\displaystyle \frac{NIR}{NIR+R}}$
		EVI	$EVI=2.5\times {\displaystyle \frac{(NIR-R)}{(NIR+6R-7.5B+1)}}$
		NDMVI	$NDMVI={\displaystyle \frac{(NIR-R)+(R_{mim}-NI{R}_{mim})}{(NIR+R)+(R_{mim}+NI{R}_{mim})}}$

Note: In the formula, $NIR$ denotes the reflectance value of the near-infrared band, $R$ denotes the reflectance value of the red band, $B$ denotes the reflectance value of the blue wave segment, and $G$ denotes the reflectance value of the green band. Furthermore, $L$ is the soil adjustment coefficient, which is 0.5 in this study. In NDMVI, the minimum reflectance values of the red band and near-infrared band of the image in the study area were extracted and assigned to $R_{mim}$ and $NI{R}_{mim}$, respectively.

Table 4

Texture feature calculation formulas based on the gray-level co-occurrence matrix.

Source	Type	Variable	Equations
Optical data	GLCM Texture	Mean (MEAN)	$MEAN={\displaystyle \sum _{i,j=0}^{N-1}i}p(i,j)$
		Variance (VAR)	$VAR={\displaystyle \sum _{i,j=0}^{N-1}p}(i,j){(i-{\mu }_i)}^2$
		Homogeneity (HOM)	$HOM={\displaystyle \sum _{i,j=0}^{N-1}\frac{p(i,j)}{1+{(i-j)}^2}}$
		Contrast (CON)	$CON={\displaystyle \sum _{i,j=0}^{N-1}p}(i,j){(i-j)}^2$
		Dissimilarity (DIS)	$DIS={\displaystyle \sum _{i,j=0}^{N-1}p}(i,j)|i-j|$
		Entropy (ENT)	$ENT={\displaystyle \sum _{i,j=0}^{N-1}p(i,j)}(-\log [p(i,j)])$
		Second moment (ASM)	$ASM={\displaystyle \sum _{i,j=0}^{N-1}p}{(i,j)}^2$
		Correlation (COR)	$COR={\displaystyle \sum _{i,j=0}^{N-1}\frac{(i-u_i)(j-u_j)}{{\phi }_i{\phi }_j}}p(i,j)$

Note: In the formula, $i$ and $j$ denote the number of rows and columns of the matrix, respectively. $N$ denotes the number of pixels. $p(i,j)$ represents the probability of the simultaneous occurrence of two gray values corresponding to row $i$ and column $j$. ${\mu }_i$ and ${\mu }_{\mathrm{j}}$ represent the mean values of rows and columns, respectively. ${\phi }_i$ and ${\phi }_j$ denote the variance of rows and columns.

Table 5

Related parameters of polarization decomposition.

Source	Type	Variable	Description
SAR data	Polarization decomposition	Freeman decomposition	Double-bounce scattering (Dbl), Volume scattering (Vol), Surface scattering (Odd)
		Pauli decomposition	Double-bounce scattering (Dbl), Volume scattering (Vol), Surface scattering (Odd)
		H/A/Alpha decomposition	Entropy, Anisotropy, Alpha, Lambda
		Yamaguchi decomposition	Double-bounce scattering (Dbl), Volume scattering (Vol), Surface scattering (Odd),Helix scattering (Hlx)
		Krogager decomposition	Spherical scattering (KS), Dihedral scattering (KD), Helicoid angles scattering (KH)

Table 6

Correlation between factors and AGB in different vertical zonality.

