Tree Species Classification for Shelterbelt Forest Based on Multi-Source Remote Sensing Data Fusion from Unmanned Aerial Vehicles

Abstract

Accurately understanding the stand composition of shelter forests is essential for the construction and benefit evaluation of shelter forest projects. This study explores classification methods for dominant tree species in shelter forests using UAV-derived RGB, hyperspectral, and LiDAR data. It also investigates the impact of individual tree crown (ITC) delineation accuracy, crown morphological parameters, and various data sources and classifiers. First, as a result of the overlap and complex structure of tree crowns in shelterbelt forests, existing ITC delineation methods often lead to over-segmentation or segmentation errors. To address this challenge, we propose a watershed and multi-feature-controlled spectral clustering (WMF-SCS) algorithm for ITC delineation based on UAV RGB and LiDAR data, which offers clearer and more reliable classification objects, features, and training data for tree species classification. Second, spectral, texture, structural, and crown morphological parameters were extracted using UAV hyperspectral and LiDAR data combined with ITC delineation results. Twenty-one classification images were constructed using RF, SVM, MLP, and SAMME for tree species classification. The results show that (1) the proposed WMF-SCS algorithm demonstrates significant performance in ITC delineation in complex mixed forest scenarios (Precision = 0.88, Recall = 0.87, F1-Score = 0.87), resulting in a 1.85% increase in overall classification accuracy; (2) the inclusion of crown morphological parameters derived from LiDAR data improves the overall accuracy of the random forest classifier by 5.82%; (3) compared to using LiDAR or hyperspectral data alone, the classification accuracy using multi-source data improves by an average of 7.94% and 7.52%, respectively; (4) the random forest classifier combined with multi-source data achieves the highest classification accuracy and consistency (OA = 90.70%, Kappa = 0.8747).

1. Introduction

Tree species diversity is one of the indicators for evaluating the stand structure of shelterbelts. The information on tree species categories and attributes is of great significance for constructing and monitoring shelterbelt ecosystems, providing a basis for planners to monitor, manage, and assess shelterbelts, thereby ensuring their roles in windbreak, sand fixation, and farmland protection [1]. Accurate tree species classification helps optimize the ecological functions of shelterbelts by ensuring the appropriate selection and management of species. It enables forest managers to make informed decisions about species conservation, regeneration, and resource management, thereby contributing to the long-term resilience and productivity of shelterbelt ecosystems. Therefore, utilizing new technological systems and model methods to conduct fine-scale monitoring of shelterbelts and acquiring timely and accurate knowledge of stand composition is of paramount importance in the construction and benefit evaluation of shelterbelt projects. In recent years, unmanned aerial vehicle (UAV) remote sensing platforms have been proven to be highly effective in global applications due to their advantages of real-time acquisition and multi-scale, repetitive monitoring [2]. Various types of flight platforms and remote sensing data, such as visible light, hyperspectral, and LiDAR, have been widely used in forestry surveys. Researchers at home and abroad have conducted extensive studies on individual tree crown (ITC) delineation and species classification using machine learning, deep learning, and other methods based on different data sources [3,4,5], greatly enriching the application scenarios.

In tree species classification using remote sensing data, ITC delineation is a crucial foundational step [6]. Effective ITC delineation can provide the necessary data support for subsequent classification tasks [7]. Currently, most studies on tree species classification based on ITC delineation use the Canopy Height Model (CHM) derived from LiDAR data to segment tree crown boundaries, followed by classification using the spectral characteristics of vegetation [8,9]. Among these, the watershed segmentation algorithm based on CHM [10] has gained wide recognition and application in the field of ITC delineation due to its efficiency, accuracy, and strong adaptability, particularly in scenarios with low-density stands and structurally simple and uniform segmentation [11,12]. However, although the watershed segmentation algorithm performs excellently in handling fine boundaries, it is sensitive to noise and prone to over-segmentation issues [13,14]. Spectral clustering algorithms [15], on the other hand, construct graphs representing the similarity between regions, analyze and identify the global structure and intrinsic patterns of the data, and can segment regions into clusters with maximum intra-cluster similarity and minimum inter-cluster similarity. Therefore, using a spectral clustering algorithm as a post-processing step for the watershed segmentation algorithm can reasonably merge over-segmented regions, effectively mitigating the common issue of over-segmentation during the watershed segmentation process. Although the combination of watershed segmentation and spectral clustering processes has been employed in ITC delineation to improve delineation accuracy [16], this approach still faces numerous challenges in complex mixed forest scenarios where tree crowns overlap and canopy structures are intricate [17]. Specifically, in dense tree crowns, complex canopy morphologies, and increased noise, the watershed segmentation algorithm may struggle to accurately extract the contours of individual trees, leading to errors and blurring in tree crown delineation boundaries. Additionally, while the spectral clustering algorithm can optimize delineation results and reduce over-segmentation, it may also result in mis-segmentation or under-segmentation in overly dense and complex canopy structures. Therefore, this study addresses the aforementioned issues and shortcomings by proposing a watershed and multi-feature-controlled spectral clustering (WMF-SCS) algorithm. This algorithm combines various image enhancement methods with the watershed segmentation algorithm to reduce image noise and enhance edge features, thereby better highlighting the details of tree crown boundaries. Additionally, by leveraging the feature similarity of the same tree, multiple features of RGB and CHM data are introduced as control conditions in the spectral clustering algorithm, fully utilizing the precise boundary detection capability of watershed segmentation and the global structure optimization capability of spectral clustering.

Numerous studies have demonstrated the significant potential of multi-source data in improving tree species classification accuracy. Researchers have conducted extensive studies on the integration and feature extraction of UAV visible light, multispectral, hyperspectral, and LiDAR data [18,19,20,21]. Wu et al. [22] classified ten dominant tree species in urban areas using high-resolution RGB images and LiDAR data acquired by UAVs. By extracting various features such as spectral, texture, and vegetation indices, and combining them with the RF classifier, their study found that integrating LiDAR and RGB data significantly improved classification accuracy. Compared to using RGB data alone, the overall accuracy (OA) and Kappa coefficient increased by 18.49% and 0.22, respectively. Zhong et al. [23] performed ITC delineation using LiDAR data and tree species classification using the support vector machine (SVM) algorithm combined with hyperspectral data. The results showed that the classification OA using fused hyperspectral and LiDAR features was the highest, reaching 89.20%, which was higher than using hyperspectral features alone (86.08%) and LiDAR features alone (76.42%). Rina et al. [24] classified birch, larch, and non-forest areas using UAV multispectral remote sensing images and LiDAR point cloud data. The results indicated that the highest OA reached 86%, and after introducing CHM, the classification accuracy improved by 7%. Wang et al. [4] designed multiple classification images using UAV LiDAR and hyperspectral data to compare the classification effects of different data sources and crown morphology features. The results showed that the classification accuracy of multi-source remote sensing data was superior to that of a single data source, and when crown morphology features extracted from LiDAR [25] (i.e., parameters describing crown shape, size, and structure) were added to the classifier, the accuracy of all classifiers improved. These studies demonstrate that applying multi-source data can provide more detailed and comprehensive information, enhancing the accuracy of tree species classification. However, differences exist between different data sources and study areas, and the complex stand structure in shelterbelts presents a challenge in fully utilizing spectral and spatial structural information to identify representative and distinguishable classification features.

