LAI Mapping of Winter Moso Bamboo Forests Using Zhuhai-1 Hyperspectral Images and a PSO-SVM Model

Abstract

Moso bamboo forests (MBFs) are unique subtropical ecosystems characterized by distinct leaf phenology, bamboo shoots, rapid growth, and carbon sequestration capability. Leaf area index (LAI) is an essential metric for evaluating the productivity and ecological quality of MBFs. However, accurate and large-scale methods for remote-sensing-based LAI monitoring during the winter growth stage remain underdeveloped. This study introduces a novel method integrating hyperspectral indices from Zhuhai-1 Orbit Hyperspectral Satellites (OHS) imagery with the particle swarm optimization-support vector machine (PSO-SVM) coupling model to estimate LAI in winter MBFs. Five traditional vegetation indices (VI_Rs) and their red-edge variants (VI_REs) were optimized to build empirical models. Machine learning algorithms, including SVM, Random Forest, extreme gradient boosting, and partial linear regression, were also applied. The PSO-SVM model, integrating three VI_Rs and three VI_REs, achieved the highest accuracy (R² = 0.721, RMSE = 0.490), outperforming traditional approaches. LAI was strongly correlated with indices, such as NDVI_R, RVI_R, EVI_RE, and SAVI_R (R > 0.77). LAI values of MBFs primarily ranged from 2.1 to 5.5 during winter, with values exceeding 4.5 indicating high winter bamboo shoot harvesting. These findings demonstrate the potential of OHS data to improve LAI retrieval models for large-scale LAI mapping, offering new insights into MBFs monitoring and contributing to sustainable forest management practices.

1. Introduction

Canopy leaf area index (LAI) is a crucial parameter for assessing the productivity and health of forest ecosystems, as well as for simulating soil erosion, water cycling, and biogeochemical processes [1,2,3]. The accurate acquisition of spatially explicit LAI data is essential for precision forestry management [4,5]. Moso bamboo (Phyllostachys edulis), a widely distributed and economically vital bamboo species in southern China (covering ~490 million ha), plays a dual role in ecological conservation and sustainable development [6,7]. As an evergreen perennial graminaceous plant, it plays a vital role in the global “Bamboo for Plastics Replacement” initiative for reducing plastic pollution and replacing plastic products [8]. Moso bamboo forests (MBFs) are typical multi-purpose forests, providing bamboo timber, forest food, and carbon sequestration [8]. Notably, the winter LAI of MBFs directly influences both winter and spring bamboo shoot yields in the following year, making winter a critical period for remote sensing-based monitoring. Previous studies related to the remote-sensing dynamic information extraction of bamboo specie, have shown promise for accurate LAI estimation [2,8]. For example, Wang et al. achieved high-precision understory bamboo mapping using winter Landsat imagery [9]. Therefore, rapid, accurate, and large-scale monitoring LAI distribution in winter is essential for better understanding the growth status across different regions, assessing the production potential of winter bamboo shoots, and guiding scientific forest management practices.

Traditional field-based LAI measurement methods, which rely on non-destructive optical instruments such as LAI-2200, are time-consuming and labor-intensive [10]. This makes them impractical for large-scale LAI mapping for remote mountainous bamboo forests. In contrast, satellite remote-sensing technology can accurately and quickly obtain ground surface characteristics over large regions, providing an effective and cost-effective way to estimate LAI in bamboo forests [11]. In recent decades, researchers have successfully utilized the vegetation index (VI) from multispectral satellite images (MSI), such as Landsat and Sentinel-2A, to retrieve the large-scale LAI of bamboo forests [2,11,12]. For instance, Yao et al. constructed the normalized difference vegetation index (NDVI) and the ratio vegetation index (RVI) using hyperspectral narrow bands, achieving high accuracy in LAI estimation with an R² value of 0.73 [12]. Vegetation spectra, especially in the red-edge region, are influenced by factors such as the tree species, geographic environment, and leaf phenology, which can significantly impact LAI retrieval accuracy [13]. These VIs can mitigate issues such as noise from the soil background and spectral saturation in dense canopy cover, producing high-accuracy LAI retrievals. For example, anti-interference vegetation indices, such as the enhanced vegetation index (EVI) and the soil-adjusted vegetation index (SAVI), have been developed for LAI inversion [14,15]. However, significant variations in chlorophyll content across different bamboo forests can reduce the accuracy of LAI inversion models. Recent studies have demonstrated that using red-edge vegetation indices (VI_RE) from multispectral data can enhance LAI estimation accuracy due to them being insensitive to chlorophyll content [16,17]. Emerging studies highlight the advantages of the VI_RE in mitigating spectral saturation and improving LAI estimation [18,19]. For instance, Sun et al. proposed a new chlorophyll-insensitive red-edge index that effectively improves the accuracy of crop LAI estimation [19]. The red-edge VIs constructed from MSI data (e.g., Sentinel-2A, Gaofen-6) can accurately estimate LAI [20,21]. Nevertheless, the red-edge MSIs generally have few bands and limited spectral information on diverse tree species with various leaf biophysical statuses for LAI inversion [22].

