Individual tree-based forest species diversity estimation by classification and clustering methods using UAV data

Li, Xiuwen; Zheng, Zhaoju; Xu, Cong; Zhao, Ping; Chen, Junhua; Wu, Jinchen; Zhao, Xueming; Mu, Xuan; Zhao, Dan; Zeng, Yuan

doi:10.3389/fevo.2023.1139458

ORIGINAL RESEARCH article

Front. Ecol. Evol., 27 March 2023

Sec. Environmental Informatics and Remote Sensing

Volume 11 - 2023 | https://doi.org/10.3389/fevo.2023.1139458

This article is part of the Research Topic Remote Sensing Advances in Biodiversity and Ecosystem Functioning Research View all 12 articles

Individual tree-based forest species diversity estimation by classification and clustering methods using UAV data

Xiuwen Li^1,2

Zhaoju Zheng¹

Cong Xu^1,2

Ping Zhao^1,2

Junhua Chen^1,2

Jinchen Wu^1,2

Xueming Zhao^1,2

Xuan Mu^1,2

Dan Zhao^1,2 $Yuan Zeng,\n$ Yuan Zeng^1,2^*

¹State Key Laboratory of Remote Sensing Science, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China
²University of Chinese Academy of Sciences, Beijing, China

Monitoring forest species diversity is essential for biodiversity conservation and ecological management. Currently, unmanned aerial vehicle (UAV) remote sensing technology has been increasingly used in biodiversity monitoring due to its flexibility and low cost. In this study, we compared two methods for estimating forest species diversity indices, namely the spectral angle mapper (SAM) classification approach based on the established species-spectral library, and the self-adaptive Fuzzy C-Means (FCM) clustering algorithm by selected biochemical and structural features. We conducted this study in two complex subtropical forest areas, Mazongling (MZL) and Gonggashan (GGS) National Nature Forest Reserves using UAV-borne hyperspectral and LiDAR data. The results showed that the classification method performed better with higher values of R² than the clustering algorithm for predicting both species richness (0.62 > 0.46 for MZL and 0.55 > 0.46 for GGS) and Shannon-Wiener index (0.64 > 0.58 for MZL, 0.52 > 0.47 for GGS). However, the Simpson index estimated by the classification method correlated less with the field measurements than the clustering algorithm (R² = 0.44 and 0.83 for MZL and R² = 0.44 and 0.62 for GGS). Our study demonstrated that the classification method could provide more accurate monitoring of forest diversity indices but requires spectral information of all dominant tree species at individual canopy scale. By comparison, the clustering method might introduce uncertainties due to the amounts of biochemical and structural inputs derived from the hyperspectral and LiDAR data, but it could acquire forest diversity patterns rapidly without distinguishing the specific tree species. Our findings underlined the advantages of UAV remote sensing for monitoring the species diversity in complex forest ecosystems and discussed the applicability of classification and clustering methods for estimating different individual tree-based species diversity indices.

1. Introduction

Forest biodiversity is essential in maintaining ecosystem patterns, functions and services (Balvanera et al., 2006; Thompson et al., 2009; Brockerhoff et al., 2013). Forest species diversity is a fundamental component of biodiversity, which refers to the uniformity of the number and distribution of tree species in forest ecosystems (Magurran, 1988; Vellend, 2004). However, with the increasing pressure of human activities and climate change, it has faced severe threats, such as accelerated species extinction and increased endangered species (Iida and Nakashizuka, 1995; Haas et al., 2011). Therefore, accurate and repeated forest species diversity monitoring is important for biodiversity conservation and ecological management.

Currently, the assessment of species diversity in a certain region is mainly based on species diversity indices, among which species richness (Gaston, 2000) emphasizes the number of various species, while Shannon-Wiener index (Shannon, 1948) and Simpson index (Simpson, 1949) take into account both the amount and evenness of species. Traditional forest species diversity monitoring relies on field surveys to investigate these diversity indices (Kerr and Ostrovsky, 2003), which are labor- and material-intensive and focus on forest species distribution at the sample scale (Myers et al., 2000; Duro et al., 2007). Remote sensing has the advantages of an extensive detection range and a short data acquisition period, extending the possibilities of forest species diversity monitoring at both temporal and spatial scales (Turner et al., 2003; Skidmore et al., 2015). Near-surface remote sensing platform equipped with hyperspectral sensors and laser scanners has been a promising tool for forest species diversity monitoring in the past decade (Turner, 2014; Guo et al., 2017; Wang and Gamon, 2019; Pu, 2021). In particular, UAV (unmanned aerial vehicle) remote sensing technology, due to its advantages of flexibility and low cost, has shown great potential in species identification and biodiversity monitoring (Anderson and Gaston, 2013; Lin et al., 2019; De Almeida et al., 2021).