In summary, this study proposes a watershed and multi-feature-controlled spectral clustering (WMF-SCS) algorithm, which is suitable for ITC delineation in complex mixed forest scenarios, based on UAV RGB data and LiDAR-derived CHM data. The goal is to improve the ITC delineation algorithm’s performance in complex scenarios, providing clearer and more reliable classification objects, features, and training data for tree species classification. Subsequently, based on UAV LiDAR and hyperspectral data combined with ITC delineation results, various classification images are constructed and compared, focusing on exploring the effectiveness of multi-source remote sensing data for tree species classification at the individual tree scale. The study also evaluates the potential application of ITC delineation and crown morphology parameters in tree species classification, aiming to provide a reference for the application of UAV multi-source remote sensing data in shelterbelt tree species classification.

2. Materials and Methods

2.1. Materials

2.1.1. Study Area Overview

The study area is located on the northern slope of the Tianshan Mountains in the Xinjiang Uygur Autonomous Region, at the southern edge of the Junggar Basin, within the Three-North Shelterbelt area of the 150th Regiment, Mosuowan Reclamation Area (45°10′ N, 85°56′ E, as shown in Figure 1). The terrain is flat, with an elevation ranging from 300 to 500 m, and the area experiences a continental warm temperate desert arid climate. The shelterbelt in the study area is a mixed forest with a vertical structure consisting of an arboreal layer, a shrub layer, and an herbaceous layer. The dominant tree species are Populus bolleana L., Ulmus pumila L., Elaeagnus angustifolia L., and Haloxylon ammodendron (C.A. Mey.) Bunge. Elm forests and mixed broadleaf forests are distributed along roadsides to serve the purpose of windbreak, sand fixation, and farmland protection. This study focuses on the classification of the dominant tree species, including Ulmus pumila, Elaeagnus angustifolia, Populus bolleana, and Dead trees.

2.1.2. Data Acquisition

The experimental LiDAR data were acquired using a UAV platform on 10 October 2023. The UAV platform was the DJI Matrice 300 RTK multirotor drone, produced by DJI in Shenzhen, China, equipped with an integrated Livox LiDAR module, a mapping camera, and the high-precision inertial navigation system ZENMUSE L1 LiDAR also manufactured by DJI in Shenzhen, China. The DJI M300 RTK was flown at a relative altitude of 50 m, with a horizontal accuracy of ±5 cm and a vertical accuracy of ±2.5 cm. The specific data parameters are given in

Table 1. LiDAR and UAV parameters.

Parameters	Values
Flight altitude/m	50
Flight speed/(m·s−1)	5.6
Lateral overlap rate/%	65
Course overlap rate/%	70
Echo times	2
Average density/(pts·m−2)	1618
Weather condition	Clear with few clouds

.

The experimental RGB data were acquired between 11:30 and 13:00 on 10 October 2023, using the DJI Matrice 300 RTK multirotor UAV platform, equipped with the ZENMUSE P1, a 45-megapixel full-frame image sensor produced by DJI in Shenzhen, China. This sensor is paired with the DJI DL 35 mm F2.8 LS ASPH Lens, with a focal length of 35 mm, also manufactured by DJI in Shenzhen, China. The relationship between Ground Sampling Distance (GSD) and shooting distance (L) is defined as GSD = L/80. The DJI M300 RTK was flown at a relative altitude of 300 m, with a flight speed of 13.7 m/s, and the gimbal angle was set perpendicular to the ground. The flight had a forward overlap rate of 75% and a side overlap rate of 75%, resulting in a GSD of 3.75 cm/pixel.

The experimental hyperspectral data were acquired between 13:30 and 14:30 on 10 October 2023, using the Matrice 600 PRO hexacopter UAV platform, produced by DJI in Shenzhen, China, equipped with the RONIN-MX multifunctional gimbal from DJI, Shenzhen, China, and a Rikola hyperspectral imager produced by SENOP in Kangasala, Finland. According to the data acquisition requirements, the spectral range was adjusted to 502 nm–903 nm, resulting in 45 spectral bands. The experiment was conducted under clear skies with minimal clouds, ensuring stable sunlight intensity and minimal shadow interference. The Matrice 600 PRO was flown at a relative altitude of 75 m, with a forward overlap rate of 75% and a side overlap rate of 75%, resulting in a GSD of 5 cm/pixel. The specific data parameters are given in

Table 2. Hyperspectral camera parameters.

Parameters	Values
Field of view/(°)	36.5
Focal length/nm	9
Spatial resolution/(cm·100 m−1)	6.5
Spectral range/nm	502~903
Amounts of band	45
Spectral resolution/nm	6
Image resolution/pixel	1010 × 1010
Weight/g	720

.

2.1.3. Data Preprocessing

For LiDAR data preprocessing, the first step involves flight strip stitching, trajectory line cropping, point cloud deduplication, and noise removal using KD-Tree-Based Gaussian Filtering (KDT-GF) on the raw point cloud data of the study area. Non-tree point clouds within the region are manually filtered out, and low points are separated. Next, the Improved Progressive TIN Densification (IPTD) algorithm embedded in the Lidar360 software, version 5.2, is used for ground point classification. A Digital Elevation Model (DEM) and a Digital Surface Model (DSM) with a resolution of 0.05 m are generated using the Inverse Distance Weighting (IDW) method. The CHM is then derived by subtracting the DEM from the DSM. Subsequently, ArcGIS software, version 10.3, is used to perform smoothing, hole filling, and data clipping to obtain the post-processed CHM. Finally, the LiDAR point cloud is height-normalized based on the plot’s DEM to eliminate the influence of terrain.

For RGB data preprocessing, aerial triangulation was performed using Pix4D Mapper software, version 4.5.6, to generate a point cloud dataset and a DSM. The images were then ortho-rectified, and a Digital Orthophoto Map (DOM) was produced.

For hyperspectral data preprocessing, the raw hyperspectral images underwent dark current correction, image format conversion, and radiometric correction. Using POS data containing image geolocation information, the images were automatically stitched in Agisoft PhotoScan, version 1.5.0. Based on the stitched images, a linear fitting model was established in ENVI using ground radiometric targets with reflectance values of 3%, 22%, 48%, and 64%, to convert radiance values to reflectance.

2.1.4. Image Registration

Image registration is a crucial step in tree species classification using multi-source remote sensing data fusion. It involves accurately aligning LiDAR, RGB, and hyperspectral data obtained from different sensors at different times to improve the precision and reliability of data fusion. The following steps are taken to ensure the effective fusion of different data sources within the same coordinate system, reduce positional errors, and enhance feature extraction. First, bilinear interpolation is used to downsample the RGB images to 5 cm/pixel, matching the spatial resolution of the LiDAR-derived CHM data and hyperspectral data. Then, using the geographic reference information between the images, landmark features such as gaps between buildings and trees are employed as references for relative image registration. Geometric correction is performed using affine transformation, with the registration error controlled within one pixel.

2.1.5. Field Survey Data

• Sample data for ITC delineation: Based on the distribution of the study area and target tree species, three regions were selected from the study area to evaluate the accuracy of ITC delineation. These regions represent the diverse terrain, vegetation types, and tree species distributions within the study area, with Site 2 characterized by highly variable and densely populated shelter forests. In total, they encompass 320 reference tree crown samples. The ground reference crowns were manually delineated by experienced professionals based on RGB images through visual interpretation, and were further verified and adjusted using hyperspectral and LiDAR point cloud data to ensure the accuracy of the tree crown boundaries.
• Sample data for tree species classification: A field survey was conducted in the study area on 10 October 2023. During the sampling process, a portable Real-Time Kinematic (RTK) device was used to record the geographic coordinates and tree species information of the samples. A total of 430 sample points were collected, including Populus bolleana, Ulmus pumila, Elaeagnus angustifolia, and Dead trees. The tree species data obtained from the entire study area were used as the training and validation sample set for tree species classification, while the ITC delineation data from three representative regions were used as the test sample set for tree species classification. The specific sample points and locations of the representative regions are shown in Figure 1.