	Middle Mountain Forest-Meadow Belt		Upper Middle Mountain Forest-Meadow Belt		Subalpine Open-Forest Belt
Optical data	Band3	−0.372 **	Band2	−0.473 **	Band3	0.336 **
	Band2	−0.345 **	Band3	−0.453 **	Band4	−0.321 **
	MEAN_3	0.432 **	MEAN_3	0.504 **	Band1	0.291 *
	COR_5	0.217 **	COR_5	0.184 **	DVI	−0.374 **
	DVI	−0.251 **	DVI	−0.353 **
			GNDVI	−0.106 **
SAR data	HV	0.363 **	HH	−0.378 **	HH	0.406 **
	VH	0.339 **	HV	−0.344 **	HV	0.383 **
	Pauli_Vol	0.353 **	Freeman_Vol	−0.342 **	Pauli_Vol	0.351 **
	Freeman_Rvi	0.338 **	Pauli_Vol	−0.337 **
			Freeman_Rvi	0.135 **
Topographic Factor			SLOPE	0.327 **

Note: * indicates significant correlation at the 0.05 level (two-tailed). ** indicates significant correlation at the 0.01 level (two-tailed). MEAN_3 represents the mean value in a 3 × 3 window. COR_5 represents the correlation in a 3 × 3 window.

Table 7

Correlation between factors and AGB in different stand ages.

	Near-Mature Forest		Mature Forest		Over-Mature Forest
Optical data	Band2	−0.571 **	Band3	−0.500 **	Band3	−0.412 **
	MEAN_3	0.568 **	Band2	−0.492 **	Band2	−0.322 **
	COR_11	0.369 **	MEAN_5	0.515 **	DVI	0.399 **
	RVI	0.346 **	COR_5	0.194 **
	DVI	0.340 **	COR_3	−0.143 **
	NDVI	0.312 **	DVI	−0.233 **
SAR data	VH	−0.424 **	HH	−0.431 **	VH	−0.423 **
	HV	−0.413 **	VH	−0.358 **	Freeman_Vol	−0.390 **
	Pauli_Odd	−0.403 **	Freeman_Rvi	0.266 **	Freeman_Rvi	−0.273 **
	Freeman_Vol	−0.397 **
	Yamaguchi_Odd	−0.369 **
Topographic Factor	SLOPE	0.325 **	SLOPE	0.427 **	DEM	0.250 **
	DEM	−0.178 **

Note: ** indicates significant correlation at the 0.01 level (two-tailed). MEAN_3 represents the mean value in a 3 × 3 window. COR_3 represents the contrast in a 3 × 3 window. MEAN_5 represents the mean value in a 5 × 5 window. COR_5 represents the correlation in a 5 × 5 window. COR_11 represents the correlation in an 11 × 11 window.

Table 8

The Accuracy of Each Model under Different Vertical Zonality and Different Data Sources.

		Optical Data + Topographic Factor		SAR Data + Topographic Factor		Optical Data + SAR Data + Topographic Factor
		R2	RMSE	R2	RMSE	R2	RMSE
XGBoost	Middle mountain forest-meadow belt	0.52	23.40	0.46	24.87	0.62	20.69
	Upper middle mountain forest-meadow belt	0.57	22.96	0.52	26.06	0.63	21.30
	Subalpine open-forest belt	0.51	24.56	0.54	18.66	0.57	13.40
BPNN	Middle mountain forest-meadow belt	0.67	20.24	0.67	18.02	0.72	16.87
	Upper middle mountain forest-meadow belt	0.66	19.65	0.65	22.21	0.71	19.21
	Subalpine open-forest belt	0.62	16.85	0.57	16.50	0.66	15.15
ResNet	Middle mountain forest-meadow belt	0.74	17.31	0.70	16.55	0.78	15.27
	Upper middle mountain forest-meadow belt	0.74	17.82	0.67	21.91	0.79	15.62
	Subalpine open-forest belt	0.63	22.19	0.67	12.31	0.69	12.63

Table 9

The Accuracy of Each Model under Different Stand Ages and Different Data Sources.