Hyperspectral data can obtain continuous spectral information of vegetation and has been increasingly used for monitoring forest species diversity (Féret and Asner, 2014; Ferreira et al., 2016; Laurin et al., 2016). The monitoring methods are mainly divided into two categories: supervised classification methods that directly identify forest species based on their spectroscopic characteristics (Féret and Asner, 2013; Fassnacht et al., 2016; Cao et al., 2018; Franklin and Ahmed, 2018), and spectral diversity metrics that indirectly link the variation of leaf or canopy spectra to species diversity (Palmer et al., 2002; Gholizadeh et al., 2018). Light Detection and Ranging (LiDAR) data can directly penetrate the vegetation canopy by actively emitting high-frequency pulses, so it is widely used for high-precision estimation of forest structural features, including tree height and crown diameter (Popescu, 2007; Morsdorf et al., 2009; Sankey et al., 2013; Wallace et al., 2014). Furthermore, advances in lidar remote sensing have enabled the accurate extraction of information from individual tree crowns (ITCs) (Ene et al., 2012; Zhao et al., 2014). Compared to the pixel-based approach, the ITC-based approach is more directly analogous to the field-based individual sampling method, which can better extract structural features of the canopy and minimize the signal confusion brought by non-tree pixels (Zheng et al., 2022). Based on LiDAR data, forest species diversity at the regional scale can be monitored by establishing relationships between structural features and field-measured species diversity indices (Lopatin et al., 2016; Torresani et al., 2020; Hu et al., 2021). However, the capability of forest species diversity monitoring using only hyperspectral data or LiDAR data might be limited by species complexity, image spectral mixing and canopy morphological variation (Koch, 2010; Guo et al., 2017). The fusion of hyperspectral and LiDAR data provides a synergistic ability, which can use vertical and horizontal information from each data source to acquire more robust diversity monitoring results.

Previous studies have integrated structural features extracted by LiDAR data and spectral characteristics from hyperspectral images for directly discriminating tree species by using classification techniques, including linear discrimination analysis (Alonzo et al., 2014), support vector machine (Dalponte et al., 2012), random forest (Liu et al., 2017) and spectral angle mapper (Zhao et al., 2020). Mayra et al. (2021) compared the performance of different classification methods for identifying the major tree species in a boreal forest based on airborne hyperspectral and LiDAR data. Assessing forest species diversity using remote sensing classification methods has the advantage of providing spatially explicit species distribution information for each ITC or pixel. However, it remains challenging to directly discriminate the species of all individuals accurately in complex subtropical or tropical forests due to the potential spectral or structural similarity among different species or differences existing for the same species (Price, 1994; Wang and Gamon, 2019). The confusion in classification usually increases with increasing biodiversity levels and more training data for species-rich forests is usually needed to improve the classification accuracy. Moreover, collecting sufficient training and validation data for each tree species in species-rich and topographically complex forests can be a challenging task. Although some methods are relatively capable of classifying trees with limited training samples (Christian et al., 2013; Awad, 2018), the classification results are achieved using specific images and algorithms with relatively lower transferability.

Many indirect approaches using spectral and structural information have shown great potential for monitoring forest species diversity, such as regression analysis and clustering. Regression analysis is to model the spectral and structural information directly with the measured species diversity indices, which is a mature and straightforward algorithm, but the applicability in different regions is poor (Ceballos et al., 2015). The clustering algorithm can evaluate species diversity by grouping trees with similar characteristics based on the biochemical and structural variation of different tree species (Asner et al., 2015; Padilla-Martinez et al., 2020; Pakgohar et al., 2021). Clustering can be used to identify patterns or trends in the distribution and abundance of different species within a forest ecosystem. Among them, the self-adaptive Fuzzy C-Means (FCM) clustering algorithm overcomes the disadvantage of traditional clustering methods, which require a pre-indication of the initial classes, and can determine the optimal number of clusters automatically (Bezdek et al., 1987; Li and Yu, 2009). Zhao et al. (2018) estimated the forest species richness and Shannon-Wiener index in a subtropical forest based on airborne LiDAR and hyperspectral data using seven biochemical components and tree height by the adaptive FCM clustering algorithm. The biochemical and structural parameters selected in each study area for clustering methods may be dependent, and their applicability to species diversity monitoring in other areas still needs to be determined, especially for forests with diverse species and complex compositions.

Several studies have used either classification or clustering methods to estimate forest species diversity by combining various features from remote sensing data, but it remains unclear which method is more effective for monitoring different aspects of diversity in different forest conditions. Although some studies have used spaceborne or airborne data to assess species diversity (Shen and Cao, 2017; Wan et al., 2021), they are often limited by spatial resolution or expensive costs. UAV-borne hyperspectral and LiDAR data could provide spatially explicit information on individual trees, so it is more advantageous to explore the applicability of advanced methods in different species-rich forests by UAV data.

Therefore, the major objectives of our study are to explore the performance of individual tree-based classification and clustering methods in estimating three commonly used forest species diversity indices (species richness, Shannon-Wiener index and Simpson index) in two typical subtropical forests in China using UAV-borne hyperspectral and LiDAR data. We aim to: (1) classify tree species using the SAM classification method based on hyperspectral image and the individual tree crown segmentation results from LiDAR data, (2) estimate forest species diversity using the self-adaptive FCM clustering algorithm based on optimal biochemical vegetation indices and structural features, and (3) further compare the performance of classification and clustering methods in these two subtropical forest sites.