2.2. Methods

The technical workflow of this study is illustrated in Figure 2. First, ITC delineation is performed using RGB data and LiDAR-derived CHM data, followed by an evaluation of ITC delineation accuracy. Next, spectral, texture, and structural features are extracted by combining hyperspectral, LiDAR data, and the ITC delineation results. Then, the features are ranked and filtered based on their importance, with redundant and irrelevant features removed, retaining those suitable for single-source and multi-source data classification. Finally, different classification images are constructed under the hyperspectral and LiDAR data sources for tree species classification, and classification accuracy is evaluated.

2.2.1. Individual Tree Crown Delineation

Due to the complex stand structure and dense canopy in mixed forests, the watershed segmentation algorithm based on CHM struggles to extract the contours of individual trees accurately in such segmentation scenarios, often leading to errors and blurring in tree crown delineation boundaries. This, in turn, affects the accurate extraction of spectral, texture, and structural features of individual trees in tree species classification. To address this issue, this paper proposes the WMF-SCS algorithm for ITC delineation. The algorithm consists of a two-stage process, as illustrated in Figure 3.

• Refining delineation boundaries and initial delineation: This step aims to reduce CHM image noise, preserve and enhance edge details in the image, and highlight tree crown boundaries without restricting over-segmentation. The fine-grained regions obtained provide richer information for subsequent feature extraction and regional optimization merging. First, Gaussian filtering and Contrast Limited Adaptive Histogram Equalization (CLAHE) [26] are applied to the CHM image for smoothing and local contrast enhancement. Next, the Otsu method [27] is used to perform global thresholding on the enhanced CHM image, generating a binary mask. Based on this, the Scharr operator [28] is employed to extract edge information from the image, resulting in a gradient magnitude map. Then, local maxima within the mask area are calculated as markers, and the watershed algorithm is applied to the gradient magnitude map to obtain the initial tree crown contours. Finally, the boundaries of the label matrix are extended, and the spatial adjacency matrix is constructed using the eight-neighbor relationship of each pixel. This matrix represents the connectivity between each pair of adjacent regions and is used to calculate the similarity between neighboring regions in subsequent steps.
• Merging adjacent and similar regions: This step aims to reduce over-segmentation and optimize the ITC delineation results from the first stage. First, for the R, G, and B bands of the RGB image and the CHM image, the mean, variance, minimum, maximum, median, range, quartiles, and interquartile range are calculated for each region, resulting in a total of 24 statistical measures as RGB color space features and 8 CHM statistical features. The RGB image is then converted to the HSV color space, and the mean and standard deviation of hue, saturation, and value are calculated for each region, resulting in 6 statistical measures as HSV color space features. Based on the Gray Level Co-occurrence Matrix (GLCM) [29], Local Binary Patterns (LBP) [30], and Gabor filtering [31] methods, GLCM features, LBP features, and GF mean and variance in the 0, $\pi /4$, and $\pi /2$ directions are calculated from the grayscale image, yielding 13 statistical measures as texture features, as detailed in

  Table 3. The textural metrics with variable description.

  Methods	Textural Metrics	Variable Description
  GLCM	Contrast	$CON={\sum }_{i,j=0}^{N-1}{P}_{i,j}{(i-j)}^2$
  	Correlation	$COR={\sum }_{i,j=0}^{N-1}{P}_{i,j}\left[{\displaystyle \frac{\left(i-{\mu }_i\right)\left(j-{\mu }_j\right)}{\sqrt{\left({\sigma }_i^2\right)\left({\sigma }_j^2\right)}}}\right]$
  	Homogeneity	$HOM={\sum }_{i,j=0}^{N-1}{\displaystyle \frac{P_{i,j}}{1+{(i-j)}^2}}$
  	Dissimilarity	$DIS={\sum }_{i,j=0}^{N-1}{P}_{i,j}|i-j|$
  	Energy	$ENE=\sqrt{{\sum }_{i,j=0}^{N-1}{P}_{i,j}^2}$
  Local Binary Pattern	Angular Second Moment	$ASM={\sum }_{i,j=0}^{N-1}{P}_{i,j}^2$
  	LBP Features	$LBP={\sum }_{i=0}^{N-1}s(i-c)\times 2^i$
  Gabor Filters	Gabor Mean	$GM={\displaystyle \frac{1}{N}}{\sum }_{i=0}^N{I}_i$
  	Gabor Variance	$GV={\displaystyle \frac{1}{N}}{\sum }_{i=0}^N{(I_i-{\mu }_i)}^2$

  Note: $N$, $c$, ${\mu }_i$, and ${\sigma }_i$ represent the total number of pixels, the gray value of the central pixel in the co-occurrence matrix, the mean gray value for pixel, and the standard deviation of pixel values, respectively. $P_{i,j}$ refers to the normalized co-occurrence matrix, which represents the probability of occurrence of pairs of pixel values ($i$, $j$) at specific spatial relationships. $s(i-c)$ is a sign function that compares the gray value of a neighboring pixel i with the central pixel $c$ (1 if $i$ ≥ $c$, 0 if $i$ < $c$). $I_i$ represents the pixel value at the $i$-th location in the image. The index $i$ ranges from 0 to N − 1, covering all pixel positions.

  .

Next, for the four feature groups calculated, Principal Component Analysis (PCA) is used to perform dimensionality reduction on the extracted features, retaining 80% of the principal components for each feature group. Based on the variance proportion explained by each component, a weighted average is calculated to generate composite values for each feature group, as shown in Equations (1) and (2). Combining these with the spatial adjacency matrix, the comprehensive similarity between each pair of adjacent regions in the RGB color space, HSV color space, texture, and CHM statistical features is calculated, as shown in Equation (3), resulting in a feature similarity matrix. Finally, the precomputed similarity matrix is used to initialize the spectral clustering algorithm, which identifies similar tree regions and merges adjacent and similar regions to obtain the final ITC delineation results.

$$R_{\left(i\right)}={\sum }_{q=1}^n{\rho }_i\times {PC}_q^i$$

(1)

$${\sum }_{i=1}^n{\rho }_i=1$$

(2)

$${\sigma }_{i,j}=w_1\times e^{{\displaystyle \frac{{-\left|\right|R_{\left(i\right)}-R_{\left(j\right)}\left|\right|}^2}{\alpha }}}+w_2\times e^{{\displaystyle \frac{{-\left|\right|H_{\left(i\right)}-H_{\left(j\right)}\left|\right|}^2}{\beta }}}+w_2\times e^{{\displaystyle \frac{{-\left|\right|T_{\left(i\right)}-T_{\left(j\right)}\left|\right|}^2}{\gamma }}}+w_4\times e^{{\displaystyle \frac{{-\left|\right|S_{\left(i\right)}-S_{\left(j\right)}\left|\right|}^2}{\delta }}}$$

(3)

where ${\sigma }_{i,j}$ represents the similarity between region $i$ and region $j$. $R_{\left(i\right)}$, $H_{\left(i\right)}$, $T_{\left(i\right)}$, $S_{\left(i\right)}$ and $R_{\left(j\right)}$, $H_{\left(j\right)}$, $T_{\left(j\right)}$, $S_{\left(j\right)}$ are the composite feature values of the RGB color space, HSV color space, texture, and CHM statistical features for region $i$ and region $j$, respectively. The scaling factors $\alpha$, $\beta$, $\gamma$, and $\delta$ are set to 0.05 times the maximum difference value for each feature based on the methodology in [3] to ensure effective normalization while preserving the feature differences in the dataset. The weight factors $w_1$, $w_2$, $w_3$, and $w_4$ represent the weights of each feature and are set to 0.3, 0.3, 0.2, and 0.2, respectively. The weight values were determined through multiple experiments and validations, based on initial tests and the specific characteristics of the dataset, to achieve the most balanced performance in similarity computation. ${\rho }_i$ is the variance proportion explained by the ith principal component retained in the PCA analysis (retaining 80% of the principal components), and ${PC}_q^i$ is the score of the $q$-th sample on the $i$-th principal component.

