Remote Sensing Estimation and Spatiotemporal Variation Characteristics of Forest Aboveground Carbon Storage in Qianjiangyuan Baishanzu National Park

Highlights

What are the main findings?

• The random forest model achieved optimal inversion performance, yielding a test set R² of 0.954, RMSE of 4.653 Mg/ha, and rRMSE of 10.087%.
• From 2009 to 2019, AGC density rose from 37.64 to 66.00 Mg/ha, exhibiting high accumulation in central, southern, and southeastern regions.

What are the implications of the main findings?

• Coupling multi-temporal Landsat data with the Boruta algorithm isolates structurally sensitive predictors, suppressing spatial modeling uncertainty in subtropical forests.
• Quantifying heterogeneous decadal carbon dynamics delivers a precise, data-driven baseline to guide differentiated carbon enhancement and functional zone management within national parks.

Abstract

National forest parks play an important role in maintaining the integrity of ecosystems, the sustainability of biodiversity, and the improvement of ecological service functions. Aboveground carbon storage (AGC) is an important indicator for evaluating forest ecosystem functions. Qianjiangyuan–Baishanzu National Park, located in the southern part of Lishui City, serves as a typical representative of the mid-subtropical evergreen broad-leaved forest ecosystem. It is characterized by high biodiversity and serves as a concentration area for rare and endangered species. Therefore, assessing the spatiotemporal dynamics of forest AGC in the typical subtropical forests of Qianjiangyuan–Baishanzu National Park is of great scientific significance. Taking Qianjiangyuan–Baishanzu National Park as a case study, this research utilized three periods of Landsat satellite remote sensing data (2009, 2014, and 2019) alongside contemporaneous ground-based AGC survey samples. Feature variables were extracted and subsequently screened using the Boruta algorithm. There were three algorithms, including least squares (LS), support vector regression (SVR), and random forest (RF), constructed to estimate forest AGC. The optimal AGC estimation model was selected, and the spatiotemporal variation characteristics of forest AGC within the national park were systematically analyzed. Research has shown that (1) texture features are important parameters for constructing forest AGC estimation models, accounting for up to 71%, with the 7 × 7 window having the greatest impact. (2) All three models can achieve high accuracy in estimating the forest AGC and its spatial distribution in Qianjiangyuan Baishanzu National Park. Among them, the RF model has the highest overall accuracy in estimating forest AGC, with a training set R² of 0.938, RMSE of 5.550 Mg/ha, rRMSE of 12.517%, a test set R² of 0.954, RMSE of 4.653 Mg/ha, and rRMSE of 10.087%. (3) The distribution of forest AGC in Qianjiangyuan Baishanzu National Park shows significant spatial heterogeneity, with higher carbon storage in the central, southern, and southeastern regions, while the northern region has relatively lower carbon storage. From 2009 to 2019, the forest AGC in the Qianjiangyuan–Baishanzu National Park exhibited a steady annual increase, with AGC density rising from 37.64 Mg/ha to 66 Mg/ha and total AGC stock increasing from 2.16 Tg C to 4.36 Tg C. These findings provide precise data support and a scientific basis for decision-making regarding differentiated ecological carbon enhancement and functional zone management within the national park.

1. Introduction

National forest parks are natural protected areas designated through legal and policy frameworks, which effectively safeguard ecosystem integrity and the sustainability of biodiversity while enhancing ecosystem service functions [1,2]. They are not only key regions for biodiversity conservation but also an essential component of ecological civilization construction. Therefore, strengthening forest monitoring and management within national parks is of profound significance for the sustainable development of ecosystems [3]. With the intensification of global climate change, forest ecosystems play a crucial role in climate change mitigation and in achieving carbon neutrality targets. Forest aboveground carbon storage (AGC) constitutes a major component of the forest carbon pool. Accurate estimation of forest AGC not only enables the assessment of forest carbon sink capacity but also provides a scientific basis for forest resource management and conservation, as well as data support for governmental forest development planning and environmental protection policies [4]. The Qianjiangyuan–Baishanzu National Park is characterized by high forest coverage and relatively high forest quality, representing a typical subtropical forest ecosystem in Zhejiang Province. It has made significant contributions to regional ecological balance maintenance and conservation [5]. Therefore, accurately quantifying and monitoring forest AGC in the Qianjiangyuan–Baishanzu National Park and analyzing its spatiotemporal variation are of great importance for evaluating forest carbon sink functions, biodiversity conservation, and ecosystem management within the park.

At present, forest AGC estimation methods can be broadly categorized into three types: field plot inventory methods, process-based model simulation methods, and remote sensing-based estimation methods [6,7]. The field plot inventory method involves measuring tree height, diameter at breast height (DBH), crown width, and other parameters for all trees within sample plots [8,9]. Individual-tree AGC is then calculated using allometric equations and aggregated to obtain total AGC at the plot level. Although this method provides relatively accurate estimates of forest AGC, it is labor-intensive, time-consuming, and costly [10]. Moreover, plot-based measurements are limited in representing regional-scale spatial patterns and may cause certain disturbances to forest ecosystems, making them unsuitable for large-area and large-scale AGC estimation. Process-based model simulation methods estimate forest AGC by simulating various physiological and ecological processes within ecosystems [11]. However, these methods often face difficulties in acquiring input parameters, and their applicability is constrained by ecosystem types and environmental conditions, making simulation results highly sensitive to uncertainties [12]. In recent years, remote sensing techniques have been increasingly applied to forest AGC estimation. Remote sensing data offer advantages such as wide spatial coverage, relatively high spatial resolution, and convenient data acquisition [13]. Consequently, remote sensing-based approaches have gradually become the dominant methods for estimating forest AGC.