Hyperspectral satellite images (HSI) provide a powerful alternative for retrieving canopy biophysical traits, such as LAI and canopy chlorophyll content, due to their high spectral resolution and multiple red-edge spectral bands, which significantly enhance LAI estimation accuracy [23]. Previous studies have demonstrated the effectiveness of hyperspectral imagery in LAI inversion, achieving high accuracy [13,24]. For example, several types of spaceborne hyperspectral data, such as EO-1 Hyperion, Gaofen-5, and CHIME, have been used to estimate canopy LAI [14,25,26]. However, these hyperspectral images are limited by their low spatial resolution (≥30 m), which is insufficient for precise LAI mapping in distribution areas of MBFs, which are characterized by fragmented and spatially heterogeneous patches due to varying tree species, management intensities, and uneven-aged mixed mother culms [18]. Additionally, acquiring these data is often hindered by challenges such as cloud cover, the prolonged rainy season in subtropical regions, and the long satellite revisit periods [27,28]. Currently, Zhuhai-1 Orbit Hyperspectral Satellite (OHS) images offer a promising solution to these limitations. With their high spatial resolution (10 m) and extensive coverage (approximately 2 d of global cover with an eight-satellite network), OHS data present a unique opportunity to provide suitable high-resolution images for LAI estimation in winter MBFs [8,29]. This capability facilitates site-specific management to assess the potential of bamboo shoot production and forest health status. However, the effectiveness of OHS data for estimating canopy traits in forest ecosystems remains largely unexplored. In particular, the influence of different band-combined VIs on the LAI estimation of MBFs has not been considered. Therefore, this study represents a pioneering effort to evaluate OHS-derived VIs, including classic red VIs (VI_Rs) and VI_REs, with the aim of improving the accuracy of winter LAI estimation in MBFs.

The retrieval model plays a crucial role in accurately quantifying bamboo LAI. Empirical and machine-learning algorithm models are important methods for constructing LAI inversion models. Empirical statistical models are used to establish regression relationships between the remote-sensing index (e.g., spectral reflectance or vegetation index) and measured data (e.g., canopy LAI). For example, Qiao et al. [21] proposed an improved VI_RE-based equation that significantly improved the prediction ability of crop LAI, and Bajocco et al. [30] found that empirical equations are most commonly used for LAI estimation in an NDVI. Studies have shown that the model of multiple variables, such as RVI and NDVI, outperform univariate empirical models for LAI estimation [31]. Empirical models are simple and user-friendly but are susceptible to multivariate collinearity, whereas machine-learning models are generally insensitive to this issue and may be a useful solution to the limitations of empirical models [32]. In recent years, machine-learning algorithms, such as random forest (RF), support vector machines (SVM), and gradient boosting machines, have been widely adopted in remote-sensing vegetation parameter inversion [3,16]. Among these models, the SVM approach excels in high-dimensional data handling, performs effectively in small-sample scenarios, and provides interpretable models with good generalization ability [33]. However, determining the optimal hyperparameters is an important condition for improving SVM models’ learning and generalization abilities. Several studies have attempted to integrate hyperparameter optimization techniques (e.g., the particle swarm optimization algorithm [PSO]) to optimize SVM modeling, significantly improving the accuracy of the soil elemental content and quantitative water quality index [33,34]. Although empirical or machine-learning algorithms have been widely used for LAI retrieval in crops or forests, there is still a lack of studies comparing the performance of bamboo LAI using both classical machine learning models and SVM hyperparameter optimization approaches based on HSI data.

Current research on LAI mapping has primarily focused on crops or forests, with few studies investigating bamboo species (in particular the P. edulis) during the winter growth stage. Given the potential applications of Zhuhai-1 OHS data in bamboo canopy traits inversion and the importance of LAI mapping for management, we selected the abundant bamboo forest area of Fujian Province as the study area and used OHS images to screen sensitive hyperspectral VIs for the retrieval of LAI by integrating the PSO-SVM optimization model and the optimal VIs to construct the optimal LAI inversion model for winter MBFs. The specific objectives of the study were to (1) identify the sensitivity of spectral bands and VIs (red-band and red-edge-band based) to LAI estimation in winter MBFs; (2) evaluate the contribution of an improved PSO-SVM model for estimating LAI; and (3) obtain the optimal model for mapping LAI information, thereby providing a scientific basis for guiding the precise management of MBFs to enhance the bamboo forest management quality.

2. Materials and Methods

2.1. Study Area

The study area is located in Yong’an City, Fujian Province, Southeast China (117°22′58″–117°35′8″ N; 25°54′40″–26°1′20″ E) (Figure 1), where there are 116,000 acres of bamboo forests [35]. It is characterized by mountainous terrain, with a peak elevation of 1574 m and a minimum elevation of 185 m, and belongs to the mid-subtropical monsoon climate zone [35]. The average annual temperature, humidity, and precipitation in this study area are 15 °C, 80%, and 2039 mm, respectively. Summer is often characterized by typhoons and torrential rain from June to August. The rainy season of continuous drizzle from March to May poses a major challenge for optical remote-sensing data acquisition for land observations [36].

2.2. Study Workflow

This study was divided into three steps (Figure 2). First, we measured the field LAI values of MBFs and prepared hyperspectral VIs based on OHS image processing. Second, to select VIs, we analyzed red-based and red-edge-based VIs sensitive to LAI, designed different VI combinations, and constructed LAI estimation models. Finally, we evaluated the optimal model and applied it to the LAI mapping of MBFs during the winter growth stage.