The performance and accuracy of the algorithm are comprehensively evaluated using three metrics: recall, precision, and F1-score, as shown in Equations (4)–(6). True Positive (TP) represents correctly segmented tree crowns, where the segmented region overlaps more than 50% with the ground reference area of an actual tree. False Positive (FP) indicates a segmented region that either does not correspond to an actual tree in the ground reference or overlaps less than 50% with the ground reference tree, constituting over-segmentation. False Negative (FN) refers to unsegmented tree crowns, where an actual tree exists in the ground reference, but no segmented region overlaps more than 50% with it.

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(4)

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(5)

$$F1=2\times {\displaystyle \frac{Precision\times Recall}{Precision+Recall}}$$

(6)

2.2.2. Individual Tree Feature Extraction

Based on the above ITC delineation results and the registration of LiDAR data, 27 LiDAR features, including height, intensity statistics, and crown morphology parameters, were extracted from the individual tree LiDAR data to classify dominant tree species in the study area. The specific classification features are detailed in

Table 4. LiDAR feature variables with variable description.

LiDAR Feature Variables	Variable Description
Max	${HI}_{max}=\mathrm{m}\mathrm{a}\mathrm{x} \left(HI\right)$
Mean	${HI}_{mean}={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{HI}_i$
Standard Deviation	${HI}_{sd}=\sqrt{{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{({HI}_i-{HI}_{mean})}^2}$
Min	${HI}_{min}=\mathrm{m}\mathrm{i}\mathrm{n} \left(HI\right)$
Median	${HI}_{median}=\mathrm{m}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{n} \left(HI\right)$
25th Percentile	${HI}_{ph25}=\mathrm{p}\mathrm{e}\mathrm{r}\mathrm{c}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{l}\mathrm{e} (HI,25)$
75th Percentile	${HI}_{ph75}=\mathrm{p}\mathrm{e}\mathrm{r}\mathrm{c}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{i}\mathrm{l}\mathrm{e} (HI,75)$
Coefficient of Variation	${HI}_{cv}={\displaystyle \frac{{HI}_{sd}}{{HI}_{mean}}}$
Skewness	${HI}_{skew}={\displaystyle \frac{{\sum }_{i=1}^n{({HI}_i-{HI}_{mean})}^3}{n\times {{HI}_{sd}}^3}}$
Kurtosis	${HI}_{kurt}={\displaystyle \frac{{\sum }_{i=1}^m{({HI}_i-{HI}_{mean})}^4}{n\times {{HI}_{sd}}^4}}$
Crown Volume	$CV={\displaystyle \frac{1}{6}}{\sum }_{i=1}^k‖\begin{array}{ccc}{x}_2^i-x_1^i& y_2^i-y_1^i& z_2^i-z_1^i\\ x_3^i-x_1^i& y_3^i-y_1^i& z_3^i-z_1^i\\ x_4^i-x_1^i& y_4^i-y_1^i& z_4^i-z_1^i\end{array}‖$
Point Density	$\mathrm{P}\mathrm{D}={\displaystyle \frac{N}{CV}}$
Crown Projection Area	$CPA={\displaystyle \frac{1}{2}}\times \left|{\sum }_{i=1}^n\left(x_i{y}_{i+1}-y_i{x}_{i+1}\right)\right|$
Crown Diameter	$CD={\displaystyle \frac{\max \left(x_i\right)-\min \left(x_i\right)+\max \left(y_i\right)-\min \left(y_i\right)}{2}}$
Crown Cover Index	$CCI={\displaystyle \frac{CPA}{(H_{max}-H_{min})\times \mathrm{C}\mathrm{D}}}$
Crown Shape Index	$CSI={\displaystyle \frac{(H_{max}-H_{min})}{\mathrm{C}\mathrm{D}}}$
Crown Volume Ratio	$CVR={\displaystyle \frac{CPA\times (H_{max}-H_{min})}{CV}}$

Note: ${HI}_i$, $N$, $x_i$, and $y_i$ represent the height or intensity of the $i$th LiDAR point, the total number of points in the point cloud, and the coordinate of the point, respectively. The index $i$ ranges from 1 to N, corresponding to the data points in the LiDAR point cloud.

.

The echo height, intensity statistics, and Point Density (PD) were calculated for point cloud data within the individual tree regions, resulting in 21 point cloud statistical features. Additionally, to capture morphological differences between tree species and obtain richer geometric information, the maximum distances in the east–west and north–south directions were used to determine the Crown Diameter (CD). The Quickhull [32] method was employed to calculate the Crown Volume (CV), and the crown projection area (CPA) was estimated using the volume of the two-dimensional convex hull. Based on these, three crown morphology parameters were calculated: Crown Cover Index (CCI), Crown Shape Index (CSI), and Crown Volume Ratio (CVR) [33].

Based on the above ITC delineation results and the registration of hyperspectral data, the Cross-Validated Support Vector Machine Recursive Feature Elimination (CV-SVM-RFE) algorithm [34] was applied to the individual tree hyperspectral data to select the optimal combination of original bands. Hyperspectral classification features were constructed by combining statistical transformation variables with texture features and vegetation indices. A total of 53 hyperspectral features were extracted for the classification of dominant tree species in the study area, with specific classification features detailed in Table 3 and

Table 5. The spectral feature variables with variable description.