2. Materials and methods

2.1. Study area

Subtropical forest in China is a hotspot of tree species richness and a priority area for forest species diversity monitoring (Li et al., 2009; Liu et al., 2018). We conducted research in two study areas, both of which are typical subtropical forests of China, but their forest species compositions and environmental conditions are different. The first study area is located in the Mazongling National Nature Reserve (MZL, 115°41′37′–115°42′5′E, 31°15′25′–31°15′44′N) in Jinzhai county, Anhui province of China (Figure 1). The study area covers about 23.8 ha with an elevation varying from 1,000 m to 1,184 m above sea level. This region is characterized by a subtropical monsoon climate. The average annual temperature is about 13 ~ 15°C, and the average annual precipitation is 1,510 mm (Fan et al., 2022). Mazongling National Nature Reserve has abundant forest resources, and the study area contains more than 10 dominant tree species, including Quercus glandulifera, Platycarya strobilacea, Castanea mollissima and Lindera glauca.

FIGURE 1

Figure 1. The location of two study areas (top left) with 10 m spatial resolution from ChinaCover2020 (Wu et al., 2017), and MZL study area (top right) and GGS study area (bottom right) with imaging spectroscopy data acquired from Cuber UHD185 Firefly imaging spectrometer (Red: 866 nm, Green: 654 nm, Blue: 566 nm). The blue triangles indicate the locations of field-measured sample plots. The green circles indicate the locations of individual tree crown (ITC) validation plots. Six photographs of these ITC plots are shown on the bottom left.

The second study area is situated in the Minya Konka National Park (also known as Gonggashan, GGS, 102°3′50′–102°4′28′E, 29°36′2′–29°36′15′N) in Ganzi (Garzê) Tibetan Autonomous Prefecture, Sichuan province of China with an elevation varying from 1959 m to 2,247 m above sea level (Figure 1). This study area is located in the transitional zone from the subtropical belt to the temperate belt of the eastern Tibetan Plateau, covering an area of approximately 20.5 ha. The average annual temperature is about 4.2°C, and the average annual precipitation is 1947 mm (Zhou et al., 2013). The forest canopy across this study area comprises more than 15 dominant tree species, including Fagus longipetiolata, Jasminum nudiflorum, Ailanthus altissima, Cercidiphyllum japonicum and Bothrocaryum controversum.

2.2. Data acquisition and preprocessing

2.2.1. UAV-borne hyperspectral and LiDAR data

The UAV-borne hyperspectral data were collected on September 18 and October 15, 2020, using the Cuber UHD185 Firefly imaging spectrometer (Cubert GmbH, Ulm, Baden-Württemberg, Germany) onboard a DJI Matrice 300 aircraft (Da Jiang innovate technology Ltd., Shenzhen, China) under cloudless conditions. The sensor comprises 125 visible and near-infrared spectral channels ranging from 450 nm to 946 nm with an 8 nm spectral resolution. The sensor was equipped on the UAV platform and flew at an altitude of 80 m, resulting in a 7 cm spatial resolution. The preprocessing of the images consisted of four preliminary steps. First, the Agisoft PhotoScan software (Agisoft LCC Co. St. Petersburg, Russia) was used for image mosaic. Spectral radiation calibration was the second step to convert the spectral response into the true spectral radiance. Then, the reflectance spectrum was calculated from the reference spectra of calibration plates and the spectral radiance. Finally, the geometric correction was performed in the Image Registration Workflow tool of ENVI5.3 software (Gai, 2019).

The UAV-borne LiDAR data were obtained simultaneously with the hyperspectral dataset acquisition using the LiAir VH Pro scanner (Green Valley Inc., Beijing, China) operating at a wavelength of 905 nm. The scanner provided a 70.4° horizontal (cross-track) and 77.2° vertical (along-track) angle of view. The height accuracy of the laser scanner was 5 cm and had an 80% flight strip overlap. The average point density in MZL is more than 117 points/m², and the average point density in MZL is more than 168 points/m². The point cloud data were noise filtered and classified into ground and vegetation returns using the software TerraSolid (Terrasolid, Helsinki, Finland). We generated the digital elevation model (DEM) based on the classified ground points and constructed the digital surface model (DSM) from the first pulse reflections of the LiDAR point clouds, and subtracted a canopy height model (CHM) with a resolution of 0.1 m (Zhao et al., 2013). The UAV-LiDAR data were normalized based on the ground points to remove the influence of terrain undulations on the height values. Besides, the vegetation point clouds with a normalized height below 2 m were removed to reduce the effect of background factors such as shrubs and grasses.