Nevertheless, remote sensing data cannot directly reflect changes in forest AGC [14]. Therefore, it is necessary to integrate remotely sensed surface information with field observation data and modeling algorithms to establish quantitative relationships between remote sensing features and AGC, enabling spatiotemporal estimation of forest AGC [15,16]. Among various AGC estimation models, Xu employed the least squares method combined with Landsat data to produce preliminary global forest above-ground biomass (AGB) estimates [17]. Wu compared stepwise linear regression, k-nearest neighbors (KNN), support vector regression (SVR), random forest (RF), and stochastic gradient boosting methods for estimating forest AGB using Landsat imagery in northwestern Zhejiang, China, and found that the RF model exhibited the best performance [18]. Ma achieved satisfactory results in retrieving both total forest AGB and component biomass at the Greater Khingan Range National Forest Ecosystem Research Station using RF and SVR algorithms [19]. Ali evaluated the performance of machine learning algorithms, including XGBoost, SVM, and RF, using data from Sentinel-2 and Landsat-9 satellites to estimate forest AGB in the Saihanba Forest Farm and its surrounding mountain ranges [20]. Numerous studies have demonstrated that least squares (LS), SVR, and RF models show strong performance in forest AGC estimation. However, significant discrepancies often exist among AGC estimates derived from different algorithms within the same study area. Therefore, constructing AGC estimation models using multiple algorithms and selecting the optimal model is a common and effective strategy for improving estimation accuracy.

The primary objective of this study is to construct a high-accuracy, multi-temporal remote sensing framework for estimating forest AGC and to map its decadal spatiotemporal trajectory (2009–2019) in the core protection zone of Qianjiangyuan–Baishanzu National Park. By coupling Landsat imagery with ground-based survey plots, this study evaluates the performance of three modeling algorithms (LS, SVR, and RF) across three periods. Crucially, to mitigate optical data saturation and suppress the spatial propagation of model uncertainty, we introduce the Boruta feature selection algorithm to isolate the optimal multi-scale texture features that best capture canopy structural heterogeneity. Ultimately, this research aims to provide a scientific methodology baseline to support differentiated carbon enhancement and functional zone management within the national park framework.

2. Study Area and Data

2.1. Study Area

The Qianjiangyuan–Baishanzu National Park (27°32′25″–27°58′28″N, 118°57′49″–119°22′09″E) is located in the upper reaches of the Oujiang River in Lishui, Zhejiang Province, China. Geomorphologically, it lies within the low- to middle-elevation mountainous region of the Min–Zhe Hills, forming part of the southeastern coastal mountainous terrain of China and belonging to the residual branch of the Wuyi Mountain system (Donggong Mountain range). The park spans the administrative boundaries of Longquan City, Qingyuan County, and Jingning County (Figure 1). The total planned area of the park is 505.88 km², of which the strictly protected core zone accounts for 52%, while the general control zone accounts for 48% [21]. Among the three administrative units, the Longquan section covers 248.90 km² (49.30%), the Qingyuan section covers 205.40 km² (40.60%), and the Jingning section covers 51.58 km² (10.10%). The park encompasses 10 townships, 33 administrative villages, and 4 state-owned forest farms or management units, comprising a total of 10 forest districts. The study area is located in western Lishui and belongs to the low- to middle-mountain region of the Min–Zhe hilly area along the southeastern coast of China. The climate is characterized as a subtropical humid monsoon climate, with warm temperatures, abundant precipitation, and high relative humidity throughout the year. The forest coverage rate of the park exceeds 94%, which is more than 30 percentage points higher than the average forest coverage of Zhejiang Province. The Baishanzu area contains some of the best-preserved natural vegetation in the subtropical region of China, making it of exceptional conservation and scientific research value. The dominant zonal vegetation types include evergreen broad-leaved forests (EBF), evergreen coniferous forests (ENF), and mixed coniferous–broadleaved forests, with main species such as Castanopsis eyrei, Schima superba, Cunninghamia lanceolata, and Pinus massoniana [22].

Figure 1. Study area: (a) Geographical location and elevation of Qianjiangyuan Baishanzu National Park; (b) Landsat 5 TM remote sensing imagery of the study area in 2009; (c) Landsat 8 OLI remote sensing imagery of the study area in 2019.

2.2. Study Data

2.2.1. Remote Sensing Data Acquisition and Preprocessing

This study was conducted using the Google Earth Engine (GEE) cloud computing platform, from which Landsat 5 TM and Landsat 8 OLI surface reflectance images of the Baishanzu area were acquired for the years 2009, 2014, and 2019. GEE provides extensive archives of remotely sensed data and powerful parallel computing capabilities, enabling efficient and accurate large-scale data processing [23]. The use of the GEE platform substantially reduced data processing time while improving processing efficiency and reproducibility. The images selected in this study underwent preprocessing steps including radiometric calibration, atmospheric correction, and geometric correction, resulting in high-quality data suitable for quantitative time series analysis. Image selection followed two main criteria: (1) images were acquired during the peak growing season (June to September) to minimize the influence of phenological variation on vegetation spectral characteristics; and (2) images with cloud cover less than 10% were selected to ensure high quality and temporal consistency across the image time series. Preprocessing was applied to all eligible images. Cloud and cloud-shadow masking was performed using the CFMask algorithm provided by the United States Geological Survey [24]. Missing pixels resulting from cloud removal were filled using cloud-free observations from adjacent months, thereby generating complete, continuous, and high-quality cloud-free image composites. Subsequently, median compositing was applied to the processed images, resulting in three cloud-free Landsat images of the Baishanzu area, which were used to construct remote sensing-based forest AGC estimation models.

2.2.2. Field Plot Data and Processing

The field-measured data used in this study were obtained from plot surveys conducted within the Baishanzu area in 2009, 2014, and 2019. A total of 30 m × 30 m sample plots were established for field investigation. For each tree with a DBH greater than 5 cm within the plots, parameters including tree height, DBH, height to crown base, and crown width were recorded. Subsequently, the AGC of each tree was calculated using species-specific allometric equations [25]. Based on these calculations, the AGC density for each sample plot was determined, and the total AGC per plot was obtained by multiplying the AGC density by the plot area. The total AGC across the study area was then calculated as the sum of AGC values for all forest pixels. Outliers were removed using the two-standard deviation method to ensure that observed AGC values fell within the 95% confidence interval [4]. A total of 66 sample plots were ultimately obtained. The samples were arranged in ascending order of AGC density and subsequently stratified into three tiers: low(0–30 Mg/ha), medium (30–70 Mg/ha), and high (>70 Mg/ha). Within each stratum, samples were randomly partitioned at a ratio of 7:3 to construct a training set and a testing set, respectively. The training set was used to calibrate the AGC estimation model, while the testing set served to validate the model’s estimation accuracy and generalization capability.