Spectral Feature Variables	Variable Description	References
B1, B2, B6, B7, B8, B11, B17, B19,	CV-SVM-RFE
B20, B21, B22, B24, B25, B31, B35, B45
PCA1, PCA2, PCA3	PCA
MNF1, MNF2, MNF3	MNF
Carotenoid Index (CRI)	$CRI={\displaystyle \frac{1}{R_{510\mathrm{nm}}}}-{\displaystyle \frac{1}{R_{550\mathrm{nm}}}}$	[35]
Modified Soil-Adjusted Vegetation Index (MSAVI)	$MSAVI={\displaystyle \frac{2\times B_{nir}+1-\sqrt{{\left(2\times B_{nir}+1\right)}^2-8\times \left(B_{nir}-B_{red}\right)}}{2}}$	[36]
Plant Senescence Reflectance Index (PSRI)	$PSRI={\displaystyle \frac{R_{680\mathrm{nm}}-R_{500\mathrm{nm}}}{R_{750\mathrm{nm}}}}$	[37]
Red Edge Position (REP)	$REP=700+40\times \left({\displaystyle \frac{R_{740\mathrm{nm}}-R_{700\mathrm{nm}}}{R_{760\mathrm{nm}}-R_{700\mathrm{nm}}}}\right)$	[38]
Difference Vegetation Index (DVI)	$DVI=B_{nir}-B_{red}$	[39]
Soil-Adjusted Vegetation Index (SAVI)	$SAVI={\displaystyle \frac{\left(B_{nir}-B_{red}\right)\times \left(1+L\right)}{B_{nir}+B_{red}+L}}$	[40]
Green Normalized Difference Vegetation Index (GNDVI)	$GNDVI={\displaystyle \frac{B_{nir}-B_{green}}{B_{nir}+B_{green}}}$	[41]
Chlorophyll Absorption Reflectance Index (CARI)	$CARI={\displaystyle \frac{R_{700\mathrm{nm}}-\left(R_{670\mathrm{nm}}+\left(R_{700\mathrm{nm}}-R_{670\mathrm{nm}}\right)\times \left(\frac{700-670}{740-670}\right)\right)}{R_{550\mathrm{nm}}}}$	[42]
Chlorophyll Index (CI)	$CI={\displaystyle \frac{B_{nir}}{B_{green}}}-1$	[43]
Chlorophyll Content Index (CCI)	$CCI={\displaystyle \frac{B_{nir}}{B_{red}}}-1$	[44]
Normalized Difference Water Index (NDWI)	$NDWI={\displaystyle \frac{B_{green}-B_{nir}}{B_{green}+B_{nir}}}$	[45]
Ratio Vegetation Index I (RVI I)	$RVI I={\displaystyle \frac{B_{nir}}{B_{red}}}$	[46]
Ratio Vegetation Index II (RVI II)	$RVI II={\displaystyle \frac{B_{green}}{B_{red}}}$	[47]
Red Edge Normalized Difference Vegetation Index (RENDVI)	$RENDVI={\displaystyle \frac{B_{nir}-B_{rededge}}{B_{nir}+B_{rededge}}}$	[48]
MERIS Terrestrial Chlorophyll Index (MTCI)	$MTCI={\displaystyle \frac{B_{nir}-B_{rededge}}{B_{rededge}-B_{red}}}$	[49]
Normalized Difference Vegetation Index (NDVI)	$NDVI={\displaystyle \frac{B_{nir}-B_{red}}{B_{nir}+B_{red}}}$	[39]
Modified Simple Ratio (MSR)	$MSR={\displaystyle \frac{\left(\frac{B_{nir}}{B_{red}}\right)-1}{\sqrt{\left(\frac{B_{nir}}{B_{red}}\right)+1}}}$	[50]
Logarithmic Ratio Spectral Index (RSI)	$RSI=\ln \left({\displaystyle \frac{B_{nir}}{B_{red}}}\right)$	[51]
Modified Triangular Vegetation Index 2 (MTVI2)	$MTVI2=1.5\times \left({\displaystyle \frac{1.2\times \left(B_{nir}-B_{green}\right)-2.5\times \left(B_{red}-B_{green}\right)}{\sqrt{{\left(2\times B_{nir}+1\right)}^2-\left(6\times B_{nir}-5\times \sqrt{B_{red}}\right)-0.5}}}\right)$	[52]

Note: $B_{green}$, $B_{red}$, $B_{rededge}$, $B_{nir}$, $R_{500\mathrm{nm}}$, $R_{510\mathrm{nm}}$, $R_{550\mathrm{nm}}$, $R_{670\mathrm{nm}}$, $R_{680\mathrm{nm}}$, $R_{700\mathrm{nm}}$, $R_{740\mathrm{nm}}$, $R_{750\mathrm{nm}}$, $R_{760\mathrm{nm}}$ represent the mean reflectance of all bands in the green, red, red-edge, and near-infrared spectral ranges, as well as the spectral reflectance at 500, 510, 550, 670, 680, 700, 740, 750, and 760 nm, respectively. L is the soil brightness correction factor, set to 0.5 by default, as recommended in [40].

.

• Selection of 16 bands: When selecting original band features, the CV-SVM-RFE algorithm was used to filter 16 bands from the original 45 bands as classification variables to reduce the computational complexity of hyperspectral data processing, remove redundant information, and minimize noise impact.
• Statistically transformed variables: In addition to the preliminary selection of original bands, the Minimum Noise Fraction (MNF) transformation and PCA methods were applied to the optimal subset of original bands. Both methods effectively reduce data redundancy and improve the efficiency of data processing and analysis. Specifically, the top three components with a variance contribution rate above 90% in PCA and a signal-to-noise ratio greater than 3 in MNF were selected as classification variables.
• Textural metrics: GLCM calculations were performed on the first principal component obtained from MNF and PCA transformations, with pixel distances set to 1, 2, 3, and 4, and pixel angles set to 0, $\pi /4$, $\pi /2$, and $3\pi /4$ radians. The average values of the texture features were computed from the combinations of different pixel distances and angles, resulting in the extraction of 12 texture features.
• Vegetation indices: Based on the spectral range of the hyperspectral data and previous research experience, 19 relevant vegetation indices were constructed for classification.

2.2.3. Classification Feature Optimization

The obtained features were filtered by eliminating redundant and irrelevant features, while retaining those with the highest discriminative power. To mitigate the influence of numerical discrepancies in data with varying attributes on feature selection and classification models, the data were standardized prior to feature optimization.

The importance of variables was ranked and selected using Out-Of-Bag (OOB) data from the random forest (RF) model [53], with Permutation Importance (PI) [54] used as the criterion for assessing the importance of feature variables, thereby removing redundant feature variables. Specifically, during the construction of the RF model, the model’s generalization ability was evaluated through OOB error estimation, and the PI for each feature was calculated using OOB samples to identify the most discriminative features.

2.2.4. Classification Image

Four classification images were designed for hyperspectral data, three classification images for LiDAR data, and twelve classification images were developed based on all data sources and classifiers, along with two classification images designed for the ITC delineation algorithm.

Images 1 and 2 were used to evaluate the applicability of the CV-SVM-RFE band selection algorithm. Images 5 and 6 explored the impact of crown morphology parameters on classification results. Images 3, 4, and 6, 7 validated the effectiveness of the feature selection method. Images 8–19 compared the classification performance of single-source and multi-source data and the advantages and disadvantages of different machine learning classifiers in individual tree species classification. Images 20 and 21 investigated the impact of delineation accuracy on tree species classification accuracy. The specific classification images are detailed in

Table 6. Classification image using hyperspectral data and the RF classifier.

Image Number	Variable Group
1	Full hyperspectral bands
2	Selection of 16 bands
3	Full hyperspectral features
4	Selected hyperspectral features

,

Table 7. Classification image using LiDAR data and the RF classifier.

Image Number	Variable Group
5	All features except crown morphological parameters
6	Full LiDAR features
7	Selected LiDAR features

,

Table 8. Classification image using LiDAR, hyperspectral data, and the RF, SVM, MLP, and SAMME classifiers.

Image Number	Classifier	Variable Group
8	SVM	Hyperspectral features
9	RF	Hyperspectral features
10	MLP	Hyperspectral features
11	SAMME	Hyperspectral features
12	SVM	LiDAR features
13	RF	LiDAR features
14	MLP	LiDAR features
15	SAMME	LiDAR features
16	SVM	Hyperspectral features + LiDAR features
17	RF	Hyperspectral features + LiDAR features
18	MLP	Hyperspectral features + LiDAR features
19	SAMME	Hyperspectral features + LiDAR features

and

Table 9. Classification image using different first-stage segmentation algorithms of WMF-SCS.

Image Number	First-Stage Segmentation Algorithm of WMF-SCS	Variable Group
20	CHM-based watershed segmentation algorithm	Hyperspectral features + LiDAR features
21	WMF-SCS algorithm	Hyperspectral features + LiDAR features

.