2.2.2. Field measurements

Field measurements were collected simultaneously with the UAV data acquisition in September–October 2020, and a supplemental survey was conducted in July 2022. A total of 26 square sample plots (30 × 30 m) within these two study areas were acquired. Differentially-corrected GPS determined the coordinates of the four corners of each sample plot. Tree parameters were measured in each sample plot, including tree species name, diameters at breast height (DBH), crown base height, tree height, crown classes (dominant, co-dominant, intermediate and suppressed trees) and crown diameters in two directions (south–north and east–west) for all individual trees with DBH ≥ 5 cm. The plot-level forest canopy closure and leaf area index (LAI) were also obtained by hemispherical photographs taken by a fish-eye camera along two diagonals. In addition, to validate individual tree segmentation and carry out classification research, we measured the location of each tree in two of the sample plots in MZL and four of the sample plots in GGS by integrating the Real Time Kinematic (RTK) GPS/GLONASS System with the total station.

We selected 10 dominant tree species in MZL and 15 dominant tree species in GGS and collected top-of-canopy leaves for these dominant tree species to measure their biochemical components and spectral properties. We measured 10 major biochemical components, including chlorophyll a and b (Chl-a, Chl-b), total carotenoids (Car), total carbon (C), nitrogen (N), phosphorus (P), cellulose (Cel), lignin (Lig), specific leaf area (SLA) and equivalent water thickness (EWT) similar as the previous study (Zheng et al., 2021). Leaves of each species with a mass of more than 150 g were selected and stored in plastic bags on ice and immediately transported to the laboratory for component analysis and spectroscopic measurement. Hemispherical reflectance spectra with 350–2,500 nm wavelengths were measured on 10 fresh leaves of each dominant tree species using a leaf clip coupled with the ASD FieldSpec 4 portable spectroradiometer (ASD Inc., Boulder, CO, United States). The bands with a wavelength of less than 400 nm and more than 2,400 nm were removed to eliminate the influence of instrument noise, and the spectra between 400 to 2,400 nm were smoothed by the Savitzky–Golay filter (Savitzky and Golay, 1964).

2.2.3. Species diversity indices

We used species richness, Shannon-Wiener index, and Simpson index to represent forest species diversity and calculated them within each sample plot based on the field measurements. Species richness refers to the total number of species in the sample plot. Shannon-Wiener index (Shannon, 1948) and Simpson index (Simpson, 1949) can reflect species richness and evenness of species distribution. They are comprehensive indicators reflecting the degree of species diversity. The Shannon-Wiener index is more sensitive to the number of species, and the Simpson index is more sensitive to the evenness of enriched species (Nagendra, 2002). The calculation formula of the Shannon-Wiener index (H) and Simpson index (D) were as follows:

\begin{array}{l} H = \sum_{i = 1}^{n} - p_{i} \ln p_{i} & (1) \end{array}

\begin{matrix} D = 1 - \sum_{i = 1}^{n} p_{i}^{2} \end{matrix} (2)

where n is the total number of species in the sample plot, and $p_{i}$ is the proportional abundance of the species i.

2.3. Methods

2.3.1. Individual tree crown segmentation

Based on the 0.1 m CHM data, we used a watershed algorithm combined with morphological crown control to separate the individual tree crowns (ITCs) (Wang et al., 2004; Chen et al., 2006; Zhao et al., 2014). Firstly, a morphological crown closing operator was used to determine the crown area and obtain the binary image of the canopy. Next, a local extremum algorithm was used to detect the positions of the potential individual treetop. The actual individual treetop positions and potential crown shapes were calibrated through two watershed transformations and image reconstruction operations. Finally, the crown shapes were determined using an adaptive optimized morphological crown opening operator.

2.3.2. Spectral angle mapper (SAM) classification

Among many supervised classification methods, the SAM classification was used for its better performance in the hyperspectral data (Park et al., 2004; Yang et al., 2008; Zhang and Li, 2014). The SAM algorithm is a physically based spectral classification that uses an n-dimensional angle to match the extracted endmember spectra (Kruse et al., 1993; Park et al., 2007; Mohajane et al., 2017). The SAM algorithm determines the spectral similarity though calculating the angle between the spectrum vectors. Smaller angles correspond to closer matches to the endmember spectrum.

We calculated the average spectrum of each canopy based on the ITC segmentation results in the sample plots. Firstly, pixels with NDVI <0.2 and canopy height < 2 m were removed from the hyperspectral images to reduce the effect of background factors such as canopy gaps. A total of 2 ITC plots with 14 tree species in MZL (covering more than 90% of local tree species) and a total of 4 ITC plots with 22 tree species in GGS (covering more than 75% of local tree species) were used to establish the endmember spectral library. Then we determined the average spectrum of each species in the two study areas and used the SAM algorithm to classify them according to the established spectral library. With the classification results, we acquired the species diversity indices of each sample plot and used for validation.