2.3. Data Acquisition and Preparation

2.3.1. Field Measurements

In 2022–2023, sixty-four 10 × 10-m sample plots were established across different altitudes zones, encompassing diverse management practices (intensive, semi-intensive, and extensive) and bamboo stands of varying ages (on-year vs. off-year stands). These sampling plots were distributed across six major bamboo forest regions in the study area, covering an elevation gradient from 200 m to 1400 m. Through spatial analysis, we ensured that the sampling points encompassed the primary sources of LAI variability within the study area, including elevation, management practices, and bamboo forest types. These plots were designed to investigate bamboo ages, diameter at breast height, and stem density in MBFs. LAI values were measured using an optical instrument (LAI-2200 C Plant Canopy Analyzer; Li-Cor Bioscience, Lincoln, NE, USA) selected from the previously set sample plots and dated from 26 November to 8 December 2023. To obtain accurate LAI values of MBFs during the winter growth stage, the following specific requirements were used: (1) survey sample plots were >20 m away from roads or villages, scattering corrections for each sample plot were performed using the one-sensor measurement model, azimuths were recorded before each measurement, and 90° caps were selected to block the sun and operator shadows in the field of view of the lens; (2) at least five LAI data points were measured at each sample plot (four corners and the center point), and the latitude and longitude of each LAI sample point were recorded using global navigation and satellite system units with a position precision < 1 m [37]; and (3) the averaged LAI value of each sample plot was calculated using FV-2200 software (https://licor.app.boxenterprise.net/s/v625lbblp93dexr2vb3ww0axqurhx1tw, access on 6 April 2023), where relevant parameters were adjusted accordingly [3]. For details on LAI measurement operations and precautions, refer to the LAI-2200C Plant Canopy Analyzer operation manual (https://www.licor.com/products/leaf-area/LAI-2200C, accessed on 6 April 2023). Measured LAI values of the 64 P. edulis sample plots were determined.

2.3.2. Zhuhai-1 Orbit Hyperspectral Images Acquisition and Preprocessing

Zhuhai-1 OHS images were acquired and dated on 23 December 2023 and were downloaded from the Zhuhai No.1 Orbita Remote Sensing Data Service Platform (https://www.obtdata.com/en/zhuhai1.html, accessed on 16 April 2023) with cloud detection of <1%. The OHS image size was 50 × 50 km and 32 bands ranged from 400 to 1000 nm [38]; detailed information is listed in

Table 1. Specific information of hyperspectral data of Orbit hyperspectral satellite (OHS) images.

Image Size	Band Type	Band Number and Band Center Wavelength (nm)
5056 rows × 5056 columns	Blue band (B)	Band 1–3: 443, 466, 490
	Green band (G)	Band 4–7: 500, 510, 531, 550, 560
	Yellow band (Y)	Band 8–9: 580, 596
	Yellow edge band (YE)	Band 10–11: 620, 630
	Red band (R)	Band 12–14: 640, 665, 670
	Red edge band (RE)	Band 15–19: 686, 700, 709, 730, 746
	Near infrared band (NIR)	Band 20–32: 760, 776, 780, 806, 820, 833, 850, 865, 880, 896, 910, 926, 940

. The OHS hyperspectral data were processed using radiometric calibration and FLAASH atmospheric correction in ENVI software (ENVI, version 5.3; Boulder, CO, USA) to address atmospheric interference within the Zhuhai-1 OHS imagery [39]; detailed operational parameters are available in the Zhuhai-1 Hyperspectral Satellite Data Product User Manual (V2.6).

2.3.3. Hyperspectral Vegetation Index Extraction

The correlation between hyperspectral data and the P. edulis LAI can be effectively explored by calculating VIs using a combination of hyperspectral bands. Previous studies have proven that the accuracy of LAI estimation can be improved by applying the combinations of two bands for RVI [14] and NDVI [40,41]. The red, red edge, and near-infrared spectral regions demonstrate heightened sensitivity to plant photosynthetic pigments and canopy structure features [13,23]. The OHS images had 32 bands and 5 red-edge bands (682–750 nm) that could be used to construct VI_REs. Based on existing literature, five distinct groups of VIs were selected as prototypes for the further design of Red-based vegetation indices (VI_Rs) and VI_REs: the NDVI, RVI, EVI, SAVI, and adjusted vegetation index (ARVI) [39,42]. The Hyperspectral VIs, formulated as combinations of two or three specific bands extracted from OHS-3 images, are listed in

Table 2. Formulas used for calculating the hyperspectral vegetation index in this study.

Vegetation Index (VI) Name	Red-Based Vegetation Index Formula (VIRs)	Red-Edge-Based Vegetation Index Formula (VIREs)
Normalized difference vegetation index (NDVI)	${NDVI}_{Rs}={\displaystyle \frac{NIRi-Rj}{NIRi+Rj}}$	${NDVI}_{REs}={\displaystyle \frac{NIRi-REj}{NIRi+REj}}$
Ratio vegetation index (RVI)	${RVI}_{Rs}={\displaystyle \frac{NIRi}{Rj}}$	${RVI}_{REs}={\displaystyle \frac{NIRi}{REj}}$
Soil adjusted vegetation index (SAVI)	${SAVI}_{Rs}={\displaystyle \frac{({NIR}_i-R_j)(1+L)}{NIRi+Rj+L}}$	${SAVI}_{REs}={\displaystyle \frac{({NIR}_i-{RE}_j)(1+L)}{NIRi+REj+L}}$
Adjusted vegetation index (ARVI)	${ARVI}_{Rs}={\displaystyle \frac{{NIR}_i-(R_j-\gamma \left({\mathrm{B}}_n-R_j\right))}{{NIR}_i+(R_j-\gamma \left({\mathrm{B}}_n-R_j\right))}}$	${ARVI}_{REs}={\displaystyle \frac{{NIR}_i-({RE}_j-\gamma \left({\mathrm{B}}_n-{RE}_j\right))}{{NIR}_i+({RE}_j-\gamma \left({\mathrm{B}}_n-{RE}_j\right))}}$
Enhanced vegetation index (EVI)	${EVI}_{Rs}=2.5{\displaystyle \frac{({NIR}_i-R_j)}{({NIR}_i+6R_j-7.5{\mathrm{B}}_{\mathrm{n}}+1)}}$	${EVI}_{REs}=2.5{\displaystyle \frac{({NIR}_i-{RE}_j)}{({NIR}_i+6{RE}_j-7.5{\mathrm{B}}_{\mathrm{n}}+1)}}$

Notes: NIR, R, B, and RE are the band types of OHS images (Table 1); NIR_i is the reflectance of B20–B32 bands; R_j is the reflectance of red bands (B12–B14); RE_j is the reflectance of red-edge bands (B15–B19); and B_n is the reflectance of blue bands (B1–B3). L is a coefficient taken as 1.1 in ARVI, and the soil condition in SAVI was taken as 0.5.