2.2.5. Classification Model

This study employed the Stratified K-Folds Cross-Validator (Stratified K-Fold CV) method to enhance the reliability and stability of model evaluation, improve the credibility of classification results, and avoid overfitting or underfitting due to class imbalance [55]. The dataset was divided into five mutually exclusive subsets, with each subset maintaining the same class proportion as the overall dataset. In each iteration, four subsets were used for model training, and the remaining subset was used for model testing, with the process repeated five times to cover all subsets. The classifier settings are as follows:

• RF Classifier: RF classifies by constructing multiple decision trees and combining their results, with strong noise resistance and the ability to handle high-dimensional data. Numerous studies have demonstrated that RF classifiers can effectively classify tree species. In this study, the number of decision trees and the maximum depth were set to 1000 and 20, respectively, to ensure that the model could fully learn the data features.
• SVM Classifier: SVM is an efficient supervised learning algorithm that finds the optimal separating hyperplane in high-dimensional space to maximize the margin between classes, making it suitable for classification tasks in high-dimensional feature spaces. A linear kernel function was chosen in this study to handle linearly separable problems in high-dimensional feature spaces. The linear kernel function helps simplify computation, making it suitable for handling large-scale datasets while maintaining efficient classification performance.
• Multilayer Perceptron (MLP) Classifier: MLP is a feedforward neural network that includes an input layer, multiple hidden layers, and an output layer. It can learn complex nonlinear data relationships through nonlinear activation functions and the backpropagation algorithm. For the complex features present in tree species classification tasks, MLP effectively captures the relationships between these high-dimensional data. In this study, one hidden layer with 100 neurons was set, with a maximum of 1000 iterations, using the Rectified Linear Unit (ReLU) as the activation function for the hidden layer to handle nonlinear feature relationships.
• Stagewise Additive Modeling using a Multiclass Exponential Loss Function (SAMME) Classifier: SAMME, based on boosting, combines multiple weak classifiers to form a robust classifier, improving performance with each iteration. Unlike Adaptive Boosting, SAMME uses a multiclass exponential loss function, making it suitable for multiclass tasks. Its adaptive weight adjustment and ensemble learning mechanism excel in handling complex multiclass tasks. In this study, decision trees with a maximum depth of 1 were chosen as base learners, with 1500 weak classifiers set to ensure adequate model learning.

The classification accuracy in this study was evaluated using the following four metrics: Producer’s Accuracy (PA) and User’s Accuracy (UA) for each class, Overall Accuracy (OA) of the classifier, and the Kappa Coefficient. These metrics provide a comprehensive evaluation of classification accuracy from multiple perspectives, with specific calculation methods shown in Equations (7)–(10).

$$OA={\displaystyle \frac{{\sum }_{i=1}^n{N}_{ii}}{N}}$$

(7)

$${PA}_i={\displaystyle \frac{N_{ii}}{{\sum }_{j=1}^n{N}_{ij}}}$$

(8)

$${UA}_i={\displaystyle \frac{N_{ii}}{{\sum }_{j=1}^n{N}_{ji}}}$$

(9)

$$Kappa={\displaystyle \frac{N\times {\sum }_{i=1}^n{N}_{ii}-{\sum }_{i=1}^n({\sum }_{j=1}^n{N}_{ij}\times {\sum }_{j=1}^n{N}_{ji})}{N^2-{\sum }_{i=1}^n({\sum }_{j=1}^n{N}_{ij}\times {\sum }_{j=1}^n{N}_{ji})}}$$

(10)

where $N$ is the total number of samples, $n$ is the number of classes, $N_{ii}$ is the number of correctly classified samples in class $i$, ${\sum }_{j=1}^n{N}_{ij}$ is the total number of actual samples in class $i$, and ${\sum }_{j=1}^n{N}_{ji}$ is the total number of samples classified as class $i$.

All methods were executed on the same system to ensure consistency and fairness in the results. The experiments were conducted on a Windows 10 machine, running Python 3.9.18 and scikit-learn 1.4.0. The system is equipped with an AMD Ryzen 7 4800H processor, 16 GB of RAM, and an NVIDIA GeForce RTX 2060 Max-Q GPU.

3. Results

3.1. Individual Tree Crown Delineation Results

The ITC delineation results based on the CHM watershed algorithm and the WMF-SCS algorithm are given in

Table 10. The accuracy assessment results of individual tree crown delineation of WMF-SCS algorithm.

Algorithm	Site	Ground- Measured Tree	True Positive	False Positive	False Negative	Precision	Recall	F1-Score
CHM-based watershed segmentation algorithm	Site 1	127	53	279	74	0.16	0.42	0.23
	Site 2	98	57	256	41	0.18	0.58	0.28
	Site 3	95	49	328	46	0.13	0.52	0.21
	All sites	320	159	863	161	0.16	0.50	0.24
WMF-SCS algorithm	Site 1	127	113	16	14	0.88	0.89	0.88
	Site 2	98	83	13	15	0.86	0.85	0.86
	Site 3	95	82	10	13	0.89	0.86	0.88
	All sites	320	278	39	42	0.88	0.87	0.87

. Comparing the ITC delineation performance of the CHM-based watershed segmentation algorithm and the WMF-SCS ITC delineation algorithm, the WMF-SCS algorithm effectively reduces over-segmentation and improves ITC delineation precision (Precision = 0.88, Recall = 0.87, F1-Score = 0.87). In Site 2, characterized by highly variable and densely populated shelter forests, the WMF-SCS algorithm performed better than the CHM watershed algorithm in handling the complex canopy structure. However, due to the increased tree density and overlapping crowns in this region, the algorithm’s performance in Site 2 was slightly less accurate compared to Sites 1 and 3. The ITC delineation results for regions 1, 2, and 3 are shown in Figure 4, Figure 5, and Figure 6, respectively.

3.2. Tree Species Classification Results

In this study, 16 bands were selected for combination based on the CV-SVM-RFE band selection algorithm. The selected band numbers are 1, 2, 6, 7, 8, 11, 17, 19, 20, 21, 22, 24, 25, 31, 35, and 45. These selected bands cover the entire range, with uniform distribution and relatively low computational requirements. A comparison of the tree species spectral response curves using the selected band combination with the original 45-band spectral response curves is shown in Figure 7. As seen in Figure 7, the spectral response curves of the band combination selected by the CV-SVM-RFE band selection algorithm fit well with the original 45-band spectral response curves, effectively preserving the original spectral features.

The distribution of feature variable importance is shown in Figure 8. When both hyperspectral and LiDAR data are used, the 14 most important features are evenly split between LiDAR and hyperspectral features. Compared to hyperspectral data, LiDAR-derived features such as crown projection area and crown diameter hold higher importance, providing more useful features for tree species classification. The importance distribution indicates that the combined data significantly contribute to tree species classification, with classification performance benefiting from the complementarity of multi-source data. This demonstrates that the morphological and structural information provided by LiDAR data, along with the spectral and texture information provided by hyperspectral data, both play critical roles in the classification process.

The box plots of the top six most important features from LiDAR and hyperspectral data are shown in Figure 9, Figure 10, and Figure 11, respectively. Overall, Populus bolleana shows stable and relatively high performance across multiple features, particularly in the LiDAR-derived CPA and CSI variables, as well as the hyperspectral CARI variable. Elaeagnus angustifolia and Ulmus pumila exhibit similar behavior across most features, showing more consistent feature distributions, with some features standing out but with greater variability. Dead trees display instability across several features, particularly in the LiDAR-derived CPA and PD variables and the hyperspectral CARI variable, showing lower values, but they exhibit some advantages in the CSI and PCA2 variables. These results indicate significant differences in the distribution and variability of LiDAR feature variables among different tree species, providing strong discriminative evidence for tree species classification tasks.

The relationship between the number of LiDAR and hyperspectral data features and classification accuracy is shown in Figure 12, Figure 13 and Figure 14. Both OA and the Kappa coefficient show an initial increase followed by a stabilization trend as the number of feature variables increases. Overall, increasing the number of feature variables can effectively improve classification accuracy; however, when the number of feature variables reaches a certain point, their impact on classification accuracy tends to stabilize or even slightly decrease. This suggests that too many feature variables may lead to data redundancy and overfitting.