2.3.3. Self-adaptive fuzzy C-means (FCM) clustering algorithm

The optimal biochemical components selection followed two principles (Zhao et al., 2016): (1) biochemical components can be well inverted by the spectrum. (2) these biochemical components are sufficient to distinguish different tree species. The partial least squares regression (PLSR) was used to determine the relationships between the in-situ leaf spectral and the biochemical measurements and explore whether the biochemical components of tree species can be quantitatively estimated by their spectral signals. The PLSR method combines the advantage of principal component analysis, canonical correlation analysis, and multiple linear regression analysis. It was performed using JMP14.0 statistical software.

After the optimal biochemical components were determined, the corresponding vegetation indices (VIs) from the hyperspectral data could be selected through the existing vegetation index models to estimate the biochemical components. Due to the lack of influential lignin invention bands, we finally identified nine canopy-scale VIs to indicate Chl (Chl-a and Chl-b), Car, C, N, P, Cel, SLA and EWT based on the literature (Table 1). Many studies have confirmed that the standard deviation of VIs in an area can reflect the species diversity in this region (Cayuela et al., 2006; Stickler and Southworth, 2008; Costanza et al., 2011), so we calculated the standard deviation of VIs for all ITCs at the plot scale, and performed Spearman correlation analysis with the species diversity indices (corrplot, R-package) to select the optimal VIs. The VI for each ITC was calculated by extracting the VI of the central pixel of each ITC. These canopy-level biochemical VIs were then converted into leaf-scale biochemical VIs by dividing the canopy-level biochemical VIs by the ITC’s LAI to eliminate the effects caused by the canopy structure (Zarco-Tejada et al., 2001; Zhao et al., 2018). ITC’s LAI was calculated by establishing the relationship between forest gap fraction (GF) and LAI according to Beer–Lambert Law (Richardson et al., 2009), as shown in Formula (3):

\begin{matrix} {\begin{matrix} G F = \frac{n_{g r o u n d}}{n_{g r o u n d} + n_{v e g e t a t i o n}} \\ L A I = - \cos (θ) * \frac{\ln (G F)}{k} \end{matrix} \end{matrix} (3)

where n_ground is the number of extracted ground points, n_vegetation is the number of vegetation points, k is the extinction coefficient and takes a value of 0.5 if the vegetation is considered to follow the spherical leaf angle distribution, θ is zenith angle (LiDAR scanning angle) and GF is gap fraction.

TABLE 1

Table 1. Vegetation indices corresponding to the biochemical components.

We extracted 58 structural features for each ITC, including canopy cover, leaf area index, and height variables (statistical parameters related to point cloud height value) based on UAV-LiDAR data and the ITC-segmented ITC boundaries using LiDAR 360 software (Supplementary Table S1). Then we calculated the standard deviation of these structural variables in each sample plot. Finally, the Spearman correlation coefficient test with species diversity indices was performed to obtain the optimal structural features.

Self-adaptive Fuzzy C-Means (FCM) clustering algorithm was applied to calculate the species richness (the number of clusters) based on the optimal biochemical VIs derived from the hyperspectral image and optimal structural features obtained from LiDAR data for each ITC. Each cluster was considered to be a specific but unidentified species. Then the Shannon-Wiener index and the Simpson index can be derived from the cluster amount and the ITC number of each cluster in the sample plot [Formula (1) and (2)]. The field-measured values of species diversity indices of 26 sample plots in two study areas were then compared with the forest biodiversity prediction results to verify the estimation accuracy of the clustering algorithms.

The standard Fuzzy C-Means algorithm transforms the cluster into a nonlinear optimization problem and achieves the number of categories through iteration (Bezdek et al., 1987). Self-adaptive Fuzzy C-Means (FCM) clustering algorithm was developed from the standard Fuzzy C-Means algorithm (Li and Yu, 2009). Self-adaptive FCM automatically determines the optimal number of clusters by using a new validity function without relying on the number of pre-set categories and prior knowledge, solving the acute problem of the clustering algorithm to the initial value. The validity function is defined as:

\begin{matrix} {\begin{matrix} L (c) = \frac{\sum_{i = 1}^{c} \sum_{j = 1}^{n} \frac{u_{i j}^{m} v_{i} - {\bar{x}}^{2}}{c - 1}}{\sum_{i = 1}^{c} \sum_{j = 1}^{n} \frac{u_{i j}^{m} x_{j} - {\bar{v_{i}}}^{2}}{n - c}} \\ \bar{x} = \frac{1}{n} \sum_{i = 1}^{c} \sum_{j = 1}^{n} u_{i j}^{m} x_{j} \end{matrix} \end{matrix} (4)

where c is the number of clusters, m is the fuzzy weighting exponent, $X = {x_{1}, x_{2}, \dots, x_{n}}$ is a sample data set, $V = {v_{1}, v_{2}, \dots, v_{n}}$ is the cluster center dataset, $u_{i j}$ represents the membership if the j-th sample point belongs to the i-th class, $\bar{x}$ is the central vector of all data, and $d_{i j} = x_{j} - v_{i}$ is the Euclidean distance between the j-th sample point and j-th cluster center.