The AGC was estimated using the following allometric equation [26]:

$$AGC=a\times b\times (DB{H}^2\times H{)}^c$$

$$Total AGC=({\displaystyle \sum _{i=1}^M}{AGC}_{density,i}\times n)\times {10}^{-6}$$

where $a$ denotes the carbon conversion coefficient, which was assigned a value of 0.47 for broadleaved forests and 0.5 for coniferous forests in this study. The coefficients $b$ and $c$ are species-specific parameters that require empirical determination. $Total AGC$ represents the total forest aboveground carbon storage, expressed in Tg C. ${AGC}_{density,i}$ represents the aboveground carbon storage density of the $i$ forest pixel, and M represents the total number of forest pixels, expressed in Mg/ha. $n$ represents the pixel area.

3. Materials and Methods

3.1. Remote Sensing Variable Configuration

As shown in Table 1, the remote sensing variables used in this study include original spectral bands, vegetation indices, tasseled cap transformation components, and texture features. The original spectral bands consist of the coastal aerosol band, blue band, green band, red band, near-infrared (NIR) band, shortwave infrared 1 (SWIR1) band, and shortwave infrared 2 (SWIR2) band. Vegetation indices are quantitative indicators derived from specific combinations of multispectral remote sensing bands and are widely used to characterize vegetation cover conditions and associated biophysical parameters [27,28]. The vegetation indices employed in this study include the Difference Vegetation Index (DVI), Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Enhanced Vegetation Index (EVI), Ratio Vegetation Index (RVI), Soil-Adjusted Vegetation Index (SAVI), Normalized Burn Ratio (NBR), and Chlorophyll Absorption Reflectance Index (CARI). Texture features were derived from the original images using gray-level co-occurrence matrix (GLCM)-based methods, which are effective in enhancing land cover characterization and classification accuracy [29]. Texture features were extracted for all spectral bands using the tool in ENVI 5.6, including mean, variance, homogeneity, contrast, dissimilarity, entropy, angular second moment, and correlation. To capture texture information at multiple spatial scales, GLCM window sizes (f) of 3 × 3, 5 × 5, 7 × 7, and 9 × 9 pixels were applied in this study. The tasseled cap transformation (Tasseled Cap Transformation, TCT) is an orthogonal linear transformation based on multispectral data that effectively separates vegetation, soil, and water information [30]. Four key TCT components were used to enhance surface feature representation: tasseled cap angle (TCA), tasseled cap brightness (TCB), tasseled cap greenness (TCG), and tasseled cap wetness (TCW).

Table 1. Remote sensing feature variable information.

Types of Remote Sensing Variables	Bands	Formula or Description
Original bands	Coastal	B1
	Blue band (B)	B2
	Green band (G)	B3
	Red band (R)	B4
	Near-infrared band (NIR)	B5
	Shortwave infrared 1 band (SWIR 1)	B6
	Shortwave infrared 1 band (SWIR 2)	B7
Spectral vegetation indices	Normalized difference vegetation index (NDVI)	$\mathrm{NDVI}={\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{E}\mathrm{D}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}\mathrm{E}\mathrm{D}}}$
	Ratio Vegetation Index (RVI)	$\mathrm{R}\mathrm{V}\mathrm{I}={\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}}{\mathrm{R}\mathrm{e}\mathrm{d}}}$
	Difference vegetation index (DVI)	$\mathrm{D}\mathrm{V}\mathrm{I}=\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{e}\mathrm{d}$
	Enhanced vegetation index (EVI)	$\mathrm{E}\mathrm{V}\mathrm{I}=\mathrm{G}·{\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{e}\mathrm{d}}{\mathrm{N}\mathrm{I}\mathrm{R}+{\mathrm{C}}_1\times \mathrm{R}\mathrm{e}\mathrm{d}-{\mathrm{C}}_2\times \mathrm{B}\mathrm{l}\mathrm{u}\mathrm{e}+\mathrm{L}}}$
	Soil adjusted vegetation index (SAVI)	$\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}={\displaystyle \frac{(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{e}\mathrm{d})\times (1+\mathrm{L})}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}\mathrm{e}\mathrm{d}+\mathrm{L}}}$
	Normalized Difference Water Index (NDWI)	$\mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}={\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}}$
Spectral vegetation indices	Normalized Burn Ratio (NBR)	$\mathrm{N}\mathrm{B}\mathrm{R}={\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}2}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}2}}$
	Chlorophyll Absorption Reflectance Index (CARI)	$\mathrm{C}\mathrm{A}\mathrm{R}\mathrm{I}={\displaystyle \frac{\left|\mathrm{R}\mathrm{e}\mathrm{d}\mathrm{E}\mathrm{d}\mathrm{g}\mathrm{e}-\mathrm{a}\times \mathrm{R}\mathrm{e}\mathrm{d}-\mathrm{b}\right|}{\sqrt{{\mathrm{a}}^2+1}}}$
Texture features based on the gray-level co-occurrence matrix (GLCM)	Mean	$\mathrm{Mean}=\mathrm{\mu }={\sum }_{i=0}^{N-1}{\displaystyle {\sum }_{j=0}^{N-1}}i\cdot P\left(i,j\right)$
	Variance	$\mathrm{Variance}={\displaystyle \sum _{i=0}^{N-1}}{\displaystyle \sum _{j=0}^{N-1}}{\left(i-\mathrm{\mu }\right)}^2\cdot P\left(i,j\right)$
	Homogeneity	$\mathrm{Homogeneity}={\displaystyle \sum _{i=0}^{N-1}}{\displaystyle \sum _{j=0}^{N-1}}{\displaystyle \frac{P\left(i,j\right)}{1+\left|i-j\right|}}$
	Contrast	$\mathrm{Contrast}={\displaystyle \sum _{i=0}^{N-1}}{\displaystyle \sum _{j=0}^{N-1}}{\left(i-j\right)}^2\cdot P\left(i,j\right)$
	Dissimilarity	$\mathrm{Dissimilarity}={\displaystyle \sum _{i=0}^{N-1}}{\displaystyle \sum _{j=0}^{N-1}}\left|i-j\right|\cdot P\left(i,j\right)$
	Entropy	$\mathrm{Entropy}=-{\displaystyle \sum _{i=0}^{N-1}}{\displaystyle \sum _{j=0}^{N-1}}P\left(i,j\right)\cdot \log P\left(i,j\right)$
	Second Moment	$\mathrm{Second} \mathrm{Moment}={\displaystyle \sum _{i=0}^{N-1}}{\displaystyle \sum _{j=0}^{N-1}}P{\left(i,j\right)}^2$
	Correlation	${\displaystyle \sum _{i=0}^{N-1}}{\displaystyle \sum _{j=0}^{N-1}}{\displaystyle \frac{\left(i-{\mathrm{\mu }}_i\right)\left(j-{\mathrm{\mu }}_j\right)\cdot P\left(i,j\right)}{{\mathrm{\sigma }}_i{\mathrm{\sigma }}_j}}$
Tasseled cap transformation	Angle (TCA)	0.3037 × B + 0.2793 × G + 0.4743 × R + 0.5585 × NIR + 0.5082 × SWIR1 + 0.1863 × SWIR2
	Brightness (TCB)	0.2043 × B × 0.4158 × G + 0.5524 × R + 0.5741 × NIR + 0.3124 × SWIR1 + 0.2303 × SWIR2
	Greenness (TCG)	−0.1603 × B − 0.2819 × G − 0.4934 × R + 0.7940 × IR − 0.0002 × SWIR1 − 0.1446 × SWIR2
	Wetness (TCW)	0.0315 × B + 0.2021 × G + 0.3102 × R + 0.1594 × NIR − 0.6806 × SWIR1 − 0.6109 × SWIR2