		Optical Data + Topographic Factor		SAR Data + Topographic Factor		Optical Data + SAR Data + Topographic Factor
		R2	RMSE	R2	RMSE	R2	RMSE
XGBoost	Near-mature forest	0.52	14.21	0.48	14.86	0.60	13.40
	Mature forest	0.60	16.47	0.54	17.61	0.65	15.93
	Over-mature forest	0.51	29.60	0.49	23.79	0.60	28.33
BPNN	Near-mature forest	0.65	12.87	0.64	13.05	0.69	12.60
	Mature forest	0.63	15.37	0.63	15.38	0.71	14.93
	Over-mature forest	0.64	21.78	0.60	22.85	0.67	25.59
ResNet	Near-mature forest	0.64	12.43	0.72	10.84	0.77	10.06
	Mature forest	0.67	14.91	0.68	14.64	0.83	10.62
	Over-mature forest	0.74	21.58	0.72	22.44	0.77	22.68

Table 10

The Accuracy of different models in estimating the whole forest using multi-source data.

		XGBoost	BPNN	ResNet
Whole forest	R2	0.55	0.66	0.71
	RMSE	27.71	23.04	20.95

Table 11

The errors of the whole-forest estimation results in this study and the AGB estimation results in Yang’s study when compared with the true value.

Number	True Value	Whole Forest Estimate	Error	Yang’s AGB	Error
1	62.3534	93.81768	−31.4643	117.565	−55.2116
2	72.1016	84.79638	−12.6948	123.5967	−51.4951
3	81.8498	102.4973	−20.6475	117.2498	−35.4
4	86.4918	96.70109	−10.2093	109.858	−23.3662
5	96.24	110.9799	−14.7399	123.005	−26.765
6	101.3462	87.37556	13.97064	106.6862	−5.33999
7	105.9882	99.72482	6.263377	118.8634	−12.8752
8	111.0944	108.8939	2.200533	126.8149	−15.7205
9	115.7364	111.9555	3.780879	143.0097	−27.2733
10	120.8426	128.0583	−7.21566	130.8246	−9.98197
11	125.4846	114.2628	11.22175	117.4728	8.011774
12	130.5908	91.12645	39.46435	105.9736	24.61725
13	133.376	138.1766	−4.80064	135.807	−2.43097
14	136.1612	166.4629	−30.3017	146.5908	−10.4296
15	140.339	134.6607	5.67831	144.5293	−4.19033
16	144.981	131.7099	13.27107	126.5128	18.4682
17	154.7292	160.7999	−6.0707	138.5166	16.21257
18	159.8354	173.726	−13.8906	135.9558	23.87965
19	164.4774	176.8097	−12.3323	139.8075	24.66987
20	169.5836	131.8373	37.74632	145.0627	24.52095
21	183.9738	150.7507	33.22307	134.5283	49.44547
22	189.08	171.5312	17.54883	154.0613	35.01869
23	193.722	149.9356	43.78642	151.4542	42.26779
24	208.5764	162.8467	45.72974	147.4223	61.15406

References

1. Mo, L.; Zohner, C.M.; Reich, P.B.; Liang, J.; de Miguel, S.; Nabuurs, G.J.; Renner, S.S.; van den Hoogen, J.; Araza, A.; Herold, M. et al. Integrated global assessment of the natural forest carbon potential. Nature; 2023; 624, pp. 92-101. [DOI: https://dx.doi.org/10.1038/s41586-023-06723-z] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37957399]

2. Li, H.; Ci, L.; Fang, J.; Chen, D.; Liu, Y. Dynamic Monitoring of Picea schrenkiana Forest Biomass in West Tianshan Mountain Region of Xinjiang. Sci. Silv. Sin.; 2008; 44, pp. 14-19.

3. Walker, W.S.; Gorelik, S.R.; Cook-Patton, S.C.; Susan, C.; Baccini, A.; Farina, M.K.; Solvik, K.K.; Ellis, P.W.; Sanderman, J.; Houghton, R.A. et al. The global potential for increased storage of carbon on land. Proc. Natl. Acad. Sci. USA; 2022; 119, e2023803119. [DOI: https://dx.doi.org/10.1073/pnas.2111312119] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35639697]

4. Liu, Q.; Yang, L.; Liu, Q.; Li, J. Review of Forest Above-Ground Biomass Inversion Methods Based on Remote Sensing Technology. J. Remote Sens.; 2015; 19, pp. 62-74.