3. Results

3.1. Individual tree crown segmentation

The ITC segmentation results of all 26 sample plots show that the amounts of segmented ITCs are quantitatively close to the ground-measured tree number (MZL: R² = 0.76, RMSE = 5.41; GGS: R² = 0.82, RMSE = 7.17; Figure 2). Due to the effects of crown overlap, point cloud density, small crowns, multi-stemmed trees and other reasons, some extra trees (over-segmentation) and missed trees (under-segmentation) can be found in the ITC segmentation results. The segmented and measured position of ITCs in sample plot 2 of MZL (71 segmented vs. 74 field-measured ITCs) and sample plot 4 of GGS (55 segmented vs. 54 field-measured ITCs) are shown in Figure 2 as an example. The over-segmentation phenomenon occurs in broad-leaved trees with large crowns and non-prominent treetops. In contrast, the under-segmentation phenomenon is caused by the overlapping crowns owing to the high forest canopy density. Multiple overlapping crowns are considered as one crown and are not isolated.

FIGURE 2

Figure 2. Scatter diagram for verification of ITCs (left) and results of individual tree separation in sample plot 2 of MZL (middle) and sample plot 4 of GGS (right) (black boxes represent the boundaries of plots, white polygons refer to the segmented tree crowns, and red points represent the field-measured positions at the base of tree stems).

3.2. Forest species diversity prediction based on classification method

The SAM classification algorithm was applied to obtain the tree species of each ITC based on hyperspectral image and ITC boundaries from LiDAR data. As illustrated in Figure 3, a total of 14 endmembers in MZL and a total of 22 endmembers in GGS were extracted directly from the hyperspectral image. These tree species’ endmembers were significantly different from each other and thus could be used for classification. Figure 4 shows the tree species classification results of two typical sample plots using the SAM algorithm.

FIGURE 3

Figure 3. Endmember spectral library of dominant tree species in two study areas (left: MZL, right: GGS).

FIGURE 4

Figure 4. Tree species classification results of one sample plot in MZL (left) and one sample plot in GGS (right). Different geometric polygons represent identified ITCs, with white borders indicating the boundaries of the ITCs.

The performance of the relationships between the predicted values and the three field-measured species diversity indices (species richness, Shannon-Wiener index, and Simpson index) is shown in Figure 5 (Blue colors). In MZL, the SAM classification algorithm demonstrated positive and significant predictive validity for species richness (R² = 0.62, RMSE = 1.44), Shannon-Wiener index (R² = 0.64, RMSE = 0.16) and Simpson index (R² = 0.44, RMSE = 0.05). In GGS, the estimated values and the measured species diversity indices were positively and significantly correlated only for species richness (R² = 0.55, RMSE = 2.87) and Shannon-Wiener index (R² = 0.52, RMSE = 0.24). The classification-based prediction of the Simpson index was positively correlated with the field measurements, but the correlation was not significant (R² = 0.44, p = 0.01).

FIGURE 5

Figure 5. Field-measured species diversity indices compared with the predicted values based on classification and clustering approaches for MZL (left) and GGS (right).

3.3. Forest species diversity validation of clustering algorithm

The estimation accuracies of biochemical components based on leaf spectra of tree species are shown in Table 2. It demonstrates that: (1) In MZL, Chl-a, Chl-b, EWT, Car, SLA and C could be strongly predicted by leaf spectra based on PLSR models (R² = 0.78–0.82). Cel, N, Lig and P are also relatively quantified by spectral reflectance (R² = 0.44–0.74). (2) In GGS, Chl-a, Chl-b, EWT and SLA could be well estimated by spectral signatures (R² = 0.62–0.73). Car, N and C also perform a relatively positive relationship with spectral properties (R² = 0.30–0.46). Cel, Lig and P have no obvious correlation with spectral reflectance (R² < 0.30).

TABLE 2

Table 2. Estimation results of leaf biochemical components.

Based on the optimal biochemical components that are spectrally obtainable (Table 2), we determined 9 biochemical VIs (TCARI/OSAVI, VOG1, CRI, WBI, CCCI, RVI, PRI, NDSI and PSRI) in MZL to indicate Chl (Chl-a and Chl-b), Car, EWT, N, SLA, Cer, P and C, respectively. As for GGS, we selected 7 biochemical VIs (TCARI/OSAVI, VOG1, WBI, RVI, CRI, CCCI and PSRI) to express Chl (Chl-a and Chl-b), EWT, SLA, Car, N and C, respectively. Supplementary Figures S1, S2 show the relationships between the standard deviation of ITC’s biochemical VIs of sample plots and the species diversity indices in the two study areas. In MZL, there was a positive correlation between the standard deviation of 7 ITC-based VIs (WBI, TCARI/OSAVI, PRI, RVI, CCCI, VOG1 and PSRI) and species richness at the sample plot scale. These 7 VIs were selected as the optimal biochemical VIs. In GGS, only 3 ITC’s VIs (TCARI/OSAVI, RVI and PSRI) partially correlated with the species diversity indices. We regarded them as the optimal biochemical VIs.