. A total of 936 candidate indices generated from band combinations were used for correlation analysis to identify sensitive bands and indices.

2.4. Model Construction Between VIs and LAI of MBFs

2.4.1. Model Variable Screening Method Using Correlation Analysis

The Pearson correlation coefficient (R) is often used to select variables that are sensitive to LAI variation, the values of which range from −1 to 1 [43]. Here, we first used the correlation analysis to calculate all R values between the LAI of MBFs and hyperspectral VIs (i.e., five VI_Rs and five VI_REs), and then screened important VIs for constructing the LAI inversion model for the winter MBFs growth stage. Variables were screened using R, which was calculated using the following formula:

$$R=C\mathrm{or} (\mathrm{x},\mathrm{y})={\displaystyle \frac{Cov(x,y)}{\sqrt{Var(\mathrm{x})\times Var(\mathrm{y})}}},$$

(1)

where Cov(x, y) is the covariance of x and y, and Var(x) and Var(y) are variances of x and y, respectively.

2.4.2. Constructing LAI Models Using the Empirical Statistical Method Based on Single-VI

To evaluate the relationship between the LAI and hyperspectral VIs of MBFs during the winter growth stage, the three VIs with the highest correlation coefficients were selected. Each of these indices was used as an independent variable (x) to develop an empirical regression model for predicting LAI (y) [14].

Table 3. Univariate linear regression models.

Model Number	Regression Modeling Method	Regression Equation
M1	Linear regression model	$\mathrm{y}=\mathrm{a}+\mathrm{b}\cdot x$
M2	Quadratic polynomial model	$\mathrm{y}=\mathrm{a}+b\cdot x+\mathrm{c}\cdot x^2$
M3	Exponential model	$\mathrm{y}=a\cdot {\mathrm{e}}^{bx}$
M4	Power model	$\mathrm{y}=a\cdot x^b$
M5	Logarithmic model	$\mathrm{y}=\mathrm{a}+\mathrm{b}\cdot \ln (x)$

Note: a, b, and c are model parameters in regression equations.

lists the univariate linear regression modeling methods used in this study.

We established five empirical statistical models: linear, quadratic polynomial, exponential, power, and logarithmic. Based on the analysis of the regression equations, we found that the VI_Rs of NDVI_RE, RVI_RE, and SAVI_RE did not improve the R² values in empirical statistical models; however, EVI_RE and ARVI_RE significantly enhanced the model fitting accuracy. Therefore, we selected multivariate VIs for constructing the LAI inversion model based on the optimal univariate linear regression results.

2.4.3. Constructing LAI Models by Combining Machine-Learning Algorithms with Hyperspectral VIs

To assess the comparative efficacy of the optimized Red-edge VI versus Red VI combinations in estimating the LAI, we used the results from the preferred univariate empirical statistical model (Section 3.2). Subsequently, we screened the VIs and formulated various combination schemes for sensitive VIs (

Table 4. Combination schemes of vegetation indexes (VIs).

Combination Code Name	Combination VI	Description of combinations
Fa1	VIR1, VIR2, VIR3	First three VIRs with highest sensitivity to LAI
Fb1	VIRE1, VIRE2, VIRE3	First three VIREs with higher sensitivity to LAI
Fab1	VIRE1, VIRE2, VIRE3, VIR1	First three VIREs and first VIR with highest sensitivity to LAI
Fab2	VIRE1, VIRE2, VIRE3, VIR1, VIR2	First three VIREs and first two VIRs with higher sensitivity to LAI
Fab3	VIRE1, VIRE2, VIRE3, VIR1, VIR2, VIR3	First three VIREs and first three VIRs with higher sensitivity to LAI

).

Machine-learning algorithms, such as SVM, PSO-SVM, RF, XGBoost, and partial least squares regression (PLSR), were used to construct the LAI estimation model for MBFs in the winter shoot growth period.

SVM and PSO-SVM Machine-Learning Algorithms

SVM is a binary model within the machine learning domain, widely recognized for its robust performance in classification and quantitative estimation tasks, particularly in scenarios with limited sample sizes [44]. The optimal SVM is shown in Equation (2) [34].

$$f(x)=sign\left[{\displaystyle \sum _{i=1}^l{y}_i{\alpha }_{i}K\left(x_i,x\right)+b}\right]$$

(2)

The PSO algorithm was employed to optimize SVM’s hyperparameters with an improved search strategy [36], which is a heuristic global optimization algorithm based on iterative optimization search. The PSO-SVM process is shown in Figure 3. The PSO algorithm first initialized a group of particles with two attributes, i.e., velocity (vi) and position (xi), and then iterated in the solution space several times; the particles update themselves according to the two extreme values (p, g) to obtain the optimal solution in the spaces, and the velocity update equations for the PSO algorithm are listed in Equations (3) and (4) [32]:

$$v_{ih}\left(\mathrm{t}+1\right)=\mathsf{\omega }{v}_i\left(t\right)+C_1{R}_1[P_{{\phi }_{ih}}\left(t\right)-X_{ih}\left(t\right)]+C_2{R}_2[P_{g_h}\left(t\right)-X_{ih}\left(t\right)], \mathrm{and}$$

(3)

$$x_{ih}\left(t+1\right)=x_i\left(t\right)+v_i(t+1),$$

(4)

where V_ih(t) is the velocity of the ith particle in dimension h at the tth iteration; x_ih(t) denotes the optimal position in the space where the history of the ith particle h is found; t is the number of iterations; $\omega$ is the inertia weight; $C_1\mathrm{and} C_2$ are learning factors; $R_{1 }{\mathrm{a}\mathrm{n}\mathrm{d} R}_2$ are random numbers from 0 to 1; $P_{{\phi }_{ih}}$ is the neighborhood best position; and $P_{g_h}$ is the global best position.