5. Cao, G. Application of remote sensing technology in forest resources investigation. Remote Sens.; 2022; 11, pp. 46-49. [DOI: https://dx.doi.org/10.18282/rs.v9i2.1373]

6. Gou, R.; Chen, J.; Duan, G.; Yang, R.; Bu, Y.; Zhao, J.; Zhao, P. Inversion of aboveground biomass of Pinus tabuliformis plantations based on GF-2 data. J. Appl. Ecol.; 2019; 30, pp. 4031-4040.

7. Naik, P.; Dalponte, M.; Bruzzone, L. Prediction of forest aboveground biomass using multitemporal multispectral remote sensing data. Remote Sens.; 2021; 13, 1282. [DOI: https://dx.doi.org/10.3390/rs13071282]

8. Li, H.; Chen, D.; Liu, S.; Liu, Y.; Huang, W.; Zhang, N.; Yang, F.; Li, X.; Xing, F.; Liu, C. et al. Research on Forest Resources and Environmental Monitoring in the Western Tianshan Mountains Based on Domestic Land Satellite Data; 1st ed. Science Press: Beijing, China, 2022; pp. 74-145.

9. Tang, Z.; Xia, X.; Huang, Y.; Lu, Y.; Guo, Z. Estimation of National Forest Aboveground Biomass from Multi-Source Remotely Sensed Dataset with Machine Learning Algorithms in China. Remote Sens.; 2022; 14, 5487. [DOI: https://dx.doi.org/10.3390/rs14215487]

10. Li, L.; Chen, E.; Li, Z.; Feng, Q.; Zhao, L. A Review on Forest Height and Aboveground Biomass Estimation Based on Synthetic Aperture Radar. Remote Sens. Technol. Appl.; 2016; 31, pp. 625-633. [DOI: https://dx.doi.org/10.3390/rs8080625]

11. Zhang, L.; Yin, X.; Wang, Y.; Chen, J. Aboveground Biomass Mapping in Semi-Arid Forests by Integrating Airborne LiDAR with Sentinel-1 and Sentinel-2 Time-Series Data. Remote Sens.; 2024; 16, 3241. [DOI: https://dx.doi.org/10.3390/rs16173241]

12. Cartus, O.; Santoro, M. Exploring combinations of multi-temporal and multi-frequency radar backscatter observations to estimate above-ground biomass of tropical forest. Remote Sens. Environ.; 2019; 232, 111313. [DOI: https://dx.doi.org/10.1016/j.rse.2019.111313]

13. Wei, J.; Fan, W.; Yu, Y.; Mao, X. Polarimetric Decomposition Parameters for Artificial Forest Canopy Biomass Estimation Using GF-3 Fully Polarimetric SAR Data. Sci. Silv. Sin.; 2020; 56, pp. 174-183.

14. Zhang, Y.; Wang, N.; Wang, Y.; Li, M. A New Strategy for Improving the Accuracy of Forest Aboveground Biomass Estimates in an Alpine Region Based on Multi-Source Remote Sensing. GISci. Remote Sens.; 2023; 60, 2163574. [DOI: https://dx.doi.org/10.1080/15481603.2022.2163574]

15. Forkuor, G.; Zoungrana, J.B.B.; Dimobe, K.; Ouattara, B.; Vadrevu, K.P.; Tondoh, J.E. Above-ground biomass mapping in West African dryland forest using Sentinel-1 and 2 datasets—A case study. Remote Sens. Environ.; 2020; 236, 111496. [DOI: https://dx.doi.org/10.1016/j.rse.2019.111496]

16. Wang, P.; Zhang, H.; Tian, X.; Zhang, J.; Zhu, Y. Estimation of Forest Stem Volume of Natural Forests Based on the Optimization of Active and Passive Remote Sensing Features. Remote Sens. Technol. Appl.; 2022; 37, pp. 672-680.