In this study, when using only hyperspectral data as the data source, the classification results obtained with the original optimal band combination and before and after feature optimization are given in

Table 11. RF classification results of different classification images for hyperspectral data.

Tree Species	Image 1 (n = 45)		Image 2 (n = 16)		Image 3 (n = 53)		Image 4 (n = 19)
	UA/%	PA/%	UA/%	PA/%	UA/%	PA/%	UA/%	PA/%
Populus bolleana	76.92	80	79.31	92	84.62	88	91.67	88
Elaeagnus angustifolia	63.33	76	61.54	61.54	69.23	69.23	78.57	84.62
Ulmus pumila	60	50	54.55	33.33	62.50	58.82	68.75	64.71
Dead trees	93.33	77.78	85	1	94.44	94.44	94.44	94.44
OA	72.09%		72.09%		77.91%		83.72%
Kappa	0.6217		0.6218		0.7015		0.7800

. Comparing Images 1 and 2, after band selection, the model’s OA and Kappa coefficient remained largely unchanged compared to the original full-band configuration. Comparing Images 3 and 4, after feature optimization, the model’s OA and Kappa coefficient increased by 5.81% and 0.0785, respectively.

When using only LiDAR data as the data source, the classification results after adding LiDAR-derived crown morphology parameters and before and after feature optimization are given in

Table 12. RF classification results of different classification images for LiDAR data.

Tree Species	Image 5 (n = 21)		Image 6 (n = 27)		Image 7 (n = 14)
	UA/%	PA/%	UA/%	PA/%	UA/%	PA/%
Populus bolleana	73.08	76	87.50	84	95.46	84
Elaeagnus angustifolia	71.43	80	72	69.23	75	80.77
Ulmus pumila	66.67	55.56	68.42	76.47	72.22	76.47
Dead trees	94.12	88.89	1	1	1	1
OA	75.58%		81.40%		84.88%
Kappa	0.6698		0.7498		0.7963

. Comparing Images 5 and 6, after adding crown morphology parameters, the model’s OA and Kappa coefficient increased by 5.82% and 0.08, respectively. When comparing Images 6 and 7, after feature optimization, the model’s OA and Kappa coefficient increased by 3.48% and 0.0465, respectively. The research results indicate that feature optimization can significantly improve the classification model’s accuracy and consistency, validating the feature selection method’s effectiveness for both hyperspectral and LiDAR data.

The tree species classification results using different data sources and classifiers are given in

Table 13. Results are classified by different data and classifiers.

Classifier	Data	PA/UA	Populus Bolleana	Elaeagnus Angustifolia	Ulmus Pumila	Dead Trees	OA	Kappa
SVM	Hyperspectral	PA	88%	76.92%	66.67%	94.12%	81.40%	0.7483
		UA	91.67%	68.97%	80%	88.89%
	LiDAR	PA	84%	73.08%	64.71%	94.44%	79.07%	0.7163
		UA	77.78%	70.37%	78.57%	94.44%
	Hyperspectral + LiDAR	PA	96%	80.77%	77.78%	94.12%	87.21%	0.8275
		UA	96%	80.77%	82.35%	88.89%
RF	Hyperspectral	PA	88%	84.62%	64.71%	94.44%	83.72%	0.7800
		UA	91.67%	78.57%	68.75%	94.44%
	LiDAR	PA	84%	80.77%	76.47%	100%	84.88%	0.7963
		UA	95.45%	75%	72.22%	100%
	Hyperspectral + LiDAR	PA	100%	76.92%	94.44%	94.12%	90.70%	0.8747
		UA	89.29%	90.91%	85%	100%
MLP	Hyperspectral	PA	84%	80.77%	55.56%	82.35%	76.74%	0.6831
		UA	75%	70%	83.33%	87.50%
	LiDAR	PA	80%	53.85%	70.59%	94.44%	73.26%	0.6425
		UA	80%	70%	60%	80.95%
	Hyperspectral + LiDAR	PA	100%	50%	77.78%	100%	80.23%	0.7356
		UA	83.33%	81.25%	66.67%	89.47%
SAMME	Hyperspectral	PA	80%	88.46%	38.89%	82.35%	74.42%	0.6510
		UA	100%	60.53%	63.64%	82.35%
	LiDAR	PA	80%	69.23%	64.71%	100%	77.91%	0.7014
		UA	83.33%	64.29%	68.75%	100%
	Hyperspectral + LiDAR	PA	96%	69.23%	88.24%	83.33%	83.72%	0.7807
		UA	82.76%	81.82%	75%	100%

. The classification performance varies among different tree species. Dead trees exhibit significant spectral differences in the near-infrared band compared to other species, and their crown morphology is distinctive, making the classifier most effective at identifying this species. Populus bolleana has a larger sample size, and its crown morphology is characterized by a cylindrical or conical shape. The LiDAR data effectively capture vertical structure and morphological differences, resulting in high accuracy across all classifiers. In contrast, Elaeagnus angustifolia and Ulmus pumila are more similar in spectral and morphological structure, increasing the classification difficulty and leading to relatively lower accuracy. The LiDAR point clouds of different tree species are shown in Figure 15.

The classification results demonstrate that the RF classifier, when integrated with multi-source data, achieves the highest classification accuracy and consistency, with an accuracy of 90.70% and a Kappa coefficient of 0.8747. In contrast, the classification performance with single-source data is inferior, with both OA and the Kappa coefficient failing to reach the levels achieved after multi-source data fusion.

Moreover, the RF classifier generally outperforms the SVM, MLP, and SAMME classifiers, demonstrating high applicability and advantages in the tree species classification task using multi-source data. The RF classifier performed best for classifying Populus bolleana and Dead trees, with OAs of 95.45% and 100%, respectively. The SVM, MLP, and SAMME classifiers also performed well, particularly when hyperspectral and LiDAR data were combined, significantly improving OA and the Kappa coefficient. These results indicate that multi-source data can provide more dimensions and richer feature information, thereby enhancing the discriminative power and stability of classifiers. The fusion of multi-source data and the selection of an appropriate classifier are key strategies for improving the accuracy of individual tree species classification.

Due to the different ITC delineation results between the WMF-SCS and CHM-based watershed segmentation algorithms, the number of generated samples is inconsistent, making direct classification comparison unfeasible. However, the second stage of the WMF-SCS algorithm only involves merging similar regions. By replacing the first stage of the WMF-SCS algorithm with the CHM-based watershed segmentation algorithm, the final number of generated samples can be made approximately consistent across both algorithms. The WMF-SCS algorithm was applied to the entire study area, and based on the ITC delineation results from both algorithms, combined with tree species information from ground survey data, 420 successfully matched samples were selected as classification samples. Tree species classification was performed using both algorithms, based on the Stratified K-Fold CV and RF classifier. The detailed accuracy assessment results are given in

Table 14. RF classification results using different first-stage segmentation algorithms of WMF-SCS.

First-Stage Segmentation Algorithm of WMF-SCS	Data	OA	Kappa
CHM-based watershed segmentation algorithm	Hyperspectral + LiDAR	87.47%	0.8413
WMF-SCS algorithm	Hyperspectral + LiDAR	89.32%	0.8631

. The research results indicate that, compared to the CHM-based watershed segmentation algorithm, the WMF-SCS algorithm improved the OA and Kappa coefficient of the classification results by 1.85% and 0.0218, respectively.

Using the optimal model based on multi-source data combined with the RF classifier, the WMF-SCS algorithm ITC delineation results of the three representative regions were used as test samples for tree species classification within the regions. The classification results are shown in Figure 16.