Based on Landsat 8 OLI data, a total of 243 remote sensing feature variables were extracted, including 7 original spectral bands, 8 vegetation indices, 4 tasseled cap transformation variables, and 224 texture features. The texture features were derived from the 7 original bands using four window sizes, with eight texture metrics extracted per band, resulting in a comprehensive multi-dimensional feature set for forest AGC estimation.

3.2. Feature Selection Based on the Boruta Algorithm

Feature selection refers to the process of identifying the most informative variables from an original feature set to reduce feature dimensionality and eliminate redundant or irrelevant information [31]. In machine learning applications, appropriate feature selection is crucial for improving model performance and interpretability. By filtering out variables that are weakly related or unrelated to the dependent variable, feature selection facilitates a clearer understanding of the contribution of explanatory variables to model outputs [32]. The Boruta algorithm, proposed by Kursa, is an efficient and robust feature selection method specifically designed to identify all relevant features in a dataset [33,34]. In this study, the 243 original features were duplicated and randomly shuffled to generate shadow features, which were then concatenated with the original features to train a random forest model. The Z-scores of the original and shadow features were calculated and compared. A binomial distribution test with a significance level of p = 0.01 was applied over 500 iterations to assess feature significance. Features with Z-scores significantly exceeding the maximum Z-score of the shadow features were retained as significant features; otherwise, they were discarded. Unlike conventional feature selection approaches that rely on manually defined importance thresholds, the Boruta algorithm performs automatic feature selection based on statistical hypothesis testing. It exhibits strong robustness against random noise and effectively reduces the risk of selecting spurious variables [35,36]. Therefore, in this study, the Boruta algorithm was employed to screen the extracted remote sensing variables, and the selected features were subsequently used to construct forest AGC estimation models for the Baishanzu area.

3.3. AGC Model Construction

Based on the feature subsets selected by the Boruta algorithm, three modeling approaches—LS, SVR, and RF—were applied to construct forest AGC estimation models for three time periods between 2009 and 2019. Model performance was evaluated separately for each period, and the model with the highest predictive accuracy was selected to perform high-precision estimation of forest AGC in the Baishanzu area and to retrieve its spatial distribution.

3.3.1. LS Model

LS regression is a classical linear regression method that estimates model parameters by minimizing the sum of squared residuals between predicted and observed values [37]. This approach has a clear analytical solution and high computational efficiency, making it suitable for processing large-scale remote sensing datasets. One of the main advantages of LS regression lies in its strong statistical interpretability, as regression coefficients can be rigorously tested and directly reflect the influence of each explanatory variable on the response variable [38]. However, LS regression assumes linear relationships among variables. When applied to forest AGC estimation, the presence of complex nonlinear relationships between predictors and AGC may limit its predictive accuracy. Nevertheless, with appropriate feature engineering and variable transformations, LS regression can still provide a solid baseline and reasonable predictive performance for forest AGC inversion.

3.3.2. SVR Model

SVR is a supervised learning method derived from statistical learning theory and is widely used for nonlinear regression problems [39,40]. Its fundamental principle is based on structural risk minimization. By introducing kernel functions, SVR maps input data into a high-dimensional feature space where a linear regression function is constructed, thereby effectively addressing nonlinear relationships in the original input space [41]. Compared with traditional regression methods, SVR demonstrates clear advantages in handling high-dimensional, small-sample, and nonlinear datasets. Through the ε-insensitive loss function, SVR allows small prediction errors to be ignored within a predefined tolerance, which enhances model generalization ability and robustness to noise [14]. In addition, SVR has a solid theoretical foundation, avoids convergence to local optima, and maintains stable predictive performance under limited sample conditions.

3.3.3. RF Model

RF is an ensemble learning algorithm that constructs a large number of decision trees using bootstrap sampling and random feature selection. The final prediction is obtained by averaging or voting across all individual trees [42,43]. In forest AGC estimation based on remote sensing data, RF is particularly effective in capturing complex nonlinear relationships and high-order interactions among variables without requiring prior assumptions about functional forms or extensive variable transformations [44,45]. By randomly selecting feature subsets at each node split, RF significantly reduces the risk of overfitting and improves model generalization performance. These characteristics enable RF to perform exceptionally well in modeling the complex and nonlinear relationships between multisource remote sensing variables and forest carbon storage, making it one of the most widely applied machine learning algorithms in forest AGC estimation studies.

3.3.4. Hyperparameter Tuning for Machine Learning Models

Model parameter optimization is key to improving model estimation performance. The key hyperparameters and their optimization configurations for the three algorithms used in this study are presented in Table 2. The original dataset was divided into a training dataset (70%) and a testing dataset (30%). A 10-fold cross-validation [46] was employed to optimize the hyperparameters, with the minimization of the root mean square error (RMSE) serving as the criterion for determining the optimal hyperparameter combination for each AGC modeling algorithm. The training dataset was used to construct the AGC model, while the testing dataset was used to validate its performance in estimating AGC. To ensure that the estimation results of different AGC modeling algorithms were not affected by the specific partitioning of the training and testing datasets, the entire procedure of hyperparameter tuning and model performance evaluation was repeated 1000 times [47].