When the adaptive capacity of a particle exceeded its self-optimal adaptation range, its position vector remained unchanged. If the adaptation ability of this particle exceeded the existing global optimal adaptation, its position vector should be updated to the global optimum. Initial values and the position and velocity of each dimension were randomly generated for each group. In the present study, the penalty factor was initially set to a value of 1.5, the maximum number of iterations was set to 50, and the initial value of the number of populations was 20. The initial values for the hyperparameters of the SVM model were C = 81, Gamma = 3, and Kernel = RBF.

RF, PLSR, and XGBoost Algorithms

The RF algorithm is a supervised learning method that uses a stochastic decision tree integrated learning technique, making it applicable to estimating agroforestry vegetation parameters based on remote sensing [43]. The XGBoost algorithm is a novel decision tree algorithm that can prevent overfitting by training and processing the classification and regression tasks of weakly supervised learning in machine learning while maintaining optimal computational efficiency through simplification and regularization to reduce the computational cost [44,45]. PLSR is a novel expression of multiple linear regression that integrates correlation analysis, principal component analysis, and multiple linear regression and can effectively reduce the redundancy of feature variables and remove the problem of covariance [46]. In the present study, SVM, XGBoost, the RF algorithm, and PLSR from the scikit-learn machine-learning library in Python 3.4 were used to construct the LAI inversion model of MBFs.

2.5. Model Evaluation

We randomly divided the Moso bamboo sample dataset for model training (70%) and testing (30%). The determination coefficient (R²) and root mean squared error (RMSE) were selected to assess the accuracy of the inversion model in estimating LAI, with higher R² and lower RMSE values indicating greater precision of the model simulation or predictive capabilities [47]. Evaluation indices were calculated using Equations (5) and (6).

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

(5)

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

(6)

3. Results

3.1. Screening VIs to Estimate the LAI of MBFs

The mean LAI value of the measured MBF samples was 3.21 (ranging from 1.08 to 6.40) with a standard deviation of 1.10 and a coefficient of variance of 34.2% (

Table 5. Descriptive statistics of measured LAI in sample plots for MBFs.

Sample Size	Mean	Minimum	Maximum	Standard Deviation	Coefficient of Variance (%)
64	3.21	1.08	6.40	1.10	34.27

). During the winter growth stage, MBFs with high LAI values presented dense green foliage, indicating that underground bamboo whips are likely to yield winter shoots (Figure 1d—High LAI and Winter shoot), and the other bamboo forests with lower LAI values presented light green or yellow foliage without winter bamboo shoots (Figure 1d—Low and Medium LAI values).

MBF loaded on the LAI affects the canopy spectral reflectance, and the VI from the optical satellite sensor image changes accordingly. Correlation analysis results between the LAI of MBF and the VI_S are shown in Figure 4. The VIs of the two band combinations (Figure 4a–c) mostly showed significant positive correlations with the LAI (R > 0.6), and the red-band-based VIs (NDVI_R, RVI_R, and SAVI_R) were significantly higher than the red-edge-based VIs (NDVI_RE, RVI_RE, and SAVI_RE). The highest correlation coefficient within the NDVI_S was for the combination of the 780-nm near infrared band (B22) and 640-nm red band (B12) (R = 0.777, p < 0.01). Nevertheless, the highest correlation coefficient in the NDVI_REs was the combination index of B22 and the 686 nm red-edge band (B15) (R = 0.721, p < 0.01). The highest correlation coefficients of RVIs were 0.783 for the RVI_R (B22 and B12) and 0.726 for RVI_RE. Similar results were found for SAVI, with the highest correlation coefficient of 0.778, and the correlation coefficient of SAVI_RE with the combination of B15 and B22 was 0.725. In Figure 4d(B1–B3), the relationship among EVI_R, EVI_RE, and LAI_M shows a significant positive correlation (p < 0.01); the EVI (B25, B12, and B2) had a great correlation coefficient (R = 0.784) within the various EVIs; and the correlation coefficients for EVI_RE indices of the participating combinations in the red-edge band of B15 or B16 (EVI_RE80 and EVI_RE83) were 0.785 and 0.786, respectively. This suggests that the combination of bands at 686, 709 or 833, and 466 nm for EVI_RE was sensitive to LAI variations. However, Figure 4e shows that the highest correlation coefficients of ARVI_R and ARVI_RE with LAI_M were −0.406 and −0.600, respectively, which were lower than the preferred VIs (NDVI, RVI, SAVI, and EVI).

3.2. Univariate Empirical Model Construction and Selection for LAI Estimation

Variance analysis and correlation analysis were conducted to screen the single vegetation index for constructing an LAI univariate empirical model for Moso bamboo. The first 12 top-rank optimal equations were selected (

Table 6. Univariate empirical models of LAI based on the preferred VI (ranked top 12). R²—determination coefficient; RMSE—root mean square error.