4. Discussion

Image registration is an essential process in tree species classification, as it facilitates the integration of multi-source remote sensing data for accurate analysis. Given the nature of our remote sensing data, bilinear interpolation was chosen for image upscaling to strike a balance between computational efficiency and the need for a smoother representation of pixel values. This method helps minimize pixelated effects, which is crucial when dealing with high-resolution data from UAV platforms. While we recognize the potential effects of smoothing on classification accuracy, bilinear interpolation was deemed more suitable than nearest neighbor interpolation for preserving fine-scale details in the image [56], ensuring better consistency in the derived features.

A comparative analysis of the results from the two ITC delineation algorithms shows that improving delineation accuracy can enhance tree species classification accuracy. Additionally, Maschler et al. [57] conducted ITC delineation and classification for 13 tree species using airborne hyperspectral data, while Qin et al. [3] used UAV-based LiDAR, hyperspectral, and ultra-high-resolution RGB data for ITC delineation and species classification in a subtropical broadleaf forest. Both studies emphasized the critical role of delineation accuracy in tree species classification, noting that improved ITC delineation increases overall classification accuracy. These findings underscore the importance of ITC delineation as a foundational step in tree species classification tasks using remote sensing data. Furthermore, high-precision ITC delineation and accurate tree species classification can enhance the reliability of monitoring forest composition and structure. These improvements can enable more informed decision making in forest management, particularly in species monitoring, resource tracking, and forest health assessment.

While the WMF-SCS algorithm significantly improves ITC delineation in complex mixed forest environments, it has some limitations. The initial delineation step may fail to accurately capture small or densely clustered tree crowns, particularly in areas with high tree density or complex canopy structures. Additionally, the second stage, which merges similar regions, relies heavily on the quality of the initial delineation, meaning errors in the first stage can impact the final results. Furthermore, the algorithm remains sensitive to canopy features like dense crowns, irregular shapes, and overlapping structures. Future improvements could include incorporating advanced techniques such as deep learning-based segmentation or multi-scale analysis to enhance accuracy and robustness.

A comparison of the classification results before and after the inclusion of crown morphology parameters extracted from LiDAR data shows that adding these parameters improved the OA and Kappa coefficient of the model by 5.82% and 0.08, respectively. Furthermore, Maurya et al. [58] and Gong et al. [59] both highlighted the irreplaceable role of crown morphology features in enhancing tree species classification accuracy by providing rich structural and spatial information. Although the height and intensity variables extracted from LiDAR data can reflect the vertical structural characteristics of individual tree point clouds, crown morphology parameters—such as crown projection area, diameter, and shape irregularities—provide a more comprehensive description of the overall crown shape. These features enable more accurate differentiation of tree species, particularly in areas with complex or overlapping tree structures.

Comparing the classification results of the four classifiers reveals that the RF classifier generally outperforms the SVM, MLP, and SAMME classifiers in most cases, showing superior performance in terms of classification accuracy and Kappa coefficient. This high performance can be attributed to the characteristics of the RF algorithm, which reduces the risk of overfitting by aggregating a large number of decision trees and effectively handles high-dimensional and complex data. In contrast, SVM relies on a small number of support vectors to determine the classification boundary, and when new features are added to the data samples, especially if these features are not as distinctive as the original ones, SVM’s classification performance may not significantly improve. Since the data sample size in this study is relatively small, neural networks like MLP require large amounts of data for effective training and to avoid overfitting, making MLP’s performance more dependent on data quantity. SAMME improves classification performance by integrating multiple weak classifiers, but the limited performance of the base classifiers makes it challenging to further improve overall classification effectiveness when dealing with complex data.

In multi-source data-based tree species classification, various indices are used to represent vegetation properties. While these indices provide valuable insights, their potential collinearity can affect the accuracy of classification models. In this study, we addressed potential multicollinearity by applying PCA and feature importance ranking, which helped reduce dimensionality and focus on the most relevant features. However, further exploration of the relationships between these indices using more advanced collinearity analysis methods [60] could provide deeper insights. Such analysis could refine index selection, reduce redundancy, and ultimately improve classification performance, offering a promising direction for future research.

Additionally, in the context of classification performance assessment, traditional metrics such as PA, UA, OA, and the Kappa coefficient are commonly used but have limitations. Phan et al. [61] emphasized that these metrics may not fully capture classification errors, particularly in complex datasets. Foody et al. [62] argued that the Kappa coefficient may be unsuitable for accurately assessing thematic map accuracy, especially in cases with class imbalances or small sample sizes. These studies highlight the need for more comprehensive evaluation methods. Future research could explore the integration of additional evaluation metrics, such as the F1 score or adjusted Rand index, to provide a more nuanced understanding of classification performance and address the limitations of traditional metrics in multi-source data tree species classification.

Although the WMF-SCS algorithm shows promising results for ITC delineation, this study did not perform a direct or indirect comparison of tree species classification accuracy between the WMF-SCS algorithm and the simple spectral clustering method. This is because the spectral clustering process involves merging similar regions, which leads to differences in the number of samples generated by the two algorithms. As a result, comparing the classification accuracy between the two methods, either directly or indirectly, would not yield meaningful conclusions. Furthermore, the main tree species in the shelterbelt are Ulmus pumila, Populus bolleana, Populus euphratica, Elaeagnus angustifolia, Haloxylon ammodendron, and Fraxinus, with understory vegetation primarily composed of mixed shrubs. Using more diverse tree species data to compare and validate the ITC delineation algorithm and tree species classification method presented in this study would be ideal. However, due to the imbalance in tree species samples within the shelterbelt and the limited scope and field survey data in this study, certain less-dominant tree species could not be included in the classification task due to insufficient sample size.

5. Conclusions

The main conclusions are as follows:

• Comparing the ITC delineation performance of the CHM-based watershed segmentation algorithm and the WMF-SCS ITC delineation algorithm, the WMF-SCS algorithm effectively reduces over-segmentation and improves ITC delineation precision (Precision = 0.88, Recall = 0.87, F1-Score = 0.87), making it suitable for ITC delineation tasks in complex mixed forest scenarios. Additionally, the OA and Kappa coefficient of the classification results increased by 1.85% and 0.0218, respectively. This indicates that high-precision ITC delineation can provide clearer and more reliable classification objects, features, and training data, thereby improving the accuracy and reliability of tree species classification.
• Comparing the classification results of hyperspectral and LiDAR data, the CV-SVM-RFE band selection algorithm effectively retains the original spectral information and reduces the dimensionality of hyperspectral data, with OA and Kappa coefficient remaining largely unchanged compared to the original full-band data. Moreover, crown morphology parameters have significant application value in tree species classification tasks; after adding these parameters, the model’s OA and Kappa coefficient increased by 5.82% and 0.08, respectively. After feature optimization, the OA of the model improved by 5.81% for hyperspectral data and 3.48% for LiDAR data.
• Multi-source data can effectively complement the features between different tree species, enhancing the model’s ability to distinguish between species, thus improving the accuracy of tree species classification. Compared to the classification results using single-source LiDAR data, the accuracy of multi-source data classification improved by an average of 7.94%. Compared to the classification results using single-source hyperspectral data, the accuracy of multi-source data classification improved by an average of 7.52%.
• Comparing the classification results of different classifiers, the RF classifier combined with multi-source data demonstrated the highest classification accuracy and consistency (OA = 90.70%, Kappa = 0.8747). Given the sample size in this study, the RF classifier generally outperformed the SVM, MLP, and SAMME classifiers, showing high applicability and advantages in tree species classification tasks involving multi-source data. These results suggest that integrating multi-source data and selecting the appropriate classifier are key strategies for improving the accuracy of individual tree species classification.