Table 2. Hyperparameters tuned and their configurations for each algorithm.

Algorithm	Hyperparameter Name	Hyperparameter Tuning Configuration
RF	n_estimators	(100, 200, 300, 400, 500, 600, 700, 800, 900)
	max_depth	(1, 5, 10, 15)
	min_samples_split	(2, 4, 6, 8, 10)
	min_samples_leaf	(1, 2, 3, 4, 5)
SVR	kernel function	(Linear, RBF, Polynomial)
	regularization penalty	(0.1, 1, 10, 100)
	kernel width parameter	(0.001, 0.01, 0.1, 1)
	insensitive loss parameter	(0.001, 0.01, 0.1, 1)

3.4. Model Accuracy Assessment

In this study, model performance was evaluated using three metrics: the coefficient of determination (R²), the RMSE, and the relative root mean square error (rRMSE) [4,45,47]. In general, a higher R² value and lower RMSE and rRMSE values indicate better model performance.

The R² and RMSE were calculated as follows:

$$\begin{gathered} R^2=1-{\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n}{\left(y_i-{\widehat{y}}_i\right)}^2}{{\displaystyle {\sum }_{i=1}^n}{\left(y_i-\overline{y}\right)}^2}} \\ \mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n}{\left(y_i-{\widehat{y}}_i\right)}^2} \\ \mathrm{r}\mathrm{RMSE}={\displaystyle \frac{RMSE}{\overline{y}}}\times 100\% \end{gathered}$$

where $y_i$ represents the measured AGC value, $\overline{y}$ is the mean observed AGC, ${\widehat{y}}_i$ denotes the model-predicted AGC, and $n$ is the number of samples. The closer the R² value is to 1, the higher the goodness of fit of the model; meanwhile, the smaller the RMSE and rRMSE values, the smaller the deviation between the observed and predicted values.

3.5. Technical Workflow

The overall workflow for estimating the spatiotemporal distribution of forest AGC using three machine learning algorithms is illustrated in Figure 2. Landsat 5 TM and Landsat 8 OLI images for the years 2009, 2014, and 2019 were acquired online via the GEE platform. The remote sensing data were subjected to radiometric calibration, atmospheric correction, geometric correction, mosaicking, and spatial clipping. Simultaneously, field plot measurements were processed by removing outliers using the two-standard-deviation method. Individual-tree AGC was calculated using species-specific allometric equations, and total observed AGC was obtained by aggregating individual-tree estimates at the plot level. The Boruta algorithm was then applied to select highly relevant variables from four categories of remote sensing predictors, including original spectral bands, vegetation indices, texture features, and tasseled cap transformation components, for each study period. The resulting datasets were randomly divided into training and testing subsets at a ratio of 7:3 and subsequently input into the LS, SVR, and RF models. Model performance was evaluated using R² and RMSE, and the algorithm with the highest predictive accuracy was selected as the optimal model for regional AGC inversion. Finally, spatially explicit maps of forest AGC for the Baishanzu area were generated for the three study periods, enabling an analysis of their spatiotemporal dynamics.

Figure 2. The methodological framework.

4. Results and Analysis

4.1. Forest Distribution in the Qianjiangyuan–Baishanzu National Park

Based on Landsat imagery acquired from the GEE platform, land use classification of the Qianjiangyuan–Baishanzu National Park was performed using the RF classifier to extract forest distribution. The forest distribution maps for the three study periods (2009, 2014, and 2019) are presented in Figure 3. Classification accuracy was evaluated using confusion matrices by comparing predicted results with reference data. For all three periods, the overall classification accuracy exceeded 90%, and the Kappa coefficient was greater than 0.90, indicating a high level of classification reliability. Forest resources in the study area are abundant, with forests occupying a dominant proportion of the total area. Coniferous forests accounted for approximately 32–37% of the total forest area and were mainly distributed in relatively high-elevation regions. Broadleaved forests represented 61–64% of the forest area and were primarily concentrated in lower-elevation areas. Comparison of the forest classification maps across the three periods (Figure 3) reveals that, from 2009 to 2019, non-forest areas decreased markedly, the area of coniferous forests increased, and the area of broadleaved forests first increased and then declined.

Figure 3. Forest distribution across three phases: (a) 2009; (b) 2014; (c) 2019.

4.2. Results of Remote Sensing Variable Selection

Based on the Boruta feature selection algorithm, the optimal combinations of remote sensing variables selected for each study year are summarized in Table 3. For 2009, a total of 40 variables were identified as highly relevant, including 5 original spectral bands, 7 vegetation indices, 23 texture features, and 5 tasseled cap transformation components. For 2014, 32 variables were selected, comprising 4 original spectral bands, 7 vegetation indices, and 21 texture features. For 2019, 26 variables were retained, including 2 original spectral bands, 7 vegetation indices, 15 texture features, and 2 tasseled cap transformation components. These results indicate that the number and composition of influential remote sensing variables varied across years, reflecting temporal differences in vegetation structure, spectral responses, and their relationships with forest AGC.

Table 3. Results of Remote Sensing Variable Selection.