17. Liu, Y.; Gong, W.; Xing, Y.; Hu, X.; Gong, J. Estimation of the Forest Stand Mean Height and Aboveground Biomass in Northeast China Using SAR Sentinel-1B, Multispectral Sentinel-2A, and DEM Imagery. ISPRS J. Photogramm. Remote Sens.; 2019; 151, pp. 277-289. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2019.03.016]

18. Jucker, T.; Bongalov, B.; Burslem, D.F.R.P.; Nilus, R.; Dalponte, M.; Lewis, S.L.; Phillips, O.L.; Qie, L.; Coomes, D.A. Topography Shapes the Structure, Composition and Function of Tropical Forest Landscapes. Ecol. Lett.; 2018; 21, pp. 989-1000. [DOI: https://dx.doi.org/10.1111/ele.12964]

19. Ni, J. Forest productivity of the Altay and Tianshan Mountains in the dryland, northwestern China. For. Ecol. Manag.; 2004; 202, pp. 13-22. [DOI: https://dx.doi.org/10.1016/j.foreco.2004.06.033]

20. Wu, Y.; Li, F.; Zhang, J.; Liu, Y.; Li, H.; Zhou, B.; Shen, B.; Hou, L.; Xu, D.; Ding, L. et al. Spatial and Temporal Patterns of Above- and Below-Ground Biomass Over the Tibet Plateau Grasslands and Their Sensitivity to Climate Change. Sci. Total Environ.; 2024; 919, 170900. [DOI: https://dx.doi.org/10.1016/j.scitotenv.2024.170900]

21. Wang, Y.; Liu, Y.; Chen, P.; Song, J.; Fu, B. Interannual Precipitation Variability Dominates the Growth of Alpine Grassland Above-ground Biomass at High Elevations on the Tibetan Plateau. Sci. Total Environ.; 2024; 931, 172745. [DOI: https://dx.doi.org/10.1016/j.scitotenv.2024.172745]

22. Wu, X.; Jiao, L.; Du, D.; Xue, R.; Wei, M.; Zhang, P. Elevation Response of Above-ground Net Primary Productivity for Picea crassifolia to Climate Change in Qilian Mountains of Northwest China Based on Tree Rings. J. Geogr. Sci.; 2024; 34, pp. 146-164. [DOI: https://dx.doi.org/10.1007/s11442-024-2199-x]

23. Maza, B.; Rodes-Blanco, M.; Rojas, E. Aboveground Biomass Along an Elevation Gradient in an Evergreen Andean–Amazonian Forest in Ecuador. Front. For. Glob. Change; 2022; 5, 738585. [DOI: https://dx.doi.org/10.3389/ffgc.2022.738585]

24. Chan, C.K.; Gomez, C.A.; Kothikar, A.; Baiz-Villafranca, P.M. Satellite-Based Carbon Estimation in Scotland: AGB and SOC. Land; 2023; 12, 818. [DOI: https://dx.doi.org/10.3390/land12040818]

25. Ye, Q.; Yu, S.; Liu, J.; Zhao, Q.; Zhao, Z. Aboveground Biomass Estimation of Black Locust Planted Forests with Aspect Variable Using Machine Learning Regression Algorithms. Ecol. Indic.; 2021; 129, 107948. [DOI: https://dx.doi.org/10.1016/j.ecolind.2021.107948]

26. Han, H.; Wan, R.; Li, B. Estimating Forest Aboveground Biomass Using Gaofen-1 Images, Sentinel-1 Images, and Machine Learning Algorithms: A Case Study of the Dabie Mountain Region, China. Remote Sens.; 2022; 14, 176. [DOI: https://dx.doi.org/10.3390/rs14010176]