The standard deviation of ITC-based structural features was weakly correlated with the species diversity indices in MZL. We finally conducted the two most relevant characteristics, namely canopy cover (CC) and density metric 30% (DM 30%) as the optimal structural features. In GGS, many structural features were positively and significantly correlated with species diversity indices. We finally determined five optimal structural features, including the interquartile range of accumulated elevation (Elev AIQ), coefficient of variance of elevation (Elev CV), the variance of elevation (Elev Var), density metric 20% (DM 20%) and density metric 30% (DM 30%), which showed high correlation with species diversity indices (Spearman correlation was above 0.5). Supplementary Figures S3, S4 show the relationships between the standard deviation of ITC’s structural features of sample plots and the species diversity indices for MZL and GGS, respectively.

We applied the Self-adaptive FCM algorithm to estimate the three species diversity indices at 26 sample plots in two study areas based on the optimal biochemical VIs and optimal structural features for each ITC. The results are shown in Figure 5 (Red colors). In MZL, the clustering algorithm demonstrated positive and significant predictive validity for Shannon-Wiener index (R² = 0.58, RMSE = 0.22) and Simpson index (R² = 0.83, RMSE = 0.06). The estimated species richness was lower than the field-measured value (RMSE = 2.47) and performed relatively unsatisfactory inversion results (R² = 0.46, p = 0.01). In GGS, the estimated value and the measured species diversity indices were positively and significantly correlated only for Simpson index (R² = 0.62, RMSE = 0.07). The prediction results for species richness (R² = 0.46, RMSE = 3.94) and Shannon-Wiener index (R² = 0.47, RMSE = 0.28) were positive but not significant enough (p = 0.01).

4. Discussion

Our results showed that the classification method performed better with higher values of R² than the clustering algorithm for predicting species richness (0.62 > 0.46 for MZL and 0.55 > 0.46 for GGS) and the Shannon-Wiener index (0.64 > 0.58 for MZL, 0.52 > 0.47 for GGS) in two study areas (Figure 5). However, the Simpson index estimated by the classification method correlated less with the field measurements than the clustering algorithm (R² = 0.44 and 0.83 for MZL and R² = 0.44 and 0.62 for GGS). This is probably due to the Simpson index weights rare species less and dominant species more than Shannon-Wiener index (Magurran, 1988; Daly et al., 2018), so the clustering algorithm taking dominant species/traits more into account are expected to predict the Simpson index accurately. Some previous studies have suggested that the Shannon-Wiener index is more closely related to species richness, while the Simpson index is more distantly correlated with richness (Nagendra, 2002; Costanza et al., 2011; Leinster and Cobbold, 2012). Our outcomes further indicated that the classification method is more advantageous in identifying rare species and estimating species richness, while the clustering method performs better in indicating the evenness of species. Constrained by the limited number of sample plots, it could be considered to use more independent validation plots to verify the advantages of classification and clustering methods in predicting species diversity indices in the future study.

We demonstrated that the individual tree-based SAM classification could be used to monitor the species diversity of complex forests and have the ability to distinguish the non-dominant species (Figure 4). This is mainly because SAM classification could distinguish similar spectra of tree species for classifying species based on hyperspectral data (e.g., Platycarya strobilacea and Tilia tuan in this experiment, Figure 3) when the endmember spectral library of dominant tree species is available (Awad, 2018; Zhao et al., 2020). However, when the spectra of non-dominant trees and dominant trees are very similar (such as Carpinus turczaninowii and Castanea mollissima in this study, Figure 3), SAM classification may also incorrectly classify them, which brings some challenges to the estimation of Shannon-Weiner and Simpson index. To better estimate species diversity using the SAM classification approach, it is necessary to extract their distinguishable bands to accurately classify these tree species. Moreover, forest structure has been identified as an essential indicator of forest species diversity (Ishii et al., 2004; Zeng et al., 2008; Guo et al., 2017; Torresani et al., 2020). We used the SAM classification to monitor species diversity based on the ITCs’ spectral signal from UAV-hyperspectral data without considering the input of structural characteristics. The fusion of spectral and structural features could increase the dissimilarity among tree species and improve classification accuracy (Torabzadeh et al., 2019). Therefore, whether integrating LiDAR-derived tree structural parameters into the supervised classification of hyperspectral data can improve species diversity monitoring is worth to be further investigated.

Our results for both study areas demonstrated that forest diversity patterns could be rapidly acquired by the Self-adaptive FCM clustering algorithm based on individual tree-based variations in biochemical and structural features without distinguishing the tree species, which is similar to the previous clustering research (Féret and Asner, 2014; Schafer et al., 2016; Zhao et al., 2018). A maximum number of 11 and 15 tree species can be identified in the sample polt of MZL and GGS based on different optimal feature compositions using the clustering algorithm (7 optimal biochemical VIs and 2 optimal structural features for MZL, 3 optimal biochemical VIs and 6 optimal structural features for GGS, Figure 6). The outcomes of our feature selection further illustrated the spectral and structural heterogeneity of different regions and also emphasized the applicability of our clustering method in subtropical forests. Compared to the previous studies using Random Forest (RF) algorithm to select the optimal features, we underlined the strength of correlation analysis between the variation of biochemical VIs or structural features and species diversity indices at sample plot scale (Xie et al., 2019; Adhikari et al., 2020; de Almeida et al., 2021). The RF algorithm filters the optimal features according to the importance of the variables, while our feature selection method considers the basic biochemical and structural principles of forest (Hall, 2000; Strobl et al., 2008). However, the biochemical composition and structural characteristics of the same tree species vary considerably depending on individual development and landscape topography, introducing much uncertainty in selecting parameters for different forests.