Year	Number of Selected Variables	Name of Selected Variables
2009	40	CVI, DVI, EVI, NDII, NDMI, NDVI, RVI, B1, B1_idm_W7, B1_var_W3, B1_var_W5, B2, B3, B3_con_W7, B3_dent_W7, B3_ent_W7, B3_idm_W7, B3_svar_W7, B3_var_W7, B5, B5_con_W7, B5_dent_W7, B6, B6_asm_W7, B6_con_W3, B6_con_W5, B6_con_W7, B6_dent_W5, B6_dent_W7, B6_ent_W7, B6_idm_W7, B6_svar_W5, B6_svar_W7, B6_var_W5, B6_var_W7, TCA, TCB, TCG, TCW, TVI
2014	32	CVI, DVI, EVI, NDII, NDMI, NDVI, RVI, B1, B1_con_W5, B1_con_W7, B1_dent_W7, B2, B2_con_W5, B2_con_W7, B2_dent_W7, B2_var_W5, B2_var_W7, B3, B4, B4_con_W5, B4_con_W7, B4_svar_W7, B4_var_W5, B4_var_W7, B5_con_W5, B7_con_W5, B7_con_W7, B7_dent_W5, B7_dent_W7, B7_svar_W7, B7_var_W5, B7_var_W7
2019	26	B3, TCA, B5_con_W9, B4_con_W5, B4_con_W7, B5_diss_W7, NDBI, NDII, B4_diss_W5, B4_prom_W5, B5_diss_W9, B4_prom_W9, B7, B4_diss_W9, B4_prom_W7, B2_idm_W9, B6, TCW, B4_var_W7, B4_con_W9, B2_diss_W5, B5_var_W7, B2_idm_W5, B5_con_W3, B4_con_W3, B3_var_W9, B4_diss_W3, B3_diss_W9, B2_diss_W9

4.3. AGC Model Construction and Prediction Results

The training and testing results of the three models across all study years were evaluated and visualized using scatter plots comparing observed and predicted AGC values (Figure 4). The results indicate that the LS model exhibited lower performance than the other two models, with both training and testing R² values being the lowest among the three approaches. In addition, the LS model produced the highest RMSE and Rrmse values. Although the LS model showed similar performance on the training and testing datasets, it exhibited systematic bias, particularly in high-AGC ranges, where substantial discrepancies were observed between predicted and measured values. The SVR model outperformed the LS model on both the training and testing datasets. Notably, on the testing dataset, the SVR model achieved an R² value of 0.934, indicating strong predictive performance and good generalization ability. Among the three models, the RF model demonstrated the best overall performance across all evaluation metrics. On the testing dataset, the RF model achieved an R² value of 0.954, representing an improvement of 2.1% over the SVR model and 8.3% over the LS model. The RMSE of the RF model was 4.653 Mg/ha, which was 17.1% lower than that of the SVR model and 35.8% lower than that of the LS model. Table 4 presents the accuracy performance of the models on the training and testing sets under random split runs across three temporal periods. The RF model achieved higher accuracy than the other two models in each period. Furthermore, the RF model consistently maintained a high level of accuracy across all three periods, demonstrating excellent stability. As shown in Figure 5, the mean residual of the LS model deviates significantly from the zero axis, indicating a systematic overestimation in its overall estimates. Moreover, this model exhibits the largest standard deviation among the three, reflecting the highest degree of dispersion in its inversion uncertainty. In contrast, the overestimation bias of the SVR model is considerably reduced, with its standard deviation decreasing markedly to 4.990, indicating an extremely compact and concentrated error distribution. The RF model shows a mean residual of 2.445, suggesting a slight overall underestimation tendency. Its standard deviation is 5.112, which falls within a compact error range comparable to that of the SVR model and is substantially superior to that of the LS model. These results indicate that the RF model not only achieved excellent fitting performance on the training dataset but also maintained superior predictive accuracy on the testing dataset. Overall, the RF model clearly outperformed the LS and SVR models in terms of prediction accuracy and generalization capability, followed by the SVR model, while the LS model showed the weakest performance among the three.

Figure 4. Accuracy assessment of three Landsat-based AGC prediction models per year.

Table 4. Training and testing set accuracy of the three models in 2009, 2014, and 2019.

Time		LS (Train)	LS (Text)	SVR (Train)	SVR (Text)	RF (Train)	RF (Text)
2009	R2	0.84	0.89	0.96	0.91	0.97	0.85
	RMSE (Mg/ha)	6.29	4.89	4.01	4.13	3.80	8.13
2014	R2	0.93	0.93	0.96	0.94	0.96	0.96
	RMSE (Mg/ha)	5.59	6.67	5.55	4.78	5.47	4.39
2019	R2	0.87	0.70	0.89	0.90	0.94	0.92
	RMSE (Mg/ha)	10.67	7.25	6.52	5.61	6.37	5.25

Figure 5. Residual distribution analysis of the three models.

In summary, the forest AGC estimation model based on the RF algorithm exhibited the highest accuracy and the best overall performance. Therefore, the RF model was selected as the optimal model for retrieving and mapping the spatiotemporal distribution of forest AGC in the Qianjiangyuan–Baishanzu National Park.

4.4. Spatiotemporal Variation of Forest AGC in the Qianjiangyuan–Baishanzu National Park

The spatial distribution of forest AGC density in the Qianjiangyuan–Baishanzu National Park from 2009 to 2019, as estimated by the RF model, is shown in Figure 6. In 2009, low AGC values dominated the study area, with most regions exhibiting AGC densities ranging from 20 to 70 Mg/ha. Relatively low AGC values were mainly observed in parts of the northeastern and southwestern areas. The mean forest AGC density across the park was 37.64 Mg/ha, corresponding to a total forest AGC stock of 2.16 Tg C. By 2014, AGC density increased overall across the study area. Several regions in the northeastern and southwestern parts of the park exhibited darker color tones in the AGC maps, indicating enhanced forest carbon storage, while areas with previously high AGC values largely maintained their elevated carbon stocks. The mean forest AGC density increased to 48.10 Mg/ha, and total forest AGC reached 2.76 Tg C, representing a 27.8% increase compared with 2009. In 2019, forest AGC continued to exhibit a pronounced upward trend, with high AGC values becoming dominant across the study area. At this stage, nearly all regions displayed AGC densities exceeding 20–40 Mg/ha, and some areas even surpassed 80 Mg/ha. The mean forest AGC density increased to 66.00 Mg/ha, and total forest AGC reached 4.36 Tg C, corresponding to an increase of 101.9% relative to 2009 and 58.0% relative to 2014. Although forest AGC generally increased over time, certain marginal areas and regions strongly influenced by human activities did not exhibit significant growth and even showed slight declines in AGC. This spatial heterogeneity highlights the combined effects of forest management practices, site conditions, and anthropogenic disturbances on forest carbon accumulation within the park.

Figure 6. Distribution of three sets of forest AGC: (a) 2009; (b) 2014; (c) 2019.