27. Ali, A.; Yan, E.R.; Chen, H.Y.H.; Chang, S.X.; Zhao, Y.T.; Yang, X.D.; Xu, M.S. Stand Structural Diversity Rather Than Species Diversity Enhances Aboveground Carbon Storage in Secondary Subtropical Forests in Eastern China. Biogeosciences; 2016; 13, pp. 4627-4635. [DOI: https://dx.doi.org/10.5194/bg-13-4627-2016]

28. Huang, X.; Chen, Y.C.; Tan, H.R.; Zhang, Y.; Yu, S.H.; Chen, X.Y.; Yu, K.Y.; Liu, J. Extraction of the Spatial Structure of Chinese Fir Plantation Stands Based on Unmanned Aerial Vehicle and Its Effect on AGB. For. Ecol. Manag.; 2024; 558, 121800. [DOI: https://dx.doi.org/10.1016/j.foreco.2024.121800]

29. Lee, Y.J.; Park, K.E.; Lee, H.I.; Lee, C.B. Stand age-driven tree size variation and stand type regulate aboveground biomass in alpine-subalpine forests, South Korea. Sci. Total Environ.; 2024; 915, 170063. [DOI: https://dx.doi.org/10.1016/j.scitotenv.2024.170063]

30. Wu, Y.; Ou, G.; Lu, T.; Huang, T.; Zhang, X.; Liu, Z.; Yu, Z.; Guo, B.; Wang, E.; Feng, Z. et al. Improving Aboveground Biomass Estimation in Lowland Tropical Forests across Aspect and Age Stratification: A Case Study in Xishuangbanna. Remote Sens.; 2024; 16, 1276. [DOI: https://dx.doi.org/10.3390/rs16071276]

31. Li, D.; Wang, M.; Jiang, J. China’s High-Resolution Optical Remote Sensing Satellites and Their Mapping Applications. Geo-Spat. Inf. Sci.; 2020; 24, pp. 85-94. [DOI: https://dx.doi.org/10.1080/10095020.2020.1838957]

32. Zeng, P.; Zhang, W.F.; Li, Y.; Shi, J.M.; Wang, Z.H. Forest Total and Component Above-Ground Biomass (AGB) Estimation through C- and L-band Polarimetric SAR Data. Forests; 2022; 13, 442. [DOI: https://dx.doi.org/10.3390/f13030442]

33. Lee, J.S.; Wen, J.H.; Ainsworth, T.L.; Chen, K.S.; Chen, A.J. Improved Sigma Filter for Speckle Filtering of SAR Imagery. IEEE Trans. Geosci. Remote Sens.; 2009; 47, pp. 202-213.

34. Rosenqvist, A.; Shimada, M.; Ito, N.; Watanabe, M. ALOS PALSAR: A Pathfinder Mission for Global-Scale Monitoring of the Environment. IEEE Trans. Geosci. Remote Sens.; 2007; 45, pp. 3307-3316. [DOI: https://dx.doi.org/10.1109/TGRS.2007.901027]

35. Zhang, X.; Zhang, Y.; Chen, W.; Zheng, J.; Chen, F.; Chen, K.; Mo, G. Types of clear-cut areas and their regeneration in Tianshan Snowy Mountains spruce forests. Sci. Silvae Sin.; 1964; 9, pp. 167-183.

36. Fang, J.; Liu, G.; Xu, S. Biomass and Net Production of Forest Vegetation in China. Acta Ecol. Sin.; 1996; 16, pp. 497-508.

37. Liu, J.F.; Chen, R.S.; Qin, W.W.; Yang, Y. Study on the vertical distribution of precipitation in mountainous regions using TRMM data. Adv. Water Sci.; 2011; 22, pp. 447-454.