FIGURE 6

Figure 6. Selected features of two study areas (left: MZL, right: GGS; red labels: optimal biochemical VIs, orange labels: optimal structural features, blue labels: species diversity indices; TR/OV: TCARI/OSAVI, SW: Shannon-Wiener index).

Our results demonstrated better performance for forest species richness estimation in complex forests based on UAV-borne data (RMSE: 1.44 to 2.47 for MZL, 2.87 to 3.94 for GGS) than previous studies using airborne data (RMSE: 4.0 and 6.74) (Hernandez-Stefanoni et al., 2014; Zhao et al., 2018). Coarser image spatial resolution (typically between 1 to 10 m) and relatively lower point density (usually between 4 to 10 points/m²) of airborne data can make it difficult to identify or segment trees with smaller canopies, and image spectral mixing may also be an issue (Medina et al., 2013; Sankey et al., 2017). This affects the accuracy of forest species diversity monitoring, as the spectral and structural differences between species may not be accurately captured (Ustin et al., 2004; Lesak et al., 2011; Naidoo et al., 2012). In contrast, UAV-borne LiDAR data with higher point cloud density (more than 100 points/m²) could discriminate and detect individual trees with satisfactory accuracies (Figure 2). The spectral mixture problem would be solved with the ultra-high resolution UAV-borne imagery (Somers et al., 2011; Ronay et al., 2022), but how to better represent the spectral features of each ITC and avoid potential noise caused by intra-crown shade still need to be further studied (Rocchini et al., 2010). Given the lower flight altitude than conventional airborne platforms, the UAV-borne hyperspectral images are less affected by the atmosphere, leading to improved image quality and easier processing. UAV remote sensing has improved the timeliness of data acquisition, but it has limitations such as limited payload, short flight life, and more fabulous mosaic and geocode efforts (Nex and Remondino, 2014; Matese et al., 2015; Pu, 2021). In addition, due to the “top-down” operation method of UAV-borne platforms, the data for the understory in dense forest areas are often missing. Therefore, combining the advantages of different monitoring tools, such as ground-based LiDAR to complement and verify each other can provide more information for related forest diversity research.

5. Conclusion

In this study, we compared the performance of individual tree-based classification and clustering methods with UAV-borne data for estimating the forest species diversity indices in the Mazongling and Gonggashan National Nature Forest Reserves of China. We proved that the SAM classification could provide more accurate predictions of species richness indices but requires spectral information of all dominant tree species. The Self-adaptive FCM clustering algorithm could achieve high-precision predictions for evenness indices (especially Simpson index), although information on specific tree species is unavailable.

The combination of UAV imaging spectroscopy and LiDAR make it possible to predict regional forest species diversity more accurately at individual canopy scale for complex forests. Future studies could improve the forest species-spectral library and explore forest species identification from multiple perspectives. Additionally, considering the variation in forest species characteristics over time, it would be valuable to further examine the accuracy of classification and clustering methods by incorporating phenological or multi-temporal features. Moreover, it would be beneficial to investigate the applicability of species diversity estimation models for forests in different ecological contexts and how high-resolution UAV data can be leveraged to bridge the scale gap between traditional field plot samplings and large-scale satellite observations.

Data availability statement

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

Author contributions

XL collected field data, calculated and analyzed the data, and wrote the manuscript. YZ and ZZ assisted in guiding the design of the experiment, discussing the results, and revising the manuscript. DZ and CX helped design the experiment and contributed to the manuscript. PZ, JC, JW, XZ, and XM supported the data investigation. All authors contributed to the article and approved the submitted version.

Funding

This work was supported by the National Key Research and Development Program of China (2020YFE0200800 and 2022YFF1302100) and the National Natural Science Foundation of China (No. 42071344).

Acknowledgments

We thank Ruiying Chang and Qiangxin Ou from the Gonggashan and Mazonglin National Nature Forest Reserves for their support with the fieldwork. We also thank Mingxing Zhang and Shaobo Yang for their assistance in the field sample collection.

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fevo.2023.1139458/full#supplementary-material

References

Adhikari, H., Valbuena, R., Pellikka, P. K. E., and Heiskanen, J. (2020). Mapping forest structural heterogeneity of tropical montane forest remnants from airborne laser scanning and Landsat time series. Ecol. Indic. 108:105739. doi: 10.1016/j.ecolind.2019.105739