5. Discussion

The construction of forest AGC estimation models is highly dependent on the input feature variables, and different types of variables can have a significant influence on model accuracy. To address this issue, the Boruta algorithm was employed in this study to screen feature variables and eliminate redundant and irrelevant information, thereby improving model performance. The Boruta algorithm was independently executed 500 times for each study year to evaluate the importance of individual features, and all variables were subsequently ranked according to their importance. Figure 7 illustrates the ranking of feature variables based on their relative importance. The results clearly indicate that texture features exert the most pronounced influence on model construction and predictive accuracy. Among the input features for 2009, 2014, and 2019, texture features accounted for 57.5%, 65.6%, and 73.1% of the selected variables, respectively, which is substantially higher than the proportions of original spectral bands, vegetation indices, and tasseled cap transformation components. Overall, texture features emerged as the most important category of predictors, followed by vegetation indices and original spectral bands. These findings demonstrate that texture information plays a central role in the development of high-accuracy forest AGC estimation models, which is consistent with the conclusions reported by Zhang [45]. As shown in Table 5, as the texture window size expanded from 3 × 3 to 9 × 9, the testing set R² steadily increased from 0.80 to the global maximum of 0.90. However, when the window size was further enlarged to 11 × 11, the R² decreased to 0.87. This decline is attributable to the oversmoothing effect of texture information caused by excessively large windows crossing different forest type patches, which consequently reduced estimation accuracy. Texture features effectively capture spatial heterogeneity and structural characteristics of forest canopies, such as crown arrangement and stand density, which are closely related to aboveground carbon storage but cannot be fully represented by spectral information alone [48]. Further analysis revealed clear temporal differences in the contribution of texture features derived from different window sizes. In 2009 and 2014, texture features extracted using a 7 × 7 window dominated the selected texture variables, accounting for 69.57% and 57.14% of the total texture features, respectively. In contrast, in 2019, texture features derived from a 9 × 9 window were the most influential, representing 47.37% of the selected texture variables. These differences likely reflect temporal changes in forest structural heterogeneity, as well as scale effects associated with variations in forest canopy structure and the spatial resolution of remote sensing imagery.

Figure 7. Boruta variable screening results of three sets.

Table 5. Quantitative comparison of RF model estimation accuracy across different GLCM texture window sizes.

Texture Window Size	Number of Input Features	Training Set R2	Testing Set R2
3 × 3	32	0.81	0.80
5 × 5	32	0.87	0.84
7 × 7	32	0.91	0.90
9 × 9	32	0.94	0.95
11 × 11	32	0.89	0.87

This study employed three machine learning algorithms to quantitatively estimate forest AGC in the Qianjiangyuan–Baishanzu National Park from 2009 to 2019. The results indicate that the RF model outperformed both the LS and SVR models in terms of predictive accuracy. The LS model is widely used in forest AGC estimation studies due to its simplicity and strong interpretability. However, it still has certain limitations. The LS model estimates parameters by minimizing the squared error, making it highly sensitive to outliers, which can substantially distort regression coefficients [49] and consequently impair the model’s generalization capability. As shown in Table 6, this results in systematic bias in the LS model’s estimates for high-AGC areas, exhibiting a notable underestimation trend, thereby limiting its accuracy and applicability in complex ecosystem modeling. SVR, as a powerful machine learning method, offers advantages in handling nonlinear relationships and has been extensively applied in various predictive studies. Nevertheless, the performance of the SVR model is highly dependent on the selection of its core hyperparameters, including the penalty coefficient, kernel function type, and associated parameters. Inappropriate parameter settings may lead to overfitting or underfitting [50]. As shown in Table 6, SVR performs well in low- to medium-AGC areas but suffers from interference due to collinear variables in high-AGC areas, resulting in an increased RMSE. In contrast, RF integrates multiple decision trees for prediction, employing averaging or voting mechanisms, which reduces its dependence on specific parameters while providing good robustness against outliers. In the context of remote sensing-based forest AGC estimation, the RF model maintains satisfactory performance even in high-AGC areas. Overall, the RF model demonstrates clear superiority over both the LS and SVR models, offering more stable and high-accuracy inversion results, which is consistent with the findings of Li [47] and Huang [4].

Table 6. Accuracy comparison of the three models across different value ranges.

Value Range	LS		SVR		RF
	R2	RMSE	R2	RMSE	R2	RMSE
0–30	0.85	5.61	0.89	2.91	0.85	4.05
30–70	0.83	5.97	0.87	5.93	0.95	2.75
>70	0.64	8.03	0.66	7.81	0.80	6.85

Uncertainty analysis provides a means to quantify the impact of uncontrollable external factors on the research process [47,51,52]. Because errors can propagate and amplify, it is necessary to consider the combined effects of multiple sources of uncertainty. In this study, an error propagation approach was employed to quantify pixel-level uncertainty in forest AGC estimates [53], analyzing uncertainties arising from both field plot measurements and model simulations. As shown in Figure 8, the overall uncertainty of forest AGC in the Baishanzu area was relatively low for all three periods between 2009 and 2019, with an average error of only 14.2%. This indicates that the RF model provides stable and reliable estimates of forest AGC in the Qianjiangyuan–Baishanzu National Park, and that the results are minimally affected by external factors, further confirming the accuracy of the study’s findings. Additionally, uncertainty statistics were calculated for different forest types in this study. The results showed that for evergreen broad-leaved forests, the mean error was 11.2%, with low-uncertainty areas accounting for 76.8% and high-uncertainty areas accounting for 6.5%. For evergreen coniferous forests, the mean error was 15.4%, with low-uncertainty areas comprising 52.1% and high-uncertainty areas 14.2%. The low-uncertainty areas were mainly concentrated in the natural forest reserve of the national park, where the forest community structure is intact and the degree of homogeneity is high, allowing the RF model to achieve excellent fitting performance.

Figure 8. Uncertainty analysis of forest AGC prediction results of Qianjiangyuan Baishanzu National Park: (a) 2009; (b) 2014; (c) 2019.