38. Xu, D.; Wang, C.; Chen, J.; Shen, M.; Shen, B.; Yan, R.; Li, Z.; Karnieli, A.; Chen, J.; Yan, Y. et al. The superiority of the normalized difference phenology index (NDPI) for estimating grassland aboveground fresh biomass. Remote Sens. Environ.; 2021; 264, 112578. [DOI: https://dx.doi.org/10.1016/j.rse.2021.112578]

39. Chen, L.Y.; He, A.Q.; Xu, Z.H.; Li, B.; Zhang, H.F.; Li, G.T.; Guo, X.Y.; Li, Z.L. Mapping aboveground biomass of Moso bamboo (Phyllostachys pubescens) forests under Pantana phyllostachysae Chao-induced stress using Sentinel-2 imagery. Ecol. Indic.; 2024; 158, 111564. [DOI: https://dx.doi.org/10.1016/j.ecolind.2024.111564]

40. Fu, B.; Wang, Y.; Campbell, A.; Li, Y.; Zhang, B.; Yin, S.; Xing, Z.; Jin, X. Comparison of object-based and pixel-based Random Forest algorithm for wetland vegetation mapping using high spatial resolution GF-1 and SAR data. Ecol. Indic.; 2017; 73, pp. 105-117. [DOI: https://dx.doi.org/10.1016/j.ecolind.2016.09.029]

41. Bu, X.; Dong, S.; Mi, W.; Li, F. Spatial-temporal change of carbon storage and sink of wetland ecosystem in arid regions, Ningxia Plain. Atmos. Environ.; 2019; 204, pp. 89-101.

42. Li, D.; Wang, C.; Hu, Y.; Liu, S. General Review on Remote Sensing-Based Biomass Estimation. Geom. Inf. Sci. Wuhan Univ.; 2012; 37, pp. 631-635.

43. Cloude, S.; Pottier, E. A review of target decomposition theorems in radar polarimetry. IEEE Trans. Geosci. Remote Sens.; 1996; 34, pp. 498-518. [DOI: https://dx.doi.org/10.1109/36.485127]

44. Wang, Q.; Zeng, Q.; Liao, J. Extraction and Application of Polarimetric Characteristic Parameters Based on Polarimetric Decomposition. Remote Sens. Nat. Resour.; 2012; 3, pp. 103-110.

45. Jia, Z.; Zhang, Z.; Cheng, Y.; Buhebaoyin, B.; Borjigin, S.; Quan, Z. Grassland biomass spatiotemporal patterns and response to climate change in eastern Inner Mongolia based on XGBoost model estimates. Ecol. Indic.; 2024; 158, 111554. [DOI: https://dx.doi.org/10.1016/j.ecolind.2024.111554]

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

47. Liu, T.; Yang, T.; Zhu, S.; Mou, N.; Zhang, W.; Wu, W.; Zhao, Y.; Yao, Z.; Sun, J.; Chen, C. et al. Estimation of wheat biomass based on phenological identification and spectral response. Comput. Electron. Agric.; 2024; 222, 109076. [DOI: https://dx.doi.org/10.1016/j.compag.2024.109076]

48. Attarchi, S.; Gloaguen, R. Improving the Estimation of Above Ground Biomass Using Dual Polarimetric PALSAR and ETM+ Data in the Hyrcanian Mountain Forest (Iran). Remote Sens.; 2014; 6, pp. 3693-3715. [DOI: https://dx.doi.org/10.3390/rs6053693]

49. Yang, Q.; Niu, C.; Liu, X.; Feng, Y.; Ma, Q.; Wang, X.; Guo, Q. Mapping high-resolution forest aboveground biomass of China using multisource remote sensing data. GISci. Remote Sens.; 2023; 60, 2203303. [DOI: https://dx.doi.org/10.1080/15481603.2023.2203303]

50. Tian, S.C.; Zhao, S.C.; Zheng, X.J.; Wang, Y.G.; Li, Y. Water Source of Spruce (Picea schrenkiana) at Different Altitudes in the Tianshan Mountains during the Growing Season. Arid Zone Res.; 2023; 40, pp. 436-444.