Optimal Model No. *	Screening Vegetation Index (x)	Optimal Bands	Optimal Prediction Equation	Training Dataset		Test Dataset
				R2	RMSE	R2	RMSE
M3.1	NDVIR7	B22, B12	y = 0.795e2.834x	0.622	0.602	0.422	0.913
M3.2	NDVIR10	B24, B12	y = 0.832e3.052x	0.605	0.615	0.347	1.134
M4.1	NDVR16	B25, B12	y = 8.096x1.212	0.617	0.614	0.453	0.900
M4.2	RVIR10	B23, B12	y = 0.944x1.146	0.609	0.607	0.443	0.996
M2.1	RVIR16	B25, B12	y = 0.058x2 + 0.97x + 0.058	0.604	0.602	0.415	0.990
M2.2	RVIR7	B22, B12	y = 0.036x2 + 0.935x + 0.151	0.615	0.615	0.427	1.102
M2.3	EVIRE77	B25, B17, B2	y = 2.471x2 + 4.75x + 1.114	0.615	0.593	0.472	0.940
M2.4	EVIRE80	B25, B16, B2	y = −0.758x2 + 6.097x + 1.126	0.617	0.593	0.474	0.938
M2.5	EVIRE83	B25, B15, B2	y = −2.235x2 + 6.464x + 1.244	0.624	0.587	0.482	0.932
M3.3	SAVIR13	B24, B12	y = 0.838e2.049x	0.606	0.614	0.440	0.908
M4.3	SAVIR16	B25, B12	y = 5.009x1.205	0.618	0.612	0.518	0.899
M4.4	SAVIR7	B22, B12	y = 5.237x1.095	0.606	0.624	0.525	0.916

Notes: y denotes the measured LAI value and x denotes the prediction variable. * ranked top 12 empirical models based on screening vegetation indices. Preferred univariate inversion models are in bold font.

). The results showed that the relationship between the preferred spectral VIs and the LAI of MBF was best described by a robust nonlinear function. Among the VIs tested, the red-edge enhanced VI (EVI_RE83) performed the best in the LAI inversion model. This model used a quadratic polynomial function with an R² value of 0.624 and an RMSE of 0.587. Additionally, the VIs such as NDVI_R7, NDVI_R13, and SAVI_R13 were used to construct optimal LAI estimation models based on an exponential function, with R² values ranging from 0.605 to 0.622. LAI inversion models employing NDVI_R16, RVI_R10, SAVI_R16, and SAVI_R7 showed a high accuracy using the power function, with R² values ranging from 0.606 to 0.618.

The preferred LAI univariate models developed using six optimized vegetation indices demonstrated relatively high fitting accuracy with R² > 0.615. EVI_RE83 was the optimal predictor variable within the different single VIs, and the others were NDVI₇, SAVI₁₆, EVI_RE80, NDVI_R16, RVI_R7, and EVI_RE77, with R² values of 0.622, 0.618, 0.617, 0.617, 0.617, and 0.615, respectively. Evaluation metrics from the test data revealed that the R² values of the M2.3, M2.4, M2.5, M4.3, and M4.4 estimation models were 0.472, 0.474, 0.482, 0.518, and 0.525, respectively. When considering both the training and test data evaluation indices (R² and RMSE), the red-edge VIs (EVI_RE77, EVI_RE80, and EVI_RE83) achieved high fitting accuracy with M2.3, M2.4, and M2.5 models; red-based VIs (SAVI_R16, NDVI_R7, and NDV_R16) demonstrated relatively great fitting accuracies with M4.3, M3.1, and M4.1 models.

3.3. Machine-Learning Model Based on Multivariate VIs

To evaluate the performance of the multiple-variable combinations of VI_RES and VI_RS in the LAI estimation model, five combination scenarios were selected to construct the multivariate LAI inversion models, with the results presented in

Table 7. Accuracy comparison of LAI inversion models based on combinations of machine-learning algorithms and multivariate VIs.

Machine-Learning Algorithm	CFab3 #		CFab2		CFab1		CFa1		CFb1
	R2	RMSE	R2	RMSE	R2	RMSE	R2	RMSE	R2	RMSE
PSO-SVM	0.721	0.490	0.719	0.493	0.715	0.498	0.712	0.501	0.684	0.520
SVM	0.470	0.943	0.466	0.945	0.497	0.918	0.379	1.020	0.519	0.898
RF	0.445	0.964	0.446	0.964	0.436	0.972	0.492	0.923	0.360	1.036
XGBoost	0.418	0.987	0.424	0.982	0.421	0.985	0.396	1.006	0.368	1.029
PLSR	0.501	0.915	0.500	0.915	0.518	0.984	0.334	1.056	0.436	0.972

Notes: # indicates multivariable combination schemes; Fa1 was the preferred combination scheme of 3EVI_REs (EVI_RE83, EVI_RE80, and EVI_RE77); Fb1 was the preferred combination scheme of three non-VI_RES (NDVI_R7, SAVI_R16, and NDVI_R16); Fab1 was the combination of 3VI_RES with one optimal non-red-edged index (NDVI_R7); Fab2 was the combination of 3VI_RES with two optimal VI_RS (NDVI_R7, NDVI_R16); and Fab3 was the combination of 3VI_RES with 3VI_RS.

. The integrated approach of the PSO-SVM algorithm demonstrated effective nonlinear generalization capabilities for estimating the LAI of MBFs. Specifically, the PSO-SVM model outperformed the other four classic machine-learning models (SVM, RF, XGBoost, and PLSR) in terms of accuracy. The combination of optimal 3VI_REs and 1VI_Rs noticeably improved the model fitting accuracy; for example, the R² of the PLSR-Fab1 model improved from 0.334 to 0.518. The PSO-SVM-Fab3 model had the highest fitting accuracy (R² = 0.721, RMSE = 0.523) for LAI estimation, which was compared with the ordinary SVM-Fb1 model (R² = 0.519, RMSE = 0.898); the R² value increased by 39.19% and the RMSE decreased by 0.408. These results indicated that the PSO-SVM model had a nonlinear generalization ability for the LAI inversion model and effectively exhibits self-learning and adaptive abilities, which to a certain extent overcame the covariance problem between the independent variables.

3.4. Model Evaluation and LAI Mapping for Moso Bamboo in the Winter Growth Stage

Figure 5 presents a comparison of test accuracy of the preferred models using different modeling methods. The combined model PSO-SVM-Fab3 exhibited the highest accuracy, with an R² value of 0.715, which, compared with the SVM-Fb1 and optimal univariate empirical (M2.5) models, improved R² values by 15.54% and 39.19%, respectively. These results indicate that the PSO effectively optimized the hyperparameters of the SVM model for LAI, significantly enhancing the evaluation metrics of the PSO-SVM model, which effectively predicted changes in the Moso bamboo LAI. Additionally, the EVI_RE83 constructed with the 686 nm red-edge band and 833 nm band explained 48.2% of the variance in the empirical statistical model (see Table 6 for R² values). Figure 5 illustrates the relationship between the predicted and field-measured LAI values for winter MBFs. Validation samples for the PSO-SVM-Fab3 showed a linear distribution and demonstrated high levels of the prediction accuracy and stability of the LAI estimation.

Figure 6 displays the LAI mapping of Moso bamboo during the winter growth period, based on the optimal LAI inversion model of PSO-SVM_Fab3. The LAI mapping results align well with field observations. The region with higher LAI prediction value was mainly distributed in the Moso bamboo intensive management productive zones at high elevations, such as the Shangping and Longgong villages. The LAI values of MBFs mainly ranged from 2.1 to 5.5. In the LAI thematic mapping, the values exceeding 4.5 in MBFs are indicative of high winter bamboo shoot yields during the winter growth stage. As illustrated in Figure 6(b2), the sample MBF with a high LAI value of 6.1 characterized by dense green canopy leaves exhibited high winter bamboo shoot production. In contrast, Figure 6(b1) reveals a sample MBF with bright green leaves and a low LAI value of 3.3, indicating no winter bamboo shoot harvesting.

4. Discussion

4.1. Effect of Hyperspectral VI on LAI Estimation

Selecting a suitable hyperspectral VI is crucial for vegetation LAI inversion modeling based on hyperspectral imagery [23,48]. In the present study, correlations between the Moso bamboo LAI and VIs were mostly positive and highly associated with the classic two-band combined VIs (NDVIs, RVIs, and SAVIs), which utilize the red band at a shorter wavelength (Figure 4). The strongest correlation was observed for the two-band combination (B22 and B12), followed by the combination (B25 and B12) across NDVI, RVI, and SAVI. Among the three-band combined VIs, the red-edge-band EVI_RE involved the red-edge bands (B15, B16, and B17) and had a high correlation coefficient (R > 0.783). The EVI_RE outperformed EVI_R in estimated LAI, exhibiting greater consistency and accuracy, which confirmed the usefulness of adding a blue band to weaken the effect of atmospheric scattering in monitoring LAI [47]. The LAI inversion model accuracy was obviously related to selecting the hyperspectral index (Table 6). The EVI_RE had a substantial impact on LAI estimation, achieving an optimal R² value of 0.624. These findings align with previous research; e.g., Yi reported that the LAI retrieval model based on the red-edge EVI of Sentinel-2 images exhibited the optimal performance [49]. Other studies have confirmed that red-edged VIs have excellent LAI estimation outcomes, and VI_RE from Sentinel-2A or Gaofen-6 multispectral imagery improves the accuracy of LAI inversion [19,32]. Dong et al. [24] explored the potential of VIs for crop LAI estimation and found that VI_RE was mainly influenced by chlorophyll. However, the broader bandwidth differences of satellite sensors are more sensitive to atmospheric contamination because of scattering or absorption of interference [47,50]. Therefore, integrating improved VI_REs based on Sentinel-2 or Gaofen-6 data with optical hyperspectral imagery (OHS) represents a promising direction for future comparative studies.

Figure 6 and Figure 7 present the LAI inversion results for the two sampled bamboo forest stands within this study. The LAI mapping results based on SAVI_R16 and NDVI_R7 closely matched the actual LAI measurements obtained from field surveys. This alignment is particularly evident in intensively managed MBFs, where the understory environment lacks vegetation cover, and bare topsoil is exposed. In such cases, SAVI-based LAI inversion results are more consistent with field observations. However, the interaction between canopy LAI and chlorophyll content, combined with the understory environment, suggests that a comprehensive index incorporating more sensitive hyperspectral band regions may be necessary to enhance the applicability of OHS data in LAI inversion. This approach could improve the robustness and accuracy of LAI estimation in complex forest environments.

4.2. Measures of Modeling LAI in Empirical and Machine-Learning Models

An empirical model based on the VI is an effective method for estimating LAI in agricultural and forestry management. Comparisons based on single-variable empirical regression showed that the quadratic equation had the highest fitting accuracy, with an R² value of 0.624. As shown in Table 6, exponential functions had higher accuracy for the LAI model when using the preferred NDVI_R7. For the optimal LAI inversion model, SAVI_R16 was best represented by a power function, with an R² value of 0.618. The explanatory ability of these VIs for LAI variation exceeded 52.7% (Figure 8). The parameters of the LAI empirical models may be influenced by the heterogeneous vegetation growth environment, leaf growth phenology, and sample size. The optimal fitting equation can vary among different tree species or crops, and even for the same vegetation species in different regions, due to differences in leaf phenology caused by plant growth [22,30]. For example, Li et al. [51] found that a power function based on RVI was the optimal model for peanut LAI estimation, while Li et al. [52] used an exponential function to fit the relationship between NDVI_RE and LAI.

The application of machine-learning algorithms significantly affected the LAI modeling of winter MBFs. As shown in Table 7 and Figure 5, the PS0-SVM model more accurately simulated LAI than traditional methods, with significantly improved model accuracy; with the combination of multiple heterogeneous VIs, the estimation precision of the model significantly improved (in the training set, R² improved from 0.624 for the optimal univariate model to 0.721 for the optimal multivariate model, and R² improved from 0.559 to 0.715 in the validation test dataset). Previous studies have consistently shown that LAI inversion models integrating multivariate approaches generally improve prediction accuracy compared to univariate models [31,53]. These results are consistent with those of He et al. [50], indicating that the PSO-SVM hybrid model exhibited superior generalization capabilities compared to those of the traditional SVM inversion model. The performance of SVM is highly dependent on the selection of its parameters, such as the penalty parameter C and the kernel function parameter γ [23,54]. Traditional methods, such as grid search and cross-validation, are commonly used but are time-consuming and inefficient [53]. In contrast, PSO offers a more efficient approach by dynamically adjusting the inertia weight and learning factors to balance global search and local search capabilities, thereby accelerating convergence [55]. The PSO-SVM algorithm integrates the global search ability of PSO with the classification or modeling performance of SVM, addressing the challenges of parameter optimization in traditional SVM methods and significantly enhancing model performance and applicability [56]. By traversing an empirical interval to search for optimal hyperparameters based on dataset characteristics [53], the PSO-SVM algorithm improves model predictive performance, as confirmed by Cai et al. [57]. Additionally, the uncertainty errors in the PSO-SVM model were lower than those in the traditional machine-learning algorithm models. We compared the optimal LAI SVM models optimized by three hyperparameter optimization algorithms. As shown in

Table 8. Accuracy comparison of LAI estimating models based on hyperparameter optimization algorithms and SVM.

Optimization Models	R2	RMSE
PSO + SVM	0.721	0.490
Bayesian + SVM	0.684	0.678
Grid search + SVM	0.647	0.717

, the optimal PSO-SVM LAI retrieval model outperformed the Bayesian and Grid search hyperparameter optimization algorithms for the SVM approach in terms of accuracy (R² = 0.721, RMSE = 0.645 for the PSO-SVM model vs. R² = 0.684, RMSE = 0.678 for the optimal Bayesian + SVM model, and vs. R² = 0.647, RMSE = 0.717 for the optimal Grid search + SVM model).

4.3. LAI Values and Management Implications for MBFs

MBFs are typically uneven-aged mixed forests with weaker photosynthesis in winter. The unique off-year and on-year leaf phenology in winter is critical for assessing winter bamboo shoots and utilization [58,59]. The LAI in winter MBFs is key to understanding their ecosystem structure and function. Previous studies found a relationship between LAI and bamboo shoot yield. Pan et al. reported that a higher LAI is closely related to the bamboo shoot yield, with the highest yield at an LAI of 10.339 for Phyllostachys pubescens [60]. In this study, bamboo forest plots with a winter LAI greater than 4.5 had winter bamboo shoot harvests. However, the specific LAI threshold for bamboo shoots and LAI variation in winter MBFs are likely closely related to bamboo species and management measures. The implications of mapping LAI values on management are substantial for winter MBFs. For areas with particularly low LAI, fertilization and irrigation should be strengthened to promote LAI and chlorophyll accumulation. If the LAI is much lower than the normal status of MBFs, the forest may have pests (e.g., Pantana phyllostachysae Chao) that affected LAI, so pest control should be strengthened [43,61]. In regions with a high LAI, avoid destroying bamboo rhizomes during winter and spring shoot excavation. Selective thinning can be carried out after the spring shoot growth period the following year to reduce competition and conserve nutrients. Proper management promotes healthy leaf growth and replacement in MBFs, leading to better bamboo quality and higher yields.

4.4. Limitations and Scope for Future Studies

Data sources of different spatial resolution remote-sensing images have significant effects on plant LAI inversion modeling [62]. For instance, Liu et al. found that the LAI inversion accuracy of Sentinel-2 images (10 m) was higher than those of Landsat-8 (30 m) and GF-2 (4 m) [63]. Additionally, studies have shown that integrating multi-source remote sensing data, including high-resolution aerial and/or lidar data, can enhance the detection of fine-scale spatial patterns of LAI in forest ecosystems [3,64]. Hence, comparative studies should be conducted to explore LAI inversions using different remote-sensing data sources. In the present study, the LAI survey was conducted in Winter, 2023, and the developed inversion model was only suitable for Moso bamboo during Winter growth stages. Therefore, the results may not reflect present conditions or temporal changes in the LAI of MBFs. In addition, the leaf status of MBFs differs in different growth seasons, and there are related leaf phenology changes between on- and off-year bamboo forests [8,24]. The present study did not discuss the potential effects of disturbances or management practices on the LAI inversion model. Therefore, considering that the MBF distribution is mostly in the central subtropical rainy zone, it is imperative that future study areas be carefully selected for comparative studies. Although the current findings were based on incomplete information, they should be understood within the constraints of these limitations, particularly when conducting investigative studies on intricate mountainous terrains. For accurate LAI mapping over large regions to monitor Moso bamboo growth status, future research should focus on addressing these limitations to enhance the practicality and accuracy of the developed models.

5. Conclusions

The present study introduced a novel method integrating Zhuhai-1 OHS imagery with the PSO-SVM coupling model to estimate MBF LAI in winter. Our approach encompassed screening hyperspectral VIs, including red-based VIs (NDVI_R7 and SAVI_R16) and a red-edge-based enhanced VI (EVI_RE83), which were sensitive to the LAI of MBF in the winter growth stage. The findings revealed a high correlation between the ground-measured and predicted LAI values in optimal empirical models. The hybrid method integrating PSO-SVM and multivariate VIs achieved higher precision in LAI prediction than the empirical models and traditional machine-learning algorithm models for MBFs. The optimal PSO-SVM_Fab3 model was proven to be effective in mapping high-resolution LAI products, with an accuracy (R²) of 0.715 and RMSE of 0.591. These results are promising for the provision of a decameter LAI dataset, which can enable more accurate fine-scale ecosystem modeling and precision bamboo forestry applications. The application of extensive dynamic OHS images enables the dynamic evaluation of the LAI status in bamboo forests, offering decision-making support to enhance the management quality of bamboo forests.