The mean AGC densities for the three periods were 37.7 Mg/ha, 48.1 Mg/ha, and 76.2 Mg/ha, respectively. These values are generally consistent with the provincial-scale AGC estimates for Zhejiang Province reported by Huang [54] and Liu [6]. From 2009 to 2019, forest AGC in the Qianjiangyuan–Baishanzu National Park exhibited a clear upward trend. As shown in Figure 9, the annual mean growth rate of forest AGC in broadleaved forest areas was 20.8%, contributing 68.64% to the quantitative increase in total carbon storage. Among these areas, the central–eastern region exhibited the most pronounced increase in forest AGC, whereas areas near roads in the central region showed relatively lower AGC. For coniferous forest areas, the annual mean growth rate of forest AGC was 11.2%, with a contribution rate of 31.36% to the quantitative increase in carbon storage. Notably, the carbon storage pattern shifted from being dominated by low values to being dominated by high values. Spatially, coniferous forests in the southwestern and central regions displayed relatively lower AGC, while coniferous forests in the eastern and northwestern regions demonstrated significantly higher AGC increases. The synergistic effect of topography and climate constitutes the core natural driving force behind the increase in carbon storage in Baishanzu National Forest Park. Topographically, the natural barrier formed by high-altitude areas effectively isolates human disturbance, which not only ensures the long-term stability of the soil carbon pool in old-growth forests but also creates favorable conditions for the sustained natural growth of young and middle-aged forests [55]. Meanwhile, the abundant hydrothermal conditions in the subtropical region significantly promote aboveground biomass accumulation in evergreen forests. This combination of heat and moisture not only extends the growing season and enhances photosynthetic efficiency but also accelerates nutrient cycling and organic matter synthesis [54]. Consequently, the topographical and climatic interactions jointly shape the carbon sink potential of the region’s forest ecosystems across both spatial and temporal scales, serving as a key natural mechanism driving the continuous increase in carbon storage. The gradual expansion of forested area within the Baishanzu region directly contributed to the observed increase in total forest AGC over time. In addition, since the early 21st century, China has implemented a series of ecological engineering initiatives, such as natural forest protection programs, which have positively influenced forest quality. Zhejiang Province, in particular, has actively promoted forest quality improvement projects, including the cultivation of young and mid-aged stands [50]. Data from the Global Ecosystem Dynamics Investigation indicate that during the transition from young to mid-aged forests, key structural parameters—including tree height, DBH, and canopy closure—show significant annual increases with stand age. The cumulative growth effects of increasing tree height, DBH, and canopy closure accelerate individual tree biomass accumulation, thereby driving substantial increases in aboveground carbon stocks at the stand and regional levels. Consequently, these structural improvements enhance overall forest quality, augment forest carbon storage, and increase associated ecological and socioeconomic benefits. In the future, active efforts should be made to promote artificially facilitated healthy forest succession. Through scientific selective thinning and replanting, the proportion of mixed coniferous–broadleaved forests and uneven-aged multi-story forests should be increased, thereby substantially enhancing photosynthetic carbon fixation efficiency in the vertical dimension and achieving high-quality carbon sink enhancement.

Figure 9. Distribution of forest AGC in EBF and ENF from 2009 to 2019: (a) 2009 EBF; (b) 2014 EBF; (c) 2019 EBF; (d) 2009 ENF; (e) 2014 ENF; (f) 2019 ENF.

This study has several limitations. Regarding the input data, Landsat series imagery was used for forest AGC estimation. While adequate for regional-scale analyses, the relatively coarse spatial resolution and limited spectral information of Landsat data become increasingly restrictive for detailed assessments. Although the RF model provides stable predictions, residual analysis reveals that it tends to overestimate AGC in high-AGC areas. This bias is primarily attributed to the high and increasingly saturated canopy leaf area index in high-biomass regions, which causes the spectral reflectance and vegetation indices of forest stands to level off, thereby becoming decoupled from further increases in tree height and DBH. This fundamentally limits the predictive capability of Landsat data for aboveground carbon storage in high-biomass, multi-layered forest stands. Future studies could employ higher-resolution and more spectrally rich remote sensing datasets, such as Sentinel-2A or GaoFen series imagery. These datasets include additional red-edge bands, which can significantly enhance vegetation identification and improve the accuracy of AGC estimation. In terms of model construction, although the RF model can effectively handle high-dimensional data and exhibits strong resistance to noise, it is prone to overfitting when building predictive models. To improve forest AGC simulation accuracy, future studies could integrate boosting-based ensemble learning strategies and introduce parameter optimization to reduce overfitting risks. Furthermore, a long-term monthly AGC continuous accounting framework based on deep learning architectures, such as Transformer or LSTM, should be established to quantitatively assess the dynamic effects of seasonal variation on the carbon sequestration rate of the national park. Such a framework will not only improve the understanding of ecosystem dynamic processes but also provide critical data support for climate change mitigation, forest management optimization, and policy formulation.

6. Conclusions

Based on Landsat imagery, this study applied LS, SVR, and RF machine learning algorithms to quantitatively estimate forest AGC in the Qianjiangyuan–Baishanzu National Park. The main findings are as follows:

(1)

Using the Boruta algorithm, 243 remote sensing variables were screened, and the most AGC-relevant features for each study period were identified. Texture features emerged as the most important predictors for forest AGC estimation. In 2009, 2014, and 2019, texture features accounted for 57.5%, 65.6%, and 73.1% of the selected variables, respectively, significantly higher than other feature types. For 2009 and 2014, texture features extracted using a 7 × 7 window dominated, representing 69.57% and 57.14% of total texture features, respectively; in 2019, features derived from a 9 × 9 window accounted for 47.37%, reflecting changes in forest structural heterogeneity and scale effects.

(2)

All three machine learning models achieved high-accuracy estimates of forest AGC in the park. The RF-based AGC model demonstrated the highest predictive performance, with training set R² = 0.938, RMSE = 5.550 Mg/ha, rRMSE = 12.52%, and testing set R² = 0.954, RMSE = 4.653 Mg/ha, rRMSE = 10.09%. Overall, the RF model outperformed both the LS and SVR models in estimating forest AGC.

(3)

Forest AGC in the Qianjiangyuan–Baishanzu National Park increased steadily from 2009 to 2019. Mean AGC density rose from 37.64 Mg/ha to 66.00 Mg/ha, and total forest AGC increased from 2.16 Tg C to 4.36 Tg C. This growth is attributable to the annual increases in key vegetation structural parameters, including tree height, diameter at breast height, and crown width, as well as to the positive effects of various forest protection and restoration policies implemented